fix(tools): reject mixed-root freeform patches

* fix(tools): make apply_patch tolerant and re-expose edit_file

Head-to-head benchmarking against Grok Build on grok-4.6 (3 tasks × 3 runs)
showed Zero solving every task but spending 2.7× the tokens and 2.3× the
time on the multi-file feature task. Every one of the 20 apply_patch calls
in those runs failed, after which the model degraded to whole-file
write_file rewrites and read→write verification loops.

Root causes and fixes:

- The sandbox and the tool both rejected an absolute path that lies inside
  the workspace ("must stay inside the workspace"). Models routinely echo
  the absolute path read_file showed them. Absolute paths now flow through
  the normal workspace-scope validation (inside → allowed, outside → still
  denied); unified-diff headers are rewritten to a/ b/ relative form before
  git apply, and platform symlink prefixes (/var → /private/var) are
  normalised the same way the other file tools do.
- The structured parser demanded a byte-exact "*** Begin Patch" /
  "*** End Patch"; grok-4.6 writes "*** Begin Patch ***". Decorated marker
  spellings are accepted, and a unified range after "@@" ("-12,4 +12,6",
  optionally with a heading) no longer becomes a literal context search.
- Parse errors now carry a one-line format reminder so a failed call is
  recoverable in one retry instead of a format-guessing loop.
- edit_file (exact search/replace with uniqueness guard, CRLF and fuzzy
  fallbacks) is advertised to the model again alongside apply_patch, and
  the apply_patch description shows the structured format.
- Writers (edit_file, write_file, apply_patch) recorded a file's line total
  with lineCount, which is one higher than read_file's count for a file
  ending in a newline. RecordSeenRange resets every observation when the
  total changes, so a partial read_file after a successful edit silently
  discarded whole-file knowledge and the next edit was refused with
  "content that has not been read exactly in this session". All writers
  now record the same total read_file reports (trackedLineTotal).
- read_file's padded "  N | " line prefix was ~19% of every read (about
  1.5k tokens per 800-line file) and was re-sent on every later call. It is
  now the compact "N→" prefix models already know; nothing parsed the old
  form.
- The editing-discipline prompt steers: edit_file for a targeted change,
  apply_patch for multi-hunk/multi-file changes, write_file only for new
  files or full rewrites; a successful edit result is confirmation, so the
  file is not re-read to verify it.

Tests: regression tests for decorated markers, writer/reader line-total
agreement (edit → partial read → edit far away still succeeds), range hunk headers,
absolute in-workspace paths (structured + unified), outside-workspace
rejection, header-line-only relativisation, and tool visibility; the
prompt and tool-budget tests updated to the new contract.

Benchmark (grok-4.6, 3 tasks x 3 runs, harness in demo/bench-zero-vs-grok):
Zero 0.8.0 617 s / 1,527k tokens -> with this change 346 s / 493k tokens;
Grok Build 407 s / 713k. Every run of every configuration solved its task.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): keep CRLF line endings when relativizing patch headers

relativizeUnifiedPatchPaths split on a normalized "\n" and rejoined with
"\n", so any CRLF patch that needed a header rewrite lost its line endings
everywhere; git apply could then fail to match a CRLF file. Split on "\n"
only and carry each line's trailing "\r" through, including on rewritten
headers.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(sandbox): share the structured-patch classifier with apply_patch

Review follow-ups for #956:

- The tool accepted decorated structured markers ("***Begin Patch") that
  the sandbox's applyPatchPaths did not recognise, so for that spelling the
  boundary scanned the patch as a unified diff, extracted no targets and
  validated nothing. The marker classifier now lives in the sandbox package
  (StructuredPatchMarker / IsStructuredPatch) and the tool delegates to it,
  so both sides accept exactly the same set; tests cover inside, traversal
  and outside targets for the no-space spelling at the boundary.
- Document the containment model at the apply site: structured patches are
  applied through an opened os.Root (descriptor-relative, no-follow);
  unified diffs keep the pre-existing git-apply pathname flow, which the
  absolute-path rewrite does not widen (they become the same root-relative
  form relative paths always used).
- OpenAI addendum: the one-file-per-call rule names read_file, edit_file and
  write_file explicitly and allows a single multi-file apply_patch.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* test(sandbox): compare slash-normalised patch paths so the check passes on Windows

structuredPatchHeaderPaths normalises separators to "/", so the absolute
outside-workspace path in the boundary test must be compared in its slash
form; on Windows the backslash original never matched and the Smoke job
failed.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* test(sandbox): validate the parsed patch path at the scope boundary

For each accepted marker spelling, pass the exact path applyPatchRequestPaths
produced to scope.validate: an outside path must be denied and an inside
path accepted, so a separator-handling regression between the parser and the
scope cannot slip past the test.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): apply unified diffs in-process through the workspace root

Review follow-up for #956. Unified diffs were validated by pathname and then
handed to git apply, which reopened the targets by name — a check-to-use
window a symlink swap could exploit — and the absolute-path rewrite added in
this PR normalised `diff --git` headers with strings.Fields, which splits a
git C-quoted path containing whitespace.

Both go away by not shelling out at all: parseUnifiedPatch translates a
unified diff into the same operations a structured patch produces, and both
formats are applied by applyPatchOperations through os.OpenRoot — every
stat, read, create and write is descriptor-relative and refuses to follow a
link out of the root. git is no longer required by apply_patch.

Unified-diff support: modifications, creations (--- /dev/null), deletions
(+++ /dev/null), multiple hunks per file, git "rename from/to" (a move,
hunks optional) and "copy from/to" (a new copy operation whose destination
inherits only the tracker credit the source had), "\ No newline at end of
file" on either side, a/ b/ prefixes, C-quoted paths, CRLF patches, and
git's diff/index/mode lines. Hunks are placed at their "@@ -a,b" range when
the expected lines match there and located by context otherwise (so a
stale range still applies and duplicate blocks are disambiguated by the
range). Binary patches are rejected.

Removed: the git apply flow, relativizeUnifiedPatchPaths,
recheckPatchWriteTargets, missingPatchTargets, completeCreatedPatchTargets
and recordCreatedPatchTargets. validatePatchPaths/changedFilesFromPatch stay
for mutation-target listing.

Tests: the existing unified-diff suite now runs through the in-process
applier unchanged; new coverage for quoted whitespace paths (inside applies,
outside is refused untouched), a workspace path swapped for a symlink that
escapes the root (both formats refused, outside file unmodified), create +
delete + no-newline markers, range hint vs context fallback, rename with a
hunk, and binary/empty rejection.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): harden unified-diff parsing and cover the copy operation

Review follow-ups for #956 (in-process unified applier):

- A hunk that ends before its declared line counts is reported as
  malformed instead of silently absorbing the next file's ---/+++ header
  pair (which then failed later against the wrong file). Both shapes are
  caught: leftover counts at a header or at end of input, and a --- line
  directly followed by +++ inside a hunk. The element after a patch's final
  newline is no longer counted as an empty context line, so a truncated
  hunk cannot look complete.
- "\ No newline at end of file" may only follow a file's final hunk; a
  marker before another hunk is rejected rather than leaking into that
  hunk's result.
- Zero-context insertions (git diff -U0) are documented as
  position-trusting by design, matching git apply without fuzz; hunks with
  context are verified at the range before it is used.

Tests: copy from/to (source kept, destination created with the hunk
applied and tracked with the source's whole-file state; existing
destination refused untouched; copy onto itself refused), over-declared and
truncated hunk counts, and the no-newline marker before vs after the final
hunk.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): require an @@ line before treating ---/+++ inside a hunk as a file boundary

A removed "-- x" line directly followed by an added "++ y" line is valid
hunk content that also matches the ---/+++ header shape; the boundary check
now additionally requires the following line to be a hunk header, and a
regression test keeps such a pair intact. End-of-input strictness for a hunk
cut short is documented as deliberate: applying a truncated hunk would make
a partial change (removals without their replacements).

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* refactor(tools): share one line-count implementation between reader and writer

sourceLineCount and trackedLineTotal computed the same count; keeping two
copies invites exactly the reader/writer drift this PR fixed. The reader now
delegates to trackedLineTotal, which carries the tracker contract.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): verify a unified deletion's expected content before removing the file

A unified diff's "+++ /dev/null" hunks state the content being removed, but
the in-process applier dropped them and deleted unconditionally, so a stale
deletion removed a file whose current content the patch never described.
Unified deletions now keep their hunks, require at least one, and run them
through the same verifier as updates: the hunks must match the current file
and leave nothing behind, otherwise the deletion is refused and the file is
untouched. A structured "*** Delete File" keeps its unconditional meaning.

Hardening from the same review: every non-create change re-reads its source
through the opened root immediately before commit and refuses to proceed if
the bytes differ from what was planned.

Tests: matching deletion, stale deletion (file byte-for-byte unchanged),
partial deletion, header-only deletion, multi-hunk deletion with one stale
hunk (nothing removed) and all-good hunks, structured delete unchanged, and
a source changed between planning and commit for update and delete.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* test(tools): a move with a missing source is refused before publishing

The pre-commit recheck now refuses a move whose source vanished after
planning instead of publishing the destination and then failing; update the
test to the safer contract (error names the recheck, nothing is created).

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): byte-exact unified deletions and git's header-only empty-file forms

Review follow-ups for #956:

- A unified deletion is verified byte for byte: the removed lines (with the
  removed side's "\ No newline at end of file" honoured) must equal the
  current file exactly, with only the file's own line-ending style
  normalised. None of the whitespace tolerance used to locate update hunks
  applies, and a whitespace-only remainder no longer passes; a deletion is
  not recoverable, so it fails closed.
- git describes an empty file's deletion or creation with header lines only
  ("diff --git", "deleted file mode"/"new file mode", "index") and no
  hunks. Both forms are recognised: the deletion is applied only when the
  current file is empty, the creation writes an empty file, and they compose
  with hunk-carrying files in the same patch.
- The pre-commit recheck test also covers the copy operation and asserts no
  destination is created on refusal.

Tests: whitespace/indentation/blank-line/space-only/missing-newline
mismatches all refused with the file unchanged, matching no-newline marker
and CRLF file accepted, header-only deletion of a non-empty file refused,
multi-file patch with header-only delete + create + a normal hunk.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* test(tools): remove the move source after planning via a deterministic hook

The refused-move test now creates source.txt, plans the move, and removes
the file through a test-only pre-commit hook (structuredPatchBeforeCommit)
so it exercises the post-planning disappearance path rather than a source
that was never there. Assertions unchanged: refused before publishing, no
destination created.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* test(tui): read_file fixtures use the compact N→ prefix

Three TUI fixtures still built fake read_file output in the legacy
"  N | text" shape; they pass either way but pinned an outdated contract.
They now mirror what appendReadFileRange emits, and the shape comment and
the expanded-card assertion are updated.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

* fix(tools): name the committed files when a patch is partially applied

Review follow-up for #956. When a later change in a multi-file patch fails
(for example the pre-commit recheck refuses a file that changed after
planning), the error now lists exactly which files were already committed
and states that the rest are unchanged, instead of a generic "partially
applied" warning. The model can then re-read only what actually changed.

Test: a three-file structured patch whose second file is altered after
planning fails with "already committed: first.txt", first.txt holds the
patched content, and the other two files are untouched.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

---------

Co-authored-by: Claude Fable 5 <[email protected]>

.release-please-manifest.json

@@ -1 +1 @@
-{".":"0.7.0"}
+{".":"0.8.0"}

AGENTS.md

@@ -111,3 +111,42 @@ These classes drive multi-round reviews. Fix them before requesting review:
   shipped. Wire advertised entry points or shrink the claim. Do not bundle
   unrelated fixes. Do not widen security allowlists by name alone; check
   classification and side effects too.
+- **Host tool version floors:** Zero shells out to `git` and other host tools,
+  and the project does not declare a minimum version for any of them. A flag,
+  subcommand, or config key that is not long-established therefore raises that
+  floor silently, and the failure on an older host is an unhelpful usage error
+  rather than a clear message. Before adding one, check when it landed and say
+  so where a caller would look. Record the floor for the exact subcommand you
+  call, not for the flag in general: `--end-of-options` reached Git's shared
+  parser in 2.24, but `git rev-parse` only accepted it from 2.30, and
+  `git checkout` and `git reset` later still. Likewise `--show-current` is
+  `git branch` 2.22, and `branch.autoSetupMerge=inherit` is 2.35. Then either
+  keep a fallback for older versions or gate the call. A test that depends on a
+  newer feature must skip below its floor instead of failing, and say which
+  version it needs.
+- **Prove the test fails without the change:** A regression test earns its
+  place only if it fails on the unfixed code, for the reason it claims. Run it
+  both ways and quote the failure in the PR. The recurring miss is not a
+  missing test but a test that never reaches the behavior it names, because an
+  earlier guard rejects the input first, so it passes with and without the fix.
+  When a test targets a specific layer, call that layer directly rather than
+  relying on a public entry point that may refuse earlier.
+- **Hermetic tests:** Tests must not read or write the developer's real config,
+  cache, or state directories. Redirect every root the code under test
+  resolves, not just config. `os.UserConfigDir` and `os.UserCacheDir` read
+  `XDG_CONFIG_HOME` and `XDG_CACHE_HOME` only on Linux and the BSDs. On macOS
+  they derive from `HOME`, and on Windows from `%AppData%` and `%LocalAppData%`.
+  Setting the XDG variables alone still leaves a macOS run writing to the real
+  home, so set `HOME` as well, or take the directory as a parameter. Apply it to
+  every test that reaches the same storage, not only the one where it was first
+  needed; a test that passes on a clean machine and pollutes a real one is a
+  blocker.
+- **Error codes vary by platform and by flag combination:** A check written
+  against the code one target returns can be silently inert on another while
+  still looking correct. `openat(O_DIRECTORY|O_NOFOLLOW)` on a symlink may
+  report `ENOTDIR` or `ELOOP` depending on the platform, because the kernel can
+  reject the directory mismatch before it ever resolves the link; accept both
+  rather than assuming the one POSIX names. On Windows, `\\?\` and `\\.\`
+  prefixes are not UNC paths, and an `NtCreateFile` information class must match
+  the struct actually passed. Confirm the code each supported target returns
+  instead of reasoning from the man page.

CHANGELOG.md

@@ -6,6 +6,27 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.
 
+## [0.8.0](https://github.com/Gitlawb/zero/compare/v0.7.0...v0.8.0) (2026-08-21)
+
+
+### Features
+
+* **security:** path-jail primitive, credential-store locking, worktree git hardening ([#891](https://github.com/Gitlawb/zero/issues/891)) ([060d38c](https://github.com/Gitlawb/zero/commit/060d38ce716feb7c5b6657b082cd35cb8dda3d75))
+* **update:** recognise Homebrew installs and add mise, go install docs ([#910](https://github.com/Gitlawb/zero/issues/910)) ([01c3498](https://github.com/Gitlawb/zero/commit/01c3498205e69d26e5f929225f9455b276444f39))
+* use Exa as default search provider ([#926](https://github.com/Gitlawb/zero/issues/926)) ([1ec7219](https://github.com/Gitlawb/zero/commit/1ec7219aee7b2840481ab0905f131644e27d2a4e))
+* **zerogit:** auto-create a conventional branch before push/pr on default branch ([#671](https://github.com/Gitlawb/zero/issues/671)) ([d065467](https://github.com/Gitlawb/zero/commit/d065467cd3ec3ed91c5dc60554c36fc4b4d6de4c))
+
+
+### Bug Fixes
+
+* **acp:** three defects a desktop ACP client hits ([#915](https://github.com/Gitlawb/zero/issues/915)) ([a78c6c3](https://github.com/Gitlawb/zero/commit/a78c6c357e1d21146d7cd2e8030a754e06f7f1f5))
+* bump Go to 1.26.6 for stdlib vulnerability fixes ([#903](https://github.com/Gitlawb/zero/issues/903)) ([dc15e82](https://github.com/Gitlawb/zero/commit/dc15e822987d8d7de467c62c375518f435b74d8c))
+* **daemon:** clean up child after startup timeout ([#774](https://github.com/Gitlawb/zero/issues/774)) ([9c29209](https://github.com/Gitlawb/zero/commit/9c292092a26c9092fd873d072b0af1b0d9d39ef0))
+* **dictation:** redact API keys from streaming transcriber errors ([#852](https://github.com/Gitlawb/zero/issues/852)) ([64a783b](https://github.com/Gitlawb/zero/commit/64a783b08caa2c7b21f188033b9ed84a14e00e84))
+* **modelregistry:** expose reasoning effort for DeepSeek V4 models ([#931](https://github.com/Gitlawb/zero/issues/931)) ([90dcfd1](https://github.com/Gitlawb/zero/commit/90dcfd127e8a6d9902ec9f4e72f6d03fef1a0fc6))
+* **providers:** discover ChatGPT capabilities ([#890](https://github.com/Gitlawb/zero/issues/890)) ([2d2450e](https://github.com/Gitlawb/zero/commit/2d2450e9a744349f0d01b1d4e9ba29c24ba5650d))
+* **sandbox:** normalize launcher names before the command-prefix denylist ([#934](https://github.com/Gitlawb/zero/issues/934)) ([6edf9a8](https://github.com/Gitlawb/zero/commit/6edf9a8b78dc030dc44598919db1c7fa9d4f809a))
+
 ## [0.7.0](https://github.com/Gitlawb/zero/compare/v0.6.0...v0.7.0) (2026-08-10)
 
 

README.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
   <br>
@@ -76,7 +76,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### From source
 
-Source builds require Go 1.26.5+.
+Source builds require Go 1.26.6+.
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -371,9 +371,13 @@ go run ./cmd/zero-perf-bench
 
 Experimental: `ZERO_OPENAI_TURN_SESSION=1` enables the optimized OpenAI turn
 session (background connection prewarm + request-prefix telemetry) for headless
-`zero exec` runs against official OpenAI profiles. Off by default; `0`/`false`
-disable. A/B-benchmark it by running the same `zero-perf-bench` suite with the
-variable unset and set.
+`zero exec` runs against official OpenAI profiles. Off by default; `0`, `false`,
+or `off` disable it. A/B-benchmark it by running the same `zero-perf-bench` suite
+with the variable unset and set.
+
+Native ChatGPT Responses sessions are enabled by default. Set
+`ZERO_CHATGPT_TURN_SESSION=0`, `false`, or `off` to restore stateless HTTP/SSE
+transport.
 
 ### Code Quality and Security Checks
 

README_ZH.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
   <br>
@@ -57,7 +57,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### 从源码构建
 
-源码构建需要 Go 1.26.5+。
+源码构建需要 Go 1.26.6+。
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git

SECURITY.md

@@ -33,4 +33,4 @@ Please include, where possible:
   depend on a user explicitly disabling the sandbox (`--skip-permissions-unsafe`), are generally
   out of scope — but please report anything you are unsure about.
 - Zero sends no telemetry. Network calls go to the model/tool providers you configure (and, for
-  keyless web search, to the documented Firecrawl endpoint — see the README).
+  keyless web search, to Exa's hosted MCP endpoint at `https://mcp.exa.ai/mcp`).

docs/INSTALL.md

@@ -1,14 +1,30 @@
 # Installing Zero
 
-Zero is distributed as:
+Pick whichever fits how you already install things. None of these is the
+blessed one.
 
-- an npm package, `@gitlawb/zero`
-- release archives on GitHub Releases
-- source builds with Go 1.26.5+
+| Method | Command | Self-update |
+| --- | --- | --- |
+| Install script (Linux, macOS) | `curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh \| bash` | `zero upgrade` |
+| Install script (Windows) | `irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 \| iex` | `zero upgrade` |
+| npm | `npm install -g @gitlawb/zero` | `zero upgrade` (runs npm) |
+| Release archive | download from [Releases](https://github.com/Gitlawb/zero/releases) | `zero upgrade` |
+| mise | `mise use -g ubi:Gitlawb/zero` | `mise upgrade` |
+| `go install` | `go install github.com/Gitlawb/zero/cmd/zero@latest` | rerun the command |
+| Source | `go build -o zero ./cmd/zero` | rebuild |
 
-The install scripts download a platform-specific release archive and require a
-published GitHub Release for the requested version. The npm package is
-self-contained: the platform binary installs from the npm registry.
+`zero upgrade` follows whatever owns the binary, with one exception it cannot
+detect. It delegates to npm for an npm install, refuses outright for a Homebrew
+keg and names `brew upgrade zero`, and replaces the binary in place for
+everything else — including a mise-managed one, which is the exception.
+
+mise is the one case it cannot detect: a mise-managed binary is an ordinary
+standalone install on disk, so `zero upgrade` will replace it and leave mise
+describing a version that is no longer there. Use `mise upgrade` instead.
+
+Release archives are the substrate for most of the above: the install scripts
+and the npm fallback both download a platform archive from a published GitHub
+Release and verify its checksum.
 
 ## npm
 
@@ -111,6 +127,45 @@ Defaults:
 - Version: latest GitHub release
 - Install path: `%LOCALAPPDATA%\zero\bin\zero.exe`
 
+## mise
+
+[mise](https://mise.jdx.dev/) installs Zero straight from the GitHub Release
+archives through its `ubi` backend, so there is nothing extra to publish and no
+registry in the middle:
+
+```bash
+mise use -g ubi:Gitlawb/zero
+```
+
+Pin a version the same way you would any other tool:
+
+```bash
+mise use -g ubi:Gitlawb/[email protected]
+```
+
+Updates come from `mise upgrade`. `zero upgrade` also works, because a
+mise-managed binary is an ordinary standalone install, but then mise's records
+describe a version that is no longer on disk. Prefer `mise upgrade`.
+
+## go install
+
+```bash
+go install github.com/Gitlawb/zero/cmd/zero@latest
+```
+
+This builds from source, so it needs Go 1.26.6+ and it does not go through the
+release archives. Two consequences worth knowing before you pick it:
+
+- On Linux you also need the sandbox helper, which is a separate binary and is
+  not installed by this command. See
+  [Sandbox Helpers For Source Builds](#sandbox-helpers-for-source-builds).
+  Without it, native sandboxing is unavailable.
+- `zero upgrade` treats the result as a standalone install and will replace the
+  binary with a release build rather than rebuilding from source. If you chose
+  `go install` deliberately, rerun it instead.
+
+macOS and Windows need no extra helper.
+
 ## From Source
 
 ```bash
@@ -125,7 +180,7 @@ Build a local binary:
 go build -o zero ./cmd/zero
 ```
 
-Source builds require Go 1.26.5+.
+Source builds require Go 1.26.6+.
 
 ### Sandbox Helpers For Source Builds
 
@@ -242,3 +297,23 @@ zero upgrade
 ```
 
 See the [update guide](UPDATE.md) for update modes, flags, and platform details.
+
+`zero upgrade` behaves differently depending on how Zero was installed, because
+overwriting a binary a package manager owns leaves that manager describing a
+version that is no longer there:
+
+- **npm**: runs `npm install -g @gitlawb/zero@latest` for you.
+- **Homebrew**: refuses, and tells you to run `brew upgrade zero`. Replacing the
+  keg binary directly would be reverted by the next `brew upgrade` or
+  `brew reinstall`.
+- **everything else**: downloads the verified release archive and replaces the
+  binary in place.
+
+Homebrew is detected by the binary living inside a Cellar keg that carries
+Homebrew's own `INSTALL_RECEIPT.json`. The keg shape alone is not enough: a
+directory tree you built yourself can have exactly that shape, and refusing to
+update an install Homebrew never touched would be a real cost for a made-up
+reason. The receipt is the thing only Homebrew writes.
+
+An ordinary install under `/usr/local/bin` is left alone even on an Intel Mac
+where that is also the Homebrew prefix.

go.mod

@@ -4,7 +4,7 @@ go 1.26.6
 
 require (
 	charm.land/bubbles/v2 v2.1.1
-	charm.land/bubbletea/v2 v2.0.8
+	charm.land/bubbletea/v2 v2.0.9
 	charm.land/lipgloss/v2 v2.0.5
 	github.com/Microsoft/go-winio v0.6.2
 	github.com/alecthomas/chroma/v2 v2.27.0

go.sum

@@ -1,7 +1,7 @@
 charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=
 charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=
-charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=
-charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
+charm.land/bubbletea/v2 v2.0.9 h1:DpJCMWKgzQK8SJv4zbKKFHAI10ymWy/evClPFk0k0f8=
+charm.land/bubbletea/v2 v2.0.9/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
 charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
 charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERoyfY=
@@ -64,8 +64,6 @@ github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavM
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976 h1:X8Hz2ImujgbmetVuW+w2YkyZChE3cBpZi2P158rTG9M=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976/go.mod h1:vnf4pv9iKZXY58sQE1L86zmNWJ4159e1RkcWiLCkeEY=
-golang.org/x/image v0.44.0 h1:+tDekMZED9+LrtB3G5xzRggpVh9CARjZqROla3R3R+I=
-golang.org/x/image v0.44.0/go.mod h1:V8K3KE9KKKE+pLpQDOeN18w9oacNSvy1tDOirTu4xtY=
 golang.org/x/image v0.45.0 h1:FMb1nTbH5H9vF55SriQHgFw5GnNL9Jg6L25BwXKzhB0=
 golang.org/x/image v0.45.0/go.mod h1:n62x/7RqlwXDvGsSU4u6IUTUf6KghUZ9Bt7cG/T9Fx4=
 golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=

internal/acp/agent.go

@@ -296,6 +296,16 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 		if errors.Is(err, context.Canceled) {
 			return StopCancelled, nil
 		}
+		// A CLIENT THAT CANCELLED A PERMISSION PROMPT DID NOT FAIL. Only
+		// context.Canceled was matched here, so dismissing a permission dialog —
+		// a deliberate action, and for apply_patch the only refusal a client is
+		// offered — came back as JSON-RPC -32603 carrying the internal sentinel
+		// text. Editors and the desktop app both render that as a crashed turn,
+		// so declining a tool looked like ZERO falling over. It is a
+		// cancellation, and StopCancelled is what ACP has for saying so.
+		if errors.Is(err, agent.ErrPermissionApprovalCanceled) {
+			return StopCancelled, nil
+		}
 		return "", err
 	}
 	if result.FinishReason == "length" {
@@ -311,10 +321,14 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 // session/request_permission request and maps the outcome back. Failure to reach
 // the client fails closed to deny.
 func (a *Agent) requestPermission(ctx context.Context, sessionID string, req agent.PermissionRequest) (agent.PermissionDecision, error) {
+	// Built ONCE and used for both halves. Sending one list and validating the
+	// reply against another is the defect this whole path had; see
+	// decisionFromOutcome.
+	options := buildPermissionOptions(req)
 	params := RequestPermissionParams{
 		SessionID: sessionID,
 		ToolCall:  permissionToolCall(req),
-		Options:   buildPermissionOptions(req),
+		Options:   options,
 	}
 	var result RequestPermissionResult
 	if err := a.conn.Call(ctx, MethodSessionRequestPerm, params, &result); err != nil {
@@ -323,7 +337,7 @@ func (a *Agent) requestPermission(ctx context.Context, sessionID string, req age
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission request failed: " + err.Error()}, nil
 	}
-	return decisionFromOutcome(result.Outcome, req.AvailableDecisions), nil
+	return decisionFromOutcome(result.Outcome, options), nil
 }
 
 func (a *Agent) emitPlan(registry *tools.Registry, note *notifier) {

internal/acp/desktopinterop_test.go

@@ -0,0 +1,217 @@
+package acp
+
+// desktopinterop_test.go covers three defects found by driving the real `zero
+// acp` binary from a desktop ACP client. Each is a case where ZERO and a
+// conforming client disagreed about the wire, and none of them failed loudly —
+// they produced a silent denial, a spurious crash, or two buttons the user
+// could not tell apart.
+//
+// A fourth, session/load restoring history without replaying it, is fixed by
+// #914 rather than here: that one also gives the replayed messages stable ids
+// and keeps session/resume replay-free, which this did not.
+
+import (
+	"errors"
+	"fmt"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+)
+
+// ---- an offered option is an acceptable answer ----
+
+// The list the options were BUILT from and the list the answer was CHECKED
+// against were different whenever ZERO did not enumerate: the client was sent
+// Allow and Reject, and its reply was validated against an empty slice, so
+// every button it could possibly show failed closed to deny. The user clicked
+// Allow and ZERO recorded a denial, with nothing on screen to say so.
+func TestAnOfferedOptionIsAccepted(t *testing.T) {
+	// No AvailableDecisions: the fallback path, which is every permission event
+	// that is not a prompt.
+	req := agent.PermissionRequest{ToolName: "bash"}
+
+	options := buildPermissionOptions(req)
+	if len(options) == 0 {
+		t.Fatal("no options were offered")
+	}
+
+	for _, option := range options {
+		decision := decisionFromOutcome(
+			RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: option.OptionID},
+			buildPermissionOptions(req),
+		)
+		if string(decision.Action) != option.OptionID {
+			t.Fatalf("option %q was answered with %q (%s) — an offered option must be accepted",
+				option.OptionID, decision.Action, decision.Reason)
+		}
+	}
+}
+
+// The narrowing this validation exists for still holds: a client must not be
+// able to return a broader grant than it was shown.
+func TestAnOptionThatWasNotOfferedIsStillRefused(t *testing.T) {
+	req := agent.PermissionRequest{ToolName: "bash"}
+
+	decision := decisionFromOutcome(
+		RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(agent.PermissionDecisionAlwaysAllow)},
+		buildPermissionOptions(req),
+	)
+	if decision.Action != agent.PermissionDecisionDeny {
+		t.Fatalf("an unoffered broader grant was accepted as %q", decision.Action)
+	}
+}
+
+// The resolver is the single source of truth for both sides, which is what
+// makes the two agree by construction rather than by being kept in step.
+func TestOfferedDecisionsMatchWhatWasSent(t *testing.T) {
+	for _, req := range []agent.PermissionRequest{
+		{ToolName: "bash"},
+		{ToolName: "bash", AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowForSession,
+			agent.PermissionDecisionDeny,
+		}},
+	} {
+		offered := buildPermissionOptions(req)
+		for _, option := range buildPermissionOptions(req) {
+			if !actionOffered(option.OptionID, offered) {
+				t.Fatalf("option %q was sent but is not in the offered set %v", option.OptionID, offered)
+			}
+		}
+	}
+}
+
+// ---- two options the user can tell apart ----
+
+// request_permissions offers plain allow and strict-review allow together, and
+// both were labelled "Allow". The label is the only thing an ACP client shows,
+// so the panel presented two identical buttons where one silently enabled
+// strict auto-review of what was granted.
+func TestEveryOfferedOptionHasADistinctLabel(t *testing.T) {
+	req := agent.PermissionRequest{
+		ToolName: "request_permissions",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowStrict,
+			agent.PermissionDecisionAllowForSession,
+			agent.PermissionDecisionDeny,
+		},
+	}
+
+	seen := map[string]string{}
+	for _, option := range buildPermissionOptions(req) {
+		if previous, clash := seen[option.Name]; clash {
+			t.Fatalf("options %q and %q share the label %q — a user cannot tell them apart",
+				previous, option.OptionID, option.Name)
+		}
+		seen[option.Name] = option.OptionID
+	}
+}
+
+// The two remain distinct decisions: the label changed, the round trip did not.
+func TestStrictAllowStillRoundTrips(t *testing.T) {
+	req := agent.PermissionRequest{
+		ToolName: "request_permissions",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowStrict,
+		},
+	}
+	decision := decisionFromOutcome(
+		RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(agent.PermissionDecisionAllowStrict)},
+		buildPermissionOptions(req),
+	)
+	if decision.Action != agent.PermissionDecisionAllowStrict {
+		t.Fatalf("strict allow round-tripped as %q", decision.Action)
+	}
+}
+
+// ---- cancelling a permission is a cancellation, not a crash ----
+
+// Only context.Canceled was recognised, so dismissing a permission dialog came
+// back as JSON-RPC -32603 carrying the internal sentinel text. Clients render
+// that as a failed turn, so declining a tool looked like ZERO falling over —
+// and for apply_patch, dismissing is the ONLY refusal a client is offered.
+func TestCancellingAPermissionEndsTheTurnAsCancelled(t *testing.T) {
+	canceled := fmt.Errorf("%w for apply_patch", agent.ErrPermissionApprovalCanceled)
+
+	reason, err := stopReasonFor(agent.Result{}, canceled)
+	if err != nil {
+		t.Fatalf("stopReasonFor returned an error for a cancellation: %v", err)
+	}
+	if reason != StopCancelled {
+		t.Fatalf("stop reason = %q, want %q", reason, StopCancelled)
+	}
+}
+
+// A genuine failure must still be an error. Mapping too much to cancelled would
+// hide real faults behind a clean ending, which is the opposite mistake.
+func TestARealFailureIsStillAnError(t *testing.T) {
+	failure := errors.New("provider exploded")
+
+	if _, err := stopReasonFor(agent.Result{}, failure); !errors.Is(err, failure) {
+		t.Fatalf("stopReasonFor(%v) = %v, want the failure preserved", failure, err)
+	}
+}
+
+// The sentinel is matched through errors.Is, so it survives the wrapping every
+// return site applies — each one adds the tool name.
+func TestPermissionCancellationSurvivesWrapping(t *testing.T) {
+	if !errors.Is(fmt.Errorf("%w for bash", agent.ErrPermissionApprovalCanceled), agent.ErrPermissionApprovalCanceled) {
+		t.Fatal("a wrapped permission cancellation was not recognised")
+	}
+	if errors.Is(errors.New("something else"), agent.ErrPermissionApprovalCanceled) {
+		t.Fatal("an unrelated error was reported as a permission cancellation")
+	}
+}
+
+// A decision that never became an option cannot be selected.
+//
+// PermissionDecisionCancel is the case that exists today: optionKindFor drops
+// it because ACP expresses cancellation through the outcome, yet ZERO
+// enumerates it in AvailableDecisions for shell commands and apply_patch. While
+// the reply was validated against the DECISIONS rather than the sent options, a
+// client could return {"outcome":"selected","optionId":"cancel"} — an id it was
+// never shown — and abort the whole turn.
+//
+// Written over every option-less decision rather than over "cancel", so a
+// future action that optionKindFor declines to render is covered the day it is
+// added rather than the day someone remembers this test.
+func TestADecisionThatIsNotAnOptionCannotBeSelected(t *testing.T) {
+	for _, toolName := range []string{"bash", "apply_patch"} {
+		req := agent.PermissionRequest{
+			ToolName: toolName,
+			AvailableDecisions: []agent.PermissionDecisionAction{
+				agent.PermissionDecisionAllow,
+				agent.PermissionDecisionDeny,
+				agent.PermissionDecisionCancel,
+			},
+		}
+		options := buildPermissionOptions(req)
+
+		for _, action := range req.AvailableDecisions {
+			if actionOffered(string(action), options) {
+				continue // it was sent; selecting it is legitimate
+			}
+			decision := decisionFromOutcome(
+				RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(action)},
+				options,
+			)
+			if decision.Action != agent.PermissionDecisionDeny {
+				t.Fatalf("%s: %q was never sent as an option but was accepted as %q",
+					toolName, action, decision.Action)
+			}
+		}
+	}
+}
+
+// Cancelling is still reachable — through the outcome, which is where ACP puts
+// it. Closing the selected-id route must not close the real one.
+func TestCancellingThroughTheOutcomeStillWorks(t *testing.T) {
+	options := buildPermissionOptions(agent.PermissionRequest{ToolName: "apply_patch"})
+
+	decision := decisionFromOutcome(RequestPermissionOutcome{Outcome: OutcomeCancelled}, options)
+	if decision.Action != agent.PermissionDecisionCancel {
+		t.Fatalf("outcome=cancelled gave %q, want cancel", decision.Action)
+	}
+}

internal/acp/permission.go

@@ -15,14 +15,7 @@ import (
 // PermissionOptions. Only the actions ZERO actually presented (AvailableDecisions)
 // are surfaced; the optionId is the ZERO action string for a clean round-trip.
 func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
-	actions := req.AvailableDecisions
-	if len(actions) == 0 {
-		// Sensible default if ZERO didn't enumerate: allow once / reject.
-		actions = []agent.PermissionDecisionAction{
-			agent.PermissionDecisionAllow,
-			agent.PermissionDecisionDeny,
-		}
-	}
+	actions := offeredDecisions(req)
 	options := make([]PermissionOption, 0, len(actions))
 	for _, action := range actions {
 		kind, name := optionKindFor(action, req.PrefixApprovalEscalates)
@@ -38,6 +31,32 @@ func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
 	return options
 }
 
+// offeredDecisions resolves what this request actually offers the client.
+//
+// ONE RESOLVER, USED BY BOTH SIDES, and that is the entire point of it existing.
+// The fallback below used to live inside buildPermissionOptions while
+// requestPermission validated the client's answer against the raw
+// req.AvailableDecisions. When ZERO did not enumerate — which is every
+// permission event that is not a prompt, and includes a shell call whose sandbox
+// decision came back deny — the two disagreed completely: the client was sent
+// "Allow" and "Reject", and validation was performed against an EMPTY list, so
+// every option the client could possibly click failed closed to deny with
+// "permission option was not offered".
+//
+// The user clicked Allow and ZERO recorded a denial. Nothing surfaced it,
+// because a denial is a legitimate answer — the tool was simply refused and the
+// turn carried on as though the user had rejected it.
+func offeredDecisions(req agent.PermissionRequest) []agent.PermissionDecisionAction {
+	if len(req.AvailableDecisions) > 0 {
+		return req.AvailableDecisions
+	}
+	// Sensible default if ZERO didn't enumerate: allow once / reject.
+	return []agent.PermissionDecisionAction{
+		agent.PermissionDecisionAllow,
+		agent.PermissionDecisionDeny,
+	}
+}
+
 // optionKindFor maps a ZERO decision action to an ACP PermissionOptionKind and a
 // human label. Returns an empty kind for actions that ACP expresses through the
 // outcome rather than an option (cancel).
@@ -49,8 +68,16 @@ func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
 // disclosure cannot be TUI-only.
 func optionKindFor(action agent.PermissionDecisionAction, escalates bool) (kind, name string) {
 	switch action {
-	case agent.PermissionDecisionAllow, agent.PermissionDecisionAllowStrict:
+	case agent.PermissionDecisionAllow:
 		return PermAllowOnce, "Allow"
+	case agent.PermissionDecisionAllowStrict:
+		// SPLIT FROM PLAIN ALLOW because the label is the only thing an ACP
+		// client shows. Both were labelled "Allow", and request_permissions
+		// offers them together — so the panel presented two identical buttons,
+		// and whichever the user picked looked the same while one of them
+		// silently turned on strict auto-review of what was granted. The option
+		// id carries no room for the distinction, so it has to be in the name.
+		return PermAllowOnce, "Allow with strict review"
 	case agent.PermissionDecisionAllowForSession:
 		return PermAllowAlways, "Allow for this session"
 	case agent.PermissionDecisionAllowPrefix:
@@ -70,19 +97,31 @@ func optionKindFor(action agent.PermissionDecisionAction, escalates bool) (kind,
 
 // decisionFromOutcome maps the client's permission outcome back to a ZERO
 // decision. A cancelled outcome cancels the run; a selected option id is the ZERO
-// action verbatim (validated against what was offered); anything unrecognized
-// fails closed to deny.
-func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.PermissionDecisionAction) agent.PermissionDecision {
+// action verbatim (validated against what was actually sent); anything
+// unrecognized fails closed to deny.
+//
+// IT TAKES THE OPTIONS THAT WERE SENT, not the decisions the request offered,
+// and the two are not the same list. Some actions never become options —
+// PermissionDecisionCancel is dropped by optionKindFor because ACP expresses
+// cancellation through the outcome instead — yet it IS enumerated in
+// AvailableDecisions for shell commands and apply_patch, the two commonest
+// prompts. Validating against the decisions therefore accepted
+// {"outcome":"selected","optionId":"cancel"}: an identifier no client was ever
+// shown, aborting the whole turn.
+//
+// Passing the sent options makes "what we offered" and "what we accept" the
+// same slice rather than two things kept in step, so anything optionKindFor
+// declines to render is excluded structurally instead of by name.
+func decisionFromOutcome(outcome RequestPermissionOutcome, offered []PermissionOption) agent.PermissionDecision {
 	switch outcome.Outcome {
 	case OutcomeCancelled:
 		return agent.PermissionDecision{Action: agent.PermissionDecisionCancel, Reason: "client cancelled"}
 	case OutcomeSelected:
-		action := agent.PermissionDecisionAction(outcome.OptionID)
 		// Bind to what was actually offered for THIS call: a client must not be able
 		// to return a broader grant (always_allow / allow_for_session) that wasn't
 		// presented. Anything not offered fails closed to deny.
-		if actionOffered(action, offered) {
-			return agent.PermissionDecision{Action: action}
+		if actionOffered(outcome.OptionID, offered) {
+			return agent.PermissionDecision{Action: agent.PermissionDecisionAction(outcome.OptionID)}
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission option was not offered"}
 	default:
@@ -90,9 +129,9 @@ func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.Permi
 	}
 }
 
-func actionOffered(action agent.PermissionDecisionAction, offered []agent.PermissionDecisionAction) bool {
-	for _, a := range offered {
-		if a == action {
+func actionOffered(optionID string, offered []PermissionOption) bool {
+	for _, option := range offered {
+		if option.OptionID == optionID {
 			return true
 		}
 	}

internal/acp/permission_escalation_test.go

@@ -56,10 +56,11 @@ func TestACPPrefixOptionsStaySilentWithoutEscalation(t *testing.T) {
 // what was offered, so decorating it would fail every prefix approval closed.
 func TestACPDisclosureDoesNotDisturbTheOptionIDRoundTrip(t *testing.T) {
 	request := escalationRequest(true)
-	for _, option := range buildPermissionOptions(request) {
+	options := buildPermissionOptions(request)
+	for _, option := range options {
 		decision := decisionFromOutcome(
 			RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: option.OptionID},
-			request.AvailableDecisions,
+			options,
 		)
 		if string(decision.Action) != option.OptionID {
 			t.Errorf("option id %q round-tripped to %q (%s)", option.OptionID, decision.Action, decision.Reason)

internal/acp/permission_test.go

@@ -39,11 +39,14 @@ func TestBuildPermissionOptionsDefault(t *testing.T) {
 }
 
 func TestDecisionFromOutcome(t *testing.T) {
-	offered := []agent.PermissionDecisionAction{
-		agent.PermissionDecisionAllow,
-		agent.PermissionDecisionAlwaysAllow,
-		agent.PermissionDecisionDeny,
-	}
+	offered := buildPermissionOptions(agent.PermissionRequest{
+		ToolName: "bash",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAlwaysAllow,
+			agent.PermissionDecisionDeny,
+		},
+	})
 	if d := decisionFromOutcome(RequestPermissionOutcome{Outcome: OutcomeCancelled}, offered); d.Action != agent.PermissionDecisionCancel {
 		t.Errorf("cancelled -> %q, want cancel", d.Action)
 	}

internal/agent/command_prefix_test.go

@@ -99,6 +99,43 @@ func TestProposedCommandPrefixRejectsUnsafeRequestedPrefix(t *testing.T) {
 	}
 }
 
+// The launcher denylist is only worth anything if it holds on the path a user
+// actually sees. A model naming its own prefix_rule is that path: before
+// launcher names were normalized, "python3.11" was offered as a one-token
+// grant, and hasStringPrefix then matched every later python3.11 invocation.
+func TestProposedCommandPrefixRejectsLauncherSpellingsInRequestedPrefix(t *testing.T) {
+	for _, testCase := range []struct {
+		command string
+		rule    []any
+	}{
+		{"python3.11 script.py", []any{"python3.11"}},
+		{"python.exe script.py", []any{"python.exe"}},
+		{"node.exe app.js", []any{"node.exe"}},
+		{"cmd /c whoami", []any{"cmd"}},
+		{"busybox sh -c id", []any{"busybox"}},
+		{"uv run main.py", []any{"uv", "run"}},
+		{"python3.11 -c import_os", []any{"python3.11", "-c"}},
+	} {
+		got := proposedCommandPrefix("bash", map[string]any{
+			"command":     testCase.command,
+			"prefix_rule": testCase.rule,
+		})
+		if got != nil {
+			t.Errorf("prefix_rule %v for %q was offered as %#v, want rejected", testCase.rule, testCase.command, got)
+		}
+	}
+}
+
+func TestProposedCommandPrefixStillOffersOrdinaryCommands(t *testing.T) {
+	got := proposedCommandPrefix("bash", map[string]any{
+		"command":     "cargo build --release",
+		"prefix_rule": []any{"cargo", "build"},
+	})
+	if len(got) != 2 || got[0] != "cargo" || got[1] != "build" {
+		t.Fatalf("ordinary prefix rule = %#v, want [cargo build]", got)
+	}
+}
+
 func TestProposedCommandPrefixRejectsUnsafeShellForms(t *testing.T) {
 	cases := []string{
 		"cat < in > out",

internal/agent/compaction.go

@@ -144,11 +144,17 @@ func estimateToolDefTokens(tools []zeroruntime.ToolDefinition) int {
 	for _, tool := range tools {
 		total += ApproxTextTokens(tool.Name)
 		total += ApproxTextTokens(tool.Description)
+		total += ApproxTextTokens(string(tool.Type))
 		if len(tool.Parameters) > 0 {
 			if encoded, err := json.Marshal(tool.Parameters); err == nil {
 				total += ApproxTextTokens(string(encoded))
 			}
 		}
+		if tool.Format != nil {
+			total += ApproxTextTokens(tool.Format.Type)
+			total += ApproxTextTokens(tool.Format.Syntax)
+			total += ApproxTextTokens(tool.Format.Definition)
+		}
 		total += 4 // per-tool overhead
 	}
 	return total

internal/agent/compaction_projection.go

@@ -126,8 +126,11 @@ func capCompactionBrief(brief string) (string, bool) {
 
 func compactionToolCallLine(call zeroruntime.ToolCall) string {
 	detail := ""
+	if call.Freeform && call.Name == "apply_patch" {
+		detail = strings.Join(patchPaths(call.Arguments), ", ")
+	}
 	var arguments map[string]any
-	if json.Unmarshal([]byte(strings.TrimSpace(call.Arguments)), &arguments) == nil {
+	if detail == "" && json.Unmarshal([]byte(strings.TrimSpace(call.Arguments)), &arguments) == nil {
 		for _, key := range []string{"file_path", "path", "url", "query", "pattern", "prompt", "description", "question"} {
 			if value, ok := arguments[key].(string); ok && strings.TrimSpace(value) != "" {
 				detail = value

internal/agent/compaction_projection_test.go

@@ -73,6 +73,7 @@ func TestProjectCompactionInputRetainsAskUserExchangeAndSemanticArguments(t *tes
 		{Role: zeroruntime.MessageRoleTool, ToolCallID: "ask", Content: "Which database should remain?: Postgres only; do not touch MySQL."},
 		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
 			{ID: "patch", Name: "apply_patch", Arguments: `{"patch":"*** Begin Patch\n*** Update File: internal/db.go\n*** End Patch"}`},
+			{ID: "raw-patch", Name: "apply_patch", Freeform: true, Arguments: "*** Begin Patch\n*** Update File: internal/api.go\n*** Add File: internal/config.go\n*** End Patch\n"},
 			{ID: "fetch", Name: "web_fetch", Arguments: `{"url":"https://example.com/spec"}`},
 			{ID: "task", Name: "task", Arguments: `{"prompt":"Inspect the Windows implementation"}`},
 		}},
@@ -80,7 +81,7 @@ func TestProjectCompactionInputRetainsAskUserExchangeAndSemanticArguments(t *tes
 	}
 
 	content := projectCompactionInput(messages).messages[0].Content
-	for _, want := range []string{"Which database should remain?", "Postgres only; do not touch MySQL", "internal/db.go", "https://example.com/spec", "Inspect the Windows implementation"} {
+	for _, want := range []string{"Which database should remain?", "Postgres only; do not touch MySQL", "internal/db.go", "internal/api.go", "internal/config.go", "https://example.com/spec", "Inspect the Windows implementation"} {
 		if !strings.Contains(content, want) {
 			t.Fatalf("projection missing %q: %q", want, content)
 		}

internal/agent/compaction_test.go

@@ -275,6 +275,27 @@ func TestEstimateToolDefTokensCountsDefinitions(t *testing.T) {
 	if estimateToolDefTokens(one) <= 0 {
 		t.Fatal("a tool definition (name + description + schema) should estimate > 0 tokens")
 	}
+	typed := append([]zeroruntime.ToolDefinition(nil), one...)
+	typed[0].Type = zeroruntime.ToolDefinitionFreeform
+	if estimateToolDefTokens(typed) <= estimateToolDefTokens(one) {
+		t.Fatal("a non-empty tool type must increase the estimated request cost")
+	}
+	for _, test := range []struct {
+		name   string
+		format zeroruntime.ToolDefinitionFormat
+	}{
+		{name: "type", format: zeroruntime.ToolDefinitionFormat{Type: "grammar"}},
+		{name: "syntax", format: zeroruntime.ToolDefinitionFormat{Syntax: "lark"}},
+		{name: "definition", format: zeroruntime.ToolDefinitionFormat{Definition: "start: WORD"}},
+	} {
+		t.Run("format "+test.name, func(t *testing.T) {
+			formatted := append([]zeroruntime.ToolDefinition(nil), one...)
+			formatted[0].Format = &test.format
+			if estimateToolDefTokens(formatted) <= estimateToolDefTokens(one) {
+				t.Fatalf("a non-empty format %s must increase the estimated request cost", test.name)
+			}
+		})
+	}
 	two := append(append([]zeroruntime.ToolDefinition{}, one...), zeroruntime.ToolDefinition{
 		Name:        "write_file",
 		Description: "Write contents to a file in the workspace.",

internal/agent/context_planner.go

@@ -102,6 +102,10 @@ func copyToolDefinitions(toolDefs []zeroruntime.ToolDefinition) []zeroruntime.To
 	for index, definition := range toolDefs {
 		copied[index] = definition
 		copied[index].Parameters = copySchemaMap(definition.Parameters)
+		if definition.Format != nil {
+			format := *definition.Format
+			copied[index].Format = &format
+		}
 	}
 	return copied
 }

internal/agent/context_planner_test.go

@@ -15,6 +15,7 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 	}
 	toolDefs := []zeroruntime.ToolDefinition{{
 		Name: "read_file", Description: "Read a file",
+		Format: &zeroruntime.ToolDefinitionFormat{Type: "grammar", Syntax: "lark", Definition: "start: PATH"},
 		Parameters: map[string]any{
 			"type":       "object",
 			"required":   []string{"path"},
@@ -53,10 +54,12 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 	toolDefs[0].Parameters["required"].([]string)[0] = "other"
 	toolDefs[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] = "integer"
 	toolDefs[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] = "array"
+	toolDefs[0].Format.Definition = "changed"
 	if plan.Request.Messages[1].Content != "inspect this" || plan.Request.Messages[1].Images[0].Data[0] != 1 {
 		t.Fatalf("planned messages alias caller state: %#v", plan.Request.Messages[1])
 	}
 	if plan.Request.Tools[0].Name != "read_file" ||
+		plan.Request.Tools[0].Format == nil || plan.Request.Tools[0].Format.Definition != "start: PATH" ||
 		plan.Request.Tools[0].Parameters["required"].([]string)[0] != "path" ||
 		plan.Request.Tools[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] != "string" ||
 		plan.Request.Tools[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] != "object" {

internal/agent/freeform_tool_test.go

@@ -0,0 +1,378 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+type customApplyPatchTool struct {
+	calls int
+}
+
+type fixedPathScope struct {
+	roots []string
+}
+
+func (scope fixedPathScope) Roots() []string {
+	return append([]string(nil), scope.roots...)
+}
+
+func newFixedPathScope(t *testing.T, roots ...string) fixedPathScope {
+	t.Helper()
+	resolved := make([]string, 0, len(roots))
+	for _, root := range roots {
+		absolute, err := filepath.Abs(root)
+		if err != nil {
+			t.Fatalf("Abs(%q): %v", root, err)
+		}
+		if physical, err := filepath.EvalSymlinks(absolute); err == nil {
+			absolute = physical
+		}
+		resolved = append(resolved, absolute)
+	}
+	return fixedPathScope{roots: resolved}
+}
+
+func (*customApplyPatchTool) Name() string        { return "apply_patch" }
+func (*customApplyPatchTool) Description() string { return "custom JSON tool" }
+func (*customApplyPatchTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", Properties: map[string]tools.PropertySchema{"value": {Type: "string"}}}
+}
+func (*customApplyPatchTool) Safety() tools.Safety {
+	return tools.Safety{Permission: tools.PermissionAllow, SideEffect: tools.SideEffectRead}
+}
+func (tool *customApplyPatchTool) Run(context.Context, map[string]any) tools.Result {
+	tool.calls++
+	return tools.Result{Status: tools.StatusOK}
+}
+
+func TestFreeformApplyPatchUsesExistingToolExecutionPath(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, nil))
+	patch := "*** Begin Patch\n*** Add File: hello.txt\n+hello\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID:        "call-1",
+		Name:      "apply_patch",
+		Arguments: patch,
+		Freeform:  true,
+	}, PermissionModeUnsafe, Options{Cwd: root, FileTracker: tools.NewFileTracker()})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusOK {
+		t.Fatalf("freeform apply_patch status = %s: %s", result.Status, result.Output)
+	}
+	content, err := os.ReadFile(filepath.Join(root, "hello.txt"))
+	if err != nil {
+		t.Fatalf("read patched file: %v", err)
+	}
+	if string(content) != "hello\n" {
+		t.Fatalf("patched content = %q, want hello newline", content)
+	}
+}
+
+func TestFreeformApplyPatchSupportsGrantedExtraRoot(t *testing.T) {
+	root := t.TempDir()
+	extra := t.TempDir()
+	scope, err := sandbox.NewScope(root, []string{extra})
+	if err != nil {
+		t.Fatalf("NewScope: %v", err)
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, scope))
+	target := filepath.Join(extra, "hello.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+hello\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+	}, PermissionModeUnsafe, Options{Cwd: root, FileTracker: tools.NewFileTracker()})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusOK {
+		t.Fatalf("freeform extra-root apply_patch status = %s: %s", result.Status, result.Output)
+	}
+	content, err := os.ReadFile(target)
+	if err != nil {
+		t.Fatalf("read extra-root patched file: %v", err)
+	}
+	if string(content) != "hello\n" {
+		t.Fatalf("patched content = %q, want hello newline", content)
+	}
+}
+
+func TestFreeformApplyPatchRejectsMixedWorkspaceAndExtraRoots(t *testing.T) {
+	for _, extraHasMatchingPath := range []bool{false, true} {
+		name := "extra target missing"
+		if extraHasMatchingPath {
+			name = "extra target exists"
+		}
+		t.Run(name, func(t *testing.T) {
+			workspace := t.TempDir()
+			extra := t.TempDir()
+			if err := os.MkdirAll(filepath.Join(workspace, "src"), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			workspaceTarget := filepath.Join(workspace, "src", "a.go")
+			if err := os.WriteFile(workspaceTarget, []byte("old\n"), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			extraTarget := filepath.Join(extra, "src", "a.go")
+			if extraHasMatchingPath {
+				if err := os.MkdirAll(filepath.Dir(extraTarget), 0o755); err != nil {
+					t.Fatal(err)
+				}
+				if err := os.WriteFile(extraTarget, []byte("old\n"), 0o644); err != nil {
+					t.Fatal(err)
+				}
+			}
+			scope, err := sandbox.NewScope(workspace, []string{extra})
+			if err != nil {
+				t.Fatal(err)
+			}
+			registry := tools.NewRegistry()
+			registry.Register(tools.NewScopedApplyPatchTool(workspace, scope))
+			extraCreated := filepath.Join(extra, "b.go")
+			patch := strings.Join([]string{
+				"*** Begin Patch",
+				"*** Update File: src/a.go",
+				"@@",
+				"-old",
+				"+new",
+				"*** Add File: " + extraCreated,
+				"+extra",
+				"*** End Patch",
+			}, "\n")
+
+			result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+				ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+			}, PermissionModeUnsafe, Options{Cwd: workspace, FileTracker: tools.NewFileTracker()})
+			if err != nil {
+				t.Fatalf("executeToolCall: %v", err)
+			}
+			if result.Status != tools.StatusError || !strings.Contains(result.Output, "mixes workspace-relative paths with an extra write root") {
+				t.Fatalf("mixed-root patch result = %s: %s", result.Status, result.Output)
+			}
+			if content, err := os.ReadFile(workspaceTarget); err != nil || string(content) != "old\n" {
+				t.Fatalf("workspace target changed: content=%q err=%v", content, err)
+			}
+			if extraHasMatchingPath {
+				if content, err := os.ReadFile(extraTarget); err != nil || string(content) != "old\n" {
+					t.Fatalf("extra-root target changed: content=%q err=%v", content, err)
+				}
+			}
+			if _, err := os.Stat(extraCreated); !os.IsNotExist(err) {
+				t.Fatalf("mixed-root patch created extra file: %v", err)
+			}
+		})
+	}
+}
+
+func TestFreeformApplyPatchAcceptsOneSemanticRoot(t *testing.T) {
+	for _, test := range []struct {
+		name         string
+		useExtraRoot bool
+	}{
+		{name: "workspace"},
+		{name: "extra root", useExtraRoot: true},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			workspace := t.TempDir()
+			extra := t.TempDir()
+			scope, err := sandbox.NewScope(workspace, []string{extra})
+			if err != nil {
+				t.Fatal(err)
+			}
+			root := workspace
+			if test.useExtraRoot {
+				root = extra
+			}
+			if err := os.MkdirAll(filepath.Join(root, "src"), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			target := filepath.Join(root, "src", "a.go")
+			if err := os.WriteFile(target, []byte("old\n"), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			updatePath := target
+			if !test.useExtraRoot {
+				updatePath = filepath.Join("src", "a.go")
+			}
+			created := filepath.Join(root, "b.go")
+			patch := strings.Join([]string{
+				"*** Begin Patch",
+				"*** Update File: " + updatePath,
+				"@@",
+				"-old",
+				"+new",
+				"*** Add File: " + created,
+				"+created",
+				"*** End Patch",
+			}, "\n")
+			registry := tools.NewRegistry()
+			registry.Register(tools.NewScopedApplyPatchTool(workspace, scope))
+			result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+				ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+			}, PermissionModeUnsafe, Options{Cwd: workspace, FileTracker: tools.NewFileTracker()})
+			if err != nil {
+				t.Fatalf("executeToolCall: %v", err)
+			}
+			if result.Status != tools.StatusOK {
+				t.Fatalf("single-root patch status = %s: %s", result.Status, result.Output)
+			}
+			if content, err := os.ReadFile(target); err != nil || string(content) != "new\n" {
+				t.Fatalf("updated content=%q err=%v", content, err)
+			}
+			if content, err := os.ReadFile(created); err != nil || string(content) != "created\n" {
+				t.Fatalf("created content=%q err=%v", content, err)
+			}
+		})
+	}
+}
+
+func TestFreeformApplyPatchRejectsAbsolutePathOutsideGrantedRoots(t *testing.T) {
+	root := t.TempDir()
+	granted := t.TempDir()
+	outside := t.TempDir()
+	scope := newFixedPathScope(t, root, granted)
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, scope))
+	target := filepath.Join(outside, "blocked.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+blocked\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+	}, PermissionModeUnsafe, Options{Cwd: root})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("ungranted absolute patch status = %s, want error: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(target); !os.IsNotExist(err) {
+		t.Fatalf("ungranted absolute patch created target: %v", err)
+	}
+}
+
+func TestPreparedFreeformPatchRejectsIntermediateSymlinkSwap(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("creating directory symlinks requires privileges on Windows")
+	}
+	root := t.TempDir()
+	extra := t.TempDir()
+	outside := t.TempDir()
+	intermediate := filepath.Join(extra, "nested")
+	if err := os.Mkdir(intermediate, 0o755); err != nil {
+		t.Fatalf("Mkdir: %v", err)
+	}
+	scope := newFixedPathScope(t, root, extra)
+	tool := tools.NewScopedApplyPatchTool(root, scope)
+	target := filepath.Join(intermediate, "blocked.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+blocked\n*** End Patch\n"
+	args, err := tools.PrepareFreeformApplyPatchArguments(tool, patch)
+	if err != nil {
+		t.Fatalf("PrepareFreeformApplyPatchArguments: %v", err)
+	}
+
+	if err := os.Remove(intermediate); err != nil {
+		t.Fatalf("Remove intermediate: %v", err)
+	}
+	if err := os.Symlink(outside, intermediate); err != nil {
+		t.Fatalf("Symlink: %v", err)
+	}
+	result := tool.Run(context.Background(), args)
+	if result.Status != tools.StatusError {
+		t.Fatalf("symlink-swapped patch status = %s, want error: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(filepath.Join(outside, "blocked.txt")); !os.IsNotExist(err) {
+		t.Fatalf("symlink-swapped patch wrote outside root: %v", err)
+	}
+}
+
+func TestUnknownFreeformToolFailsClosed(t *testing.T) {
+	result, err := executeToolCall(context.Background(), tools.NewRegistry(), zeroruntime.ToolCall{
+		ID: "call-1", Name: "unknown", Arguments: "raw", Freeform: true,
+	}, PermissionModeUnsafe, Options{})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("unknown freeform status = %s, want error", result.Status)
+	}
+}
+
+func TestCustomApplyPatchCannotUseFreeformContract(t *testing.T) {
+	custom := &customApplyPatchTool{}
+	registry := tools.NewRegistry()
+	registry.Register(custom)
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: "raw patch", Freeform: true,
+	}, PermissionModeUnsafe, Options{})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError || !strings.Contains(result.Output, "Unsupported freeform tool call") {
+		t.Fatalf("custom freeform apply_patch result = %#v, want unsupported error", result)
+	}
+	if custom.calls != 0 {
+		t.Fatalf("custom apply_patch executed %d times", custom.calls)
+	}
+}
+
+func TestDeniedFreeformApplyPatchDoesNotWrite(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, nil))
+	patch := "*** Begin Patch\n*** Add File: denied.txt\n+must not exist\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID:        "call-1",
+		Name:      "apply_patch",
+		Arguments: patch,
+		Freeform:  true,
+	}, PermissionModeAsk, Options{
+		Cwd: root,
+		OnPermissionRequest: func(context.Context, PermissionRequest) (PermissionDecision, error) {
+			return PermissionDecision{Action: PermissionDecisionDeny, Reason: "test denial"}, nil
+		},
+	})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("denied freeform status = %s, want error", result.Status)
+	}
+	if _, err := os.Stat(filepath.Join(root, "denied.txt")); !os.IsNotExist(err) {
+		t.Fatalf("denied freeform patch changed the workspace: %v", err)
+	}
+}
+
+func TestHistorySafeToolCallsPreservesRawFreeformArguments(t *testing.T) {
+	raw := "*** Begin Patch\n*** Update File: internal/main.go\n@@\n-old\n+new\n*** End Patch\n"
+	safe := historySafeToolCalls([]ToolCall{{
+		ID: "call-1", ProviderCallID: "provider-1", Name: "apply_patch", Arguments: raw, Freeform: true,
+	}})
+	if len(safe) != 1 || safe[0].Arguments != raw || safe[0].ProviderCallID != "provider-1" || !safe[0].Freeform {
+		t.Fatalf("freeform history changed raw call: %#v", safe)
+	}
+}
+
+func TestAbortedToolResultPreservesProviderCallID(t *testing.T) {
+	messages := appendAbortedToolResults(nil, []ToolCall{{
+		ID: "call-1", ProviderCallID: "provider-1", Name: "read_file",
+	}})
+	if len(messages) != 1 || messages[0].ToolCallID != "call-1" || messages[0].ToolCallProviderID != "provider-1" || !messages[0].IsError {
+		t.Fatalf("aborted tool result lost provider identity: %#v", messages)
+	}
+}

internal/agent/guardrails.go

@@ -33,11 +33,11 @@ const (
 	// knows before spending more tool turns.
 	toolOnlyProgressReminderAt = 6
 
-	// planReminderTurn is the turn (1-based) by the end of which a multi-step
-	// task should have called update_plan; if it hasn't, a one-time reminder is
-	// injected. Set to 3 (not 2) so short, legitimate two-step tasks finish
-	// without a spurious planning nag.
-	planReminderTurn = 3
+	// planReminderTurn and planReminderToolThreshold keep bounded implementation
+	// work plan-free while still nudging a genuinely sustained run that has not
+	// established any structured progress. Both thresholds must be met.
+	planReminderTurn          = 7
+	planReminderToolThreshold = 7
 
 	// planToolName is the planning tool the loop watches for by name.
 	planToolName = "update_plan"
@@ -591,13 +591,12 @@ func (state *guardState) planReminder(turn int) string {
 		return planStaleReminder(state.toolCallsSincePlanUpdate)
 	}
 
-	// NOT-CALLED reminder: by the end of planReminderTurn the model should have
-	// called update_plan if it's doing a multi-step task (>=1 other tool call).
-	// One-shot for the whole run.
+	// NOT-CALLED reminder: once both the turn and tool-call thresholds are met,
+	// sustained work should have called update_plan. One-shot for the whole run.
 	if !state.notCalledReminderSent &&
 		!state.planEverCalled &&
 		turn >= planReminderTurn &&
-		state.totalToolCalls >= 1 {
+		state.totalToolCalls >= planReminderToolThreshold {
 		state.notCalledReminderSent = true
 		return planNotCalledReminder()
 	}

internal/agent/guardrails_test.go

@@ -257,19 +257,16 @@ func TestRunDoesNotCountDroppedToolCallTurnsAsEmpty(t *testing.T) {
 }
 
 func TestRunInjectsPlanNotCalledReminderForMultiStepTask(t *testing.T) {
+	const expectedThreshold = 7
+	if planReminderToolThreshold != expectedThreshold {
+		t.Fatalf("plan reminder tool threshold = %d, want %d", planReminderToolThreshold, expectedThreshold)
+	}
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
 	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
-	provider := &mockProvider{
-		turns: [][]zeroruntime.StreamEvent{
-			toolTurn("call-1", "read_file", `{"path":"notes.txt"}`), // turn 1: other tool call
-			toolTurn("call-2", "read_file", `{"path":"notes.txt"}`), // turn 2: still no update_plan
-			toolTurn("call-3", "read_file", `{"path":"notes.txt"}`), // turn 3: reminder fires here
-			textTurn("done"),
-		},
-	}
+	provider := &mockProvider{turns: append(repeatedReadTurns(expectedThreshold), textTurn("done"))}
 
 	result, err := Run(context.Background(), "go", provider, Options{
 		Registry: registry,
@@ -287,6 +284,44 @@ func TestRunInjectsPlanNotCalledReminderForMultiStepTask(t *testing.T) {
 	}
 }
 
+func TestRunDoesNotInjectPlanReminderForBoundedToolSequence(t *testing.T) {
+	const expectedThreshold = 7
+	root := t.TempDir()
+	writeAgentTestFile(t, root+"/notes.txt", "alpha")
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
+
+	provider := &mockProvider{turns: append(repeatedReadTurns(expectedThreshold-1), textTurn("done"))}
+	result, err := Run(context.Background(), "go", provider, Options{Registry: registry, MaxTurns: 12})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("expected final answer, got %q", result.FinalAnswer)
+	}
+	if count := countUserMessagesContaining(result.Messages, planNotCalledReminderMarker); count != 0 {
+		t.Fatalf("bounded tool sequence must not draw a plan reminder, got %d", count)
+	}
+}
+
+func TestPlanReminderWaitsForTurnThresholdAfterEnoughToolCalls(t *testing.T) {
+	state := guardState{totalToolCalls: 7}
+	if got := state.planReminder(6); got != "" {
+		t.Fatalf("planReminder(6) = %q, want no reminder before turn threshold", got)
+	}
+	if state.notCalledReminderSent {
+		t.Fatal("early turn marked the not-called reminder as sent")
+	}
+}
+
+func repeatedReadTurns(count int) [][]zeroruntime.StreamEvent {
+	turns := make([][]zeroruntime.StreamEvent, 0, count)
+	for range count {
+		turns = append(turns, toolTurn("call", "read_file", `{"path":"notes.txt"}`))
+	}
+	return turns
+}
+
 func TestRunDoesNotInjectPlanReminderForTrivialTask(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")

internal/agent/loop.go

@@ -40,7 +40,15 @@ const (
 	toolResultControlSpecReview = "spec_review_required"
 )
 
-var errPermissionApprovalCanceled = errors.New("permission approval cancelled")
+// ErrPermissionApprovalCanceled ends a run because the caller cancelled a
+// permission prompt, rather than because anything failed.
+//
+// EXPORTED FOR THE SURFACES. A cancel is a normal ending and each surface says
+// so its own way — the TUI unwinds, and ACP has a "cancelled" stop reason for
+// exactly this. While it was unexported ACP could not distinguish it from a
+// fault and reported it as an internal error, so declining a tool looked like
+// the agent crashing.
+var ErrPermissionApprovalCanceled = errors.New("permission approval cancelled")
 
 // isImageRejectionError reports whether err is a provider 400 that rejects
 // image/multimodal content. This is checked BEFORE the compaction-retry path
@@ -133,6 +141,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 			options.Trace.Counter(trace.CounterInputTokens, int64(usage.InputTokens))
 			options.Trace.Counter(trace.CounterCachedInputTokens, int64(usage.CachedInputTokens))
+			options.Trace.Counter(trace.CounterCacheWriteTokens, int64(usage.CacheWriteTokens))
 			options.Trace.Counter(trace.CounterOutputTokens, int64(usage.OutputTokens))
 		}
 	}
@@ -689,11 +698,12 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				turnRequestedModel = toolResult.RequestedModel
 			}
 			messages = append(messages, zeroruntime.Message{
-				Role:         zeroruntime.MessageRoleTool,
-				Content:      toolResult.ModelOutput(),
-				ToolCallID:   toolResult.ToolCallID,
-				IsError:      toolResult.Status == tools.StatusError,
-				ChangedFiles: append([]string(nil), toolResult.ChangedFiles...),
+				Role:               zeroruntime.MessageRoleTool,
+				Content:            toolResult.ModelOutput(),
+				ToolCallID:         toolResult.ToolCallID,
+				ToolCallProviderID: call.ProviderCallID,
+				IsError:            toolResult.Status == tools.StatusError,
+				ChangedFiles:       append([]string(nil), toolResult.ChangedFiles...),
 			})
 			// Images ride a following USER message rather than the tool result
 			// above. Every provider drops images on a tool-role message —
@@ -1027,6 +1037,9 @@ func historySafeToolCalls(calls []ToolCall) []ToolCall {
 	safe := make([]ToolCall, len(calls))
 	for i, call := range calls {
 		safe[i] = call
+		if call.Freeform {
+			continue
+		}
 		args := strings.TrimSpace(call.Arguments)
 		switch {
 		case args == "":
@@ -1110,8 +1123,28 @@ func decodeToolArguments(arguments string, v any) error {
 }
 
 func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCall, permissionMode PermissionMode, options Options) (ToolResult, error) {
+	tool, toolFound := registry.Get(call.Name)
 	args := map[string]any{}
-	if call.Arguments != "" {
+	if call.Freeform {
+		if !toolFound || !tools.IsBuiltInApplyPatch(tool) {
+			return ToolResult{
+				ToolCallID: call.ID,
+				Name:       call.Name,
+				Status:     tools.StatusError,
+				Output:     "Error: Unsupported freeform tool call for " + call.Name + ".",
+			}, nil
+		}
+		prepared, err := tools.PrepareFreeformApplyPatchArguments(tool, call.Arguments)
+		if err != nil {
+			return ToolResult{
+				ToolCallID: call.ID,
+				Name:       call.Name,
+				Status:     tools.StatusError,
+				Output:     "Error: Invalid freeform apply_patch input: " + err.Error(),
+			}, nil
+		}
+		args = prepared
+	} else if call.Arguments != "" {
 		if err := decodeToolArguments(call.Arguments, &args); err != nil {
 			return ToolResult{
 				ToolCallID: call.ID,
@@ -1139,7 +1172,6 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			DenialReason: DenialFiltered,
 		}, nil
 	}
-	tool, toolFound := registry.Get(call.Name)
 	if (permissionMode == PermissionModeSpecDraft || permissionMode == PermissionModePlan) && toolFound && !ToolAdvertised(tool, permissionMode) {
 		modeName := string(permissionMode)
 		return ToolResult{
@@ -1364,7 +1396,7 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			}
 		case PermissionDecisionCancel:
 			emitCanceledPermission(options, call, requestEvent, decisionReason)
-			return canceledPermissionResult(call, decisionReason, requestEvent), fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+			return canceledPermissionResult(call, decisionReason, requestEvent), fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 		default:
 			emitDeniedPermission(options, call, requestEvent, decisionReason)
 			return deniedPermissionResult(call, decisionReason, requestEvent), nil
@@ -1548,7 +1580,7 @@ func maybeRetryUnsandboxedAfterSandboxRestriction(ctx context.Context, registry
 	case PermissionDecisionCancel:
 		emitCanceledPermission(options, call, requestEvent, reason)
 		canceled := canceledPermissionResult(call, reason, requestEvent)
-		return result, &canceled, true, decision.Action, reason, nil, fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+		return result, &canceled, true, decision.Action, reason, nil, fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 	default:
 		emitDeniedPermission(options, call, requestEvent, reason)
 		denied := deniedPermissionResult(call, reason, requestEvent)
@@ -1596,7 +1628,7 @@ func maybeRetryWithNetworkAfterSandboxDenial(ctx context.Context, registry *tool
 	case PermissionDecisionCancel:
 		emitCanceledPermission(options, call, requestEvent, reason)
 		canceled := canceledPermissionResult(call, reason, requestEvent)
-		return result, &canceled, true, decision.Action, reason, fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+		return result, &canceled, true, decision.Action, reason, fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 	default:
 		emitDeniedPermission(options, call, requestEvent, reason)
 		denied := deniedPermissionResult(call, reason, requestEvent)
@@ -3202,11 +3234,15 @@ func cachedRuntimeToolDefinition(defCache map[string]zeroruntime.ToolDefinition,
 // runtimeToolDefinition renders a tool's advertised definition (name, description,
 // JSON-schema parameters) as sent to the provider.
 func runtimeToolDefinition(tool tools.Tool) zeroruntime.ToolDefinition {
-	return zeroruntime.ToolDefinition{
+	definition := zeroruntime.ToolDefinition{
 		Name:        tool.Name(),
 		Description: tool.Description(),
 		Parameters:  schemaToRuntimeMap(tool.Parameters()),
 	}
+	if tools.IsBuiltInApplyPatch(tool) {
+		definition.Type = zeroruntime.ToolDefinitionFreeform
+	}
+	return definition
 }
 
 func ToolVisible(tool tools.Tool, permissionMode PermissionMode, enabledTools []string, disabledTools []string) bool {
@@ -3361,10 +3397,11 @@ func loadedToolsFromResult(meta map[string]string) []string {
 func appendAbortedToolResults(messages []Message, remaining []ToolCall) []Message {
 	for _, call := range remaining {
 		messages = append(messages, zeroruntime.Message{
-			Role:       zeroruntime.MessageRoleTool,
-			Content:    abortedToolResultNotice,
-			ToolCallID: call.ID,
-			IsError:    true,
+			Role:               zeroruntime.MessageRoleTool,
+			Content:            abortedToolResultNotice,
+			ToolCallID:         call.ID,
+			ToolCallProviderID: call.ProviderCallID,
+			IsError:            true,
 		})
 	}
 	return messages

internal/agent/loop_test.go

@@ -1853,7 +1853,7 @@ func TestRunAbortsWhenPermissionRequestCanceled(t *testing.T) {
 		},
 	})
 
-	if !errors.Is(err, errPermissionApprovalCanceled) {
+	if !errors.Is(err, ErrPermissionApprovalCanceled) {
 		t.Fatalf("expected permission approval cancel error, got %v", err)
 	}
 	if result.FinalAnswer != "" {
@@ -4096,7 +4096,7 @@ func TestRunDoesNotFlagCleanToolOutput(t *testing.T) {
 // caller's callback AND stamps token counters, and the run completes.
 func TestRunTracingWrapperStampsUsage(t *testing.T) {
 	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
-		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 100, CachedInputTokens: 20, OutputTokens: 40}},
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{PromptTokens: 100, CachedInputTokens: 20, CacheWriteTokens: 10, CompletionTokens: 40}},
 		{Type: zeroruntime.StreamEventText, Content: "done"},
 		{Type: zeroruntime.StreamEventDone},
 	}}}
@@ -4135,6 +4135,9 @@ func TestRunTracingWrapperStampsUsage(t *testing.T) {
 	if got := tr.Counter(trace.CounterCachedInputTokens); got != 20 {
 		t.Fatalf("cached input token counter = %d, want 20", got)
 	}
+	if got := tr.Counter(trace.CounterCacheWriteTokens); got != 10 {
+		t.Fatalf("cache-write token counter = %d, want 10", got)
+	}
 	if got := tr.Counter(trace.CounterOutputTokens); got != 40 {
 		t.Fatalf("output token counter = %d, want 40", got)
 	}
@@ -4340,3 +4343,95 @@ func TestPlanModeHonorsBeforeToolVeto(t *testing.T) {
 		t.Fatalf("expected tool result to mention the beforeTool veto, got %q", combined)
 	}
 }
+
+// A cancelled permission on either sandbox-retry path must abort the run with a
+// recognisable cancellation, and must not run the escalated retry.
+//
+// BOTH PATHS, because both wrap ErrPermissionApprovalCanceled and each is
+// reached by a different kind of denial: one by an unrestricted platform denial
+// asking for approval, the other by an external-network denial. Surfaces map
+// this sentinel onto their own idea of a cancel — ACP returns stopReason
+// "cancelled" for it — so a path that produced a bare error instead would show
+// the user a crash where they had simply declined.
+func TestCancellingASandboxRetryAbortsWithoutRetrying(t *testing.T) {
+	denial := func(capability execution.CapabilityKind, scope string) tools.Result {
+		return tools.Result{
+			Status: tools.StatusError,
+			ExecutionOutcome: &execution.Outcome{
+				State: execution.StateDenied,
+				Kind:  execution.OutcomeEnforcementDenied,
+				Denial: &execution.Denial{
+					Capability:  execution.Capability{Kind: capability, Scope: scope},
+					Source:      execution.DenialSourcePlatformSandbox,
+					Reason:      "denied by the sandbox",
+					Recoverable: true,
+					NextAction:  execution.DenialNextActionRequestApproval,
+				},
+			},
+		}
+	}
+
+	for _, testCase := range []struct {
+		name   string
+		result tools.Result
+		run    func(context.Context, *tools.Registry, ToolCall, tools.Tool, map[string]any, tools.Result, Options) error
+	}{
+		{
+			name:   "unsandboxed retry",
+			result: denial(execution.CapabilityUnrestricted, "host"),
+			run: func(ctx context.Context, registry *tools.Registry, call ToolCall, tool tools.Tool, args map[string]any, result tools.Result, options Options) error {
+				_, _, _, _, _, _, abortErr := maybeRetryUnsandboxedAfterSandboxRestriction(
+					ctx, registry, call, tool, args, result, PermissionModeAsk, options, nil)
+				return abortErr
+			},
+		},
+		{
+			name:   "network retry",
+			result: denial(execution.CapabilityExternalNetwork, ""),
+			run: func(ctx context.Context, registry *tools.Registry, call ToolCall, tool tools.Tool, args map[string]any, result tools.Result, options Options) error {
+				_, _, _, _, _, abortErr := maybeRetryWithNetworkAfterSandboxDenial(
+					ctx, registry, call, tool, args, result, PermissionModeAsk, options, nil)
+				return abortErr
+			},
+		},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			retryTool := &sandboxNamespaceLimitedRetryTool{}
+			registry := tools.NewRegistry()
+			registry.Register(retryTool)
+
+			asked := 0
+			options := Options{
+				Registry:       registry,
+				PermissionMode: PermissionModeAsk,
+				Autonomy:       "medium",
+				Sandbox: sandbox.NewEngine(sandbox.EngineOptions{
+					WorkspaceRoot: t.TempDir(),
+					Policy:        sandbox.DefaultPolicy(),
+				}),
+				OnPermissionRequest: func(context.Context, PermissionRequest) (PermissionDecision, error) {
+					asked++
+					return PermissionDecision{Action: PermissionDecisionCancel, Reason: "client cancelled"}, nil
+				},
+			}
+
+			call := ToolCall{ID: "call-1", Name: "bash", Arguments: `{"command":"curl example.com"}`}
+			args := map[string]any{"command": "curl example.com"}
+
+			abortErr := testCase.run(context.Background(), registry, call, retryTool, args, testCase.result, options)
+
+			if asked != 1 {
+				t.Fatalf("permission was requested %d times, want exactly 1", asked)
+			}
+			if !errors.Is(abortErr, ErrPermissionApprovalCanceled) {
+				t.Fatalf("abort error = %v, want it to match ErrPermissionApprovalCanceled", abortErr)
+			}
+			// The escalated retry is the thing the permission was gating. A
+			// cancelled prompt that still ran it would have run the command the
+			// user just declined.
+			if len(retryTool.calls) != 0 {
+				t.Fatalf("the escalated retry ran %d times after a cancel: %+v", len(retryTool.calls), retryTool.calls)
+			}
+		})
+	}
+}

internal/agent/partition_cache_test.go

@@ -31,6 +31,19 @@ func (t countingSchemaTool) Run(_ context.Context, _ map[string]any) tools.Resul
 	return tools.Result{Status: tools.StatusOK}
 }
 
+func TestRuntimeToolDefinitionMarksOnlyBuiltInApplyPatchFreeform(t *testing.T) {
+	builtIn := runtimeToolDefinition(tools.NewScopedApplyPatchTool(t.TempDir(), nil))
+	if builtIn.Type != zeroruntime.ToolDefinitionFreeform {
+		t.Fatalf("built-in apply_patch type = %q, want freeform", builtIn.Type)
+	}
+
+	calls := 0
+	custom := runtimeToolDefinition(countingSchemaTool{name: "apply_patch", calls: &calls})
+	if custom.Type != "" {
+		t.Fatalf("custom apply_patch type = %q, want legacy function shape", custom.Type)
+	}
+}
+
 // The cache renders a tool's schema once and reuses it across calls, and its
 // output is identical to the uncached path.
 func TestPartitionToolsCacheRendersOnceAndMatchesUncached(t *testing.T) {

internal/agent/prompt_budget_test.go

@@ -68,7 +68,7 @@ func TestEagerToolSchemaTokenBudget(t *testing.T) {
 	}
 }
 
-func TestAgentAdvertisesPatchInsteadOfAmbiguousStringReplacement(t *testing.T) {
+func TestAgentAdvertisesBothEditTools(t *testing.T) {
 	registry := tools.NewRegistry()
 	for _, tool := range tools.CoreToolsScoped(t.TempDir(), nil) {
 		registry.Register(tool)
@@ -79,9 +79,9 @@ func TestAgentAdvertisesPatchInsteadOfAmbiguousStringReplacement(t *testing.T) {
 		names[definition.Name] = true
 	}
 	if !names["apply_patch"] {
-		t.Fatal("agent must retain apply_patch for existing-file changes")
+		t.Fatal("agent must retain apply_patch for multi-hunk changes")
 	}
-	if names["edit_file"] {
-		t.Fatal("agent must not receive the ambiguous string-replacement tool")
+	if !names["edit_file"] {
+		t.Fatal("agent must receive edit_file for targeted exact replacements")
 	}
 }

internal/agent/prompt_fingerprint.go

@@ -58,12 +58,18 @@ func toolSubstrings(exposed []zeroruntime.ToolDefinition) (toolsSubstr, schemaSu
 	for _, def := range exposed {
 		writeLengthPrefixed(&toolsSB, def.Name)
 		writeLengthPrefixed(&toolsSB, def.Description)
+		writeLengthPrefixed(&toolsSB, string(def.Type))
 		// Parameters is rendered to a canonical JSON string per tool. The
 		// schema-render cache in internal/agent (used by partitionToolsCached)
 		// guarantees the same render is byte-identical across turns for the
 		// same tool, so this substring is a stable hash input.
 		writeLengthPrefixed(&schemaSB, def.Name)
 		writeLengthPrefixed(&schemaSB, canonicalSchemaString(def.Parameters))
+		if def.Format != nil {
+			writeLengthPrefixed(&schemaSB, def.Format.Type)
+			writeLengthPrefixed(&schemaSB, def.Format.Syntax)
+			writeLengthPrefixed(&schemaSB, def.Format.Definition)
+		}
 	}
 	return toolsSB.String(), schemaSB.String()
 }

internal/agent/prompt_fingerprint_test.go

@@ -73,6 +73,38 @@ func TestComputePrefixFingerprintSchemaChangeChangesSchemaHash(t *testing.T) {
 	}
 }
 
+func TestComputePrefixFingerprintIncludesToolTypeAndFormat(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	base := zeroruntime.ToolDefinition{Name: "apply_patch", Description: "Patch files"}
+	baseFingerprint := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{base})
+
+	typed := base
+	typed.Type = zeroruntime.ToolDefinitionFreeform
+	if got := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{typed}); got.ToolsHash == baseFingerprint.ToolsHash {
+		t.Fatal("ToolsHash must change when tool Type changes")
+	}
+
+	formatCases := []struct {
+		name   string
+		format zeroruntime.ToolDefinitionFormat
+	}{
+		{name: "type", format: zeroruntime.ToolDefinitionFormat{Type: "grammar"}},
+		{name: "syntax", format: zeroruntime.ToolDefinitionFormat{Syntax: "lark"}},
+		{name: "definition", format: zeroruntime.ToolDefinitionFormat{Definition: "start: PATCH"}},
+		{name: "empty", format: zeroruntime.ToolDefinitionFormat{}},
+	}
+	for _, test := range formatCases {
+		t.Run(test.name, func(t *testing.T) {
+			withFormat := base
+			withFormat.Format = &test.format
+			got := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{withFormat})
+			if got.SchemaHash == baseFingerprint.SchemaHash {
+				t.Fatalf("SchemaHash must change for %s format input", test.name)
+			}
+		})
+	}
+}
+
 // TestComputePrefixFingerprintSystemPromptChangeChangesBaseHash asserts the
 // most common drift path: a system-prompt edit (e.g. an agent.md update, a
 // model addendum change) must move the base-instructions hash, which moves

internal/agent/system_prompt.md

@@ -30,13 +30,10 @@ work.
    read-before-edit discipline: inspect the target file and nearby callers,
    tests, or config before you modify behavior. Never edit a file you have not
    read.
-2. **Plan.** Use update_plan for implementation or investigation that has at
-   least three meaningful dependent steps, then keep it current at milestones.
-   Mark completed work and the next in-progress step together when practical;
-   do not spend separate turns updating it after every file or command. Keep at
-   most one item in_progress. Skip plans for simple lookups, explanations, code
-   navigation, and other short tasks. Never create a plan after the work is
-   already complete merely to describe what you did.
+2. **Plan.** Use update_plan for multi-component, sequencing-uncertain, or long
+   work. Skip it for bounded changes in one component. Update only when the
+   overall phase changes, keep at most one item in_progress, and never create a
+   plan after the work is complete.
 3. **Implement.** Make focused changes that match the surrounding code's style,
    naming, and conventions. Prefer the smallest change that fully solves the
    problem. Avoid broad refactors, unrelated rewrites, dependency churn, and
@@ -52,38 +49,48 @@ work.
 
 - Choose the narrowest tool that safely accomplishes the step. Prefer native
   file tools - read_file, read_minified_file, list_directory, glob, grep,
-  write_file, apply_patch - over shelling out to
+  edit_file, apply_patch, write_file - over shelling out to
   cat/sed/awk/python for file operations.
   They are safer, reviewable, and produce clean diffs.
-- Prefer read_minified_file when initially exploring source code; it preserves
-  code while removing comments and redundant whitespace. Use read_file when
-  exact text, comments, or line numbers are needed.
+- Use read_minified_file only to explore large or unfamiliar source. Use
+  read_file for small or likely edit targets and exact text, comments, or line
+  numbers. Do not read both unless a new need for exact content appears.
 - Keep edits focused and reviewable. A single patch may update several related
   files when they form one coherent change; do not hide unrelated edits in a
   bulk shell or script rewrite.
-- For edits to existing files, prefer apply_patch with minimal, targeted hunks
-  and enough unchanged context to identify the intended location. Match the
-  existing indentation, imports, and idioms. Match the file's comment density:
-  do not add explanatory comments unless the user asks or the code is already
+- For edits to existing files, use edit_file for a targeted change (old_string
+  must match the file exactly and be unique, so include a few surrounding
+  lines) and apply_patch when one coherent change spans several hunks or
+  files. Use write_file only to create a file or when most of it changes; do
+  not rewrite a whole file to change a few lines. Match the existing
+  indentation, imports, and idioms. Match the file's comment density: do not
+  add explanatory comments unless the user asks or the code is already
   comment-dense.
+- A successful edit result already confirms the change; do not re-read a file
+  just to verify an edit that succeeded. If an edit fails, read the error, fix
+  the old_string or hunk, and retry the same tool rather than switching to a
+  full rewrite.
 - Solve the problem as posed, not a more general version of it. Add no
   speculative abstraction, configurability, or handling for cases that cannot
   occur, and nothing the user did not ask for. A small diff can still be
-  over-built; if a 200-line solution could be 50, rewrite it.
+  over-built; if a 200-line solution could be 50, rewrite it. Once the required
+  behavior passes validation, stop instead of adding optional cleanup.
 - Preserve behavior you were not asked to change. Do not delete or rewrite code
   you did not author unless the task requires it; if you must, say so.
 
 ## Testing gate (mandatory)
 
-- After any change to code, verify after edits by running the project's
-  validators before you summarize or commit: tests, type-checks, linters, and/or
-  the build, as appropriate. Scope them to the change while iterating; reserve
-  full-suite runs for milestones.
-- If you are unsure which validators apply, search the repo (Makefile, package
-  manifests, CI config) to find them.
+- After code changes, run the project's documented appropriate validators.
+  Scope iteration checks, combine compatible commands when diagnostics stay
+  useful, and run the final set once after the last edit. Rerun only after a
+  change or failure.
+- If unsure which entry points apply, inspect the Makefile, manifests, or CI;
+  do not invent a parallel flow.
 - Never claim a task is done, and never commit, while validators are failing. If
   they fail, fix the cause and rerun; do not paper over it. If you could not run
   a validator, say so explicitly rather than implying success.
+- A successful native patch confirms the edit. Reread only when the next change
+  needs exact content or the result was ambiguous.
 
 ## Tool use
 
@@ -105,7 +112,7 @@ work.
   you need to clean up a running foreground command yourself, use write_stdin.
 - write_stdin's session_id is only ever an id returned by a still-running
   exec_command; never guess or probe ids. If you have no such session, start one
-  with exec_command, or use write_file/apply_patch for file changes.
+  with exec_command, or use edit_file/apply_patch/write_file for file changes.
 - write_stdin with empty input polls an existing exec_command session, and
   `\u0003` interrupts it. Sending other stdin bytes may require approval because
   it can drive the running process beyond the original command. Non-tty sessions
@@ -141,6 +148,7 @@ work.
 - Default to concise, skimmable output. Lead with the answer or the result.
 - Use GitHub-flavored Markdown: headings to structure longer replies, fenced
   code blocks for code, and `inline code` for file paths, commands, symbols, and
-  short snippets. Reference code as `file:line` so it is clickable.
+  short snippets. Reference code with file paths and include line numbers when
+  already known; do not look them up solely for the final summary.
 - Report outcomes faithfully: if tests failed, show it; if a step was skipped,
   say so; when something is done and verified, state it plainly without hedging.

internal/agent/system_prompt_models.go

@@ -60,14 +60,17 @@ const openAIPromptAddendum = `<model_guidance>
   longer answers, fenced code blocks for code, and ` + "`inline code`" + ` for paths,
   commands, and symbols.
 - Strongly prefer the native file tools (read_file, list_directory, grep, glob,
-  write_file, apply_patch) over shelling out to cat/sed/awk/python for
-  file work. Make one tool call per file; do not batch file writes into a script.
+  edit_file, apply_patch, write_file) over shelling out to cat/sed/awk/python for
+  file work. Use them directly; never run apply_patch through exec_command.
+  read_file, edit_file and write_file take one file per call, while one coherent
+  apply_patch may span several files. Keep independent native calls in the same
+  turn, and do not batch file writes into a script.
 - Persist until the task is fully handled this turn: gather context, implement,
   run the validators, and report — do not stop at a partial result.
 </model_guidance>`
 
 const geminiPromptAddendum = `<model_guidance>
-- Prefer the dedicated tools (read_file, grep, glob, apply_patch) over
+- Prefer the dedicated tools (read_file, grep, glob, edit_file, apply_patch) over
   equivalent shell commands; they are safer and produce cleaner diffs.
 - Be concise and concrete. When you run a shell command with side effects, state
   in one short clause why it is needed.

internal/agent/system_prompt_models_test.go

@@ -29,6 +29,9 @@ func TestBuildSystemPromptAppendsModelAddendum(t *testing.T) {
 	if got := buildSystemPrompt(Options{Model: "gpt-5"}); !strings.Contains(got, openAIPromptAddendum) {
 		t.Fatalf("expected the OpenAI addendum in the gpt-5 prompt")
 	}
+	if got := strings.Join(strings.Fields(strings.ToLower(buildSystemPrompt(Options{Model: "gpt-5"}))), " "); !strings.Contains(got, "never run apply_patch through exec_command") {
+		t.Fatalf("OpenAI prompt must route edits through the native patch tool:\n%s", got)
+	}
 	// Claude is aligned with the core prompt and gets no family addendum now that
 	// comment discipline is universal; it must not pick up another family's block.
 	claude := buildSystemPrompt(Options{Model: "claude-opus-4-6"})

internal/agent/system_prompt_test.go

@@ -29,14 +29,17 @@ func TestMain(m *testing.M) {
 }
 
 func TestCoreSystemPromptIncludesCodingQualityRules(t *testing.T) {
-	prompt := strings.ToLower(buildSystemPrompt(Options{}))
+	prompt := strings.Join(strings.Fields(strings.ToLower(buildSystemPrompt(Options{}))), " ")
+	if strings.Contains(prompt, "clickable file paths") {
+		t.Fatalf("system prompt must not claim inline-code file paths are clickable: %s", prompt)
+	}
 
 	for _, want := range []string{
 		"read-before-edit",
 		"inspect the target file",
 		"plan then act",
 		"choose the narrowest tool",
-		"for edits to existing files, prefer apply_patch",
+		"for edits to existing files, use edit_file",
 		"verify after edits",
 		"honor the active permission mode",
 		"avoid broad refactors",
@@ -44,6 +47,13 @@ func TestCoreSystemPromptIncludesCodingQualityRules(t *testing.T) {
 		"do not recognize",
 		"scaled to the work",
 		"comment density",
+		"skip it for bounded changes in one component",
+		"do not read both unless a new need for exact content appears",
+		"combine compatible commands when diagnostics stay useful",
+		"do not invent a parallel flow",
+		"a successful native patch confirms the edit",
+		"once the required behavior passes validation, stop",
+		"do not look them up solely for the final summary",
 	} {
 		if !strings.Contains(prompt, want) {
 			t.Fatalf("expected core system prompt to include %q, got:\n%s", want, buildSystemPrompt(Options{}))

internal/cli/app.go

@@ -839,24 +839,26 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		mcpTokenStore = nil
 		err = nil
 	}
+	criticalMCPConfig, optionalMCPConfig := splitMCPStartupConfig(mcpConfig)
 	mcpRuntime := mcpToolRuntime(noopMCPRuntime{})
-	if len(mcpConfig.Servers) > 0 {
-		mcpRuntime, err = deps.registerMCPTools(context.Background(), registry, mcpConfig, mcp.RegisterOptions{
+	if len(criticalMCPConfig.Servers) > 0 {
+		runtime, registerErr := deps.registerMCPTools(context.Background(), registry, criticalMCPConfig, mcp.RegisterOptions{
 			PermissionStore: mcpPermissionStore,
 			Autonomy:        mcp.AutonomyLow,
 			Execution:       executionRunner,
 			WorkspaceRoot:   workspaceRoot,
 		})
-	}
-	if err != nil {
-		closeMCPRuntime(stderr, mcpRuntime)
-		return writeAppError(stderr, err.Error(), 1)
+		if registerErr != nil {
+			closeMCPRuntime(stderr, runtime)
+			return writeAppError(stderr, redaction.ErrorMessage(registerErr, redaction.Options{}), 1)
+		}
+		mcpRuntime = runtime
 	}
 	defer closeMCPRuntime(stderr, mcpRuntime)
 	// A server that could not be reached or validated is skipped, not fatal (one
 	// bad MCP server must not abort startup) — surface each so a missing tool set is
 	// explained rather than silently absent. A built-in default the user never
-	// configured (e.g. keyless Firecrawl with no credentials) is the exception: it
+	// configured (e.g. keyless Exa with no credentials) is the exception: it
 	// was never asked for, so its failure is not worth a startup warning — only
 	// servers the user actually configured warn on failure (issue #552).
 	for _, skipped := range mcpRuntime.Skipped() {
@@ -909,11 +911,35 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	}
 	// Activate deferred MCP-tool loading for the interactive run only when the
 	// VISIBLE deferred-eligible count meets the resolved threshold, matching exec.
-	// The registry is complete (core + specialist + MCP + plugins) here, so the
-	// count is accurate; below threshold this is a no-op and the surface is
-	// unchanged. The interactive surface applies no operator tool filters, so
-	// enabled/disabled are nil — matching the AgentOptions below.
+	// This first pass covers every prompt-critical tool. Optional built-in MCP
+	// defaults re-run the same gate after publishing their atomic batch, so a
+	// later catalog generation cannot miss its loader. The interactive surface
+	// applies no operator tool filters, so enabled/disabled are nil — matching the
+	// AgentOptions below.
 	registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, permissionMode, nil, nil)
+	var optionalMCPRuntime *optionalMCPStartup
+	var awaitToolReadiness func(context.Context)
+	if len(optionalMCPConfig.Servers) > 0 {
+		optionalMCPRuntime = startOptionalMCP(
+			context.Background(),
+			registry,
+			optionalMCPConfig,
+			mcp.RegisterOptions{
+				PermissionStore: mcpPermissionStore,
+				Autonomy:        mcp.AutonomyLow,
+				Execution:       executionRunner,
+				WorkspaceRoot:   workspaceRoot,
+			},
+			deps.registerMCPTools,
+			func() {
+				registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, permissionMode, nil, nil)
+			},
+		)
+		defer closeMCPRuntime(stderr, optionalMCPRuntime)
+		awaitToolReadiness = func(ctx context.Context) {
+			optionalMCPRuntime.Await(ctx, optionalMCPPromptGrace)
+		}
+	}
 	lastKnownMCPConfig := mcpConfig
 	fileTracker := tools.NewFileTracker()
 	var scratchBaseline scratchFileBaseline
@@ -950,8 +976,17 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		RecapsEnabled:        resolved.Preferences.RecapsEnabled(),
 		Provider:             provider,
 		NewProvider:          deps.newProvider,
-		ProbeProviderHealth:  deps.probeProviderHealth,
-		UserAgent:            userAgent(),
+		NewTurnSessionProvider: func(profile config.ProviderProfile, provider zeroruntime.Provider) zeroruntime.TurnSessionProvider {
+			if provider == nil {
+				return nil
+			}
+			if optimized, ok := providers.OptimizedTurnSessions(profile, provider, providers.Options{}); ok {
+				return optimized
+			}
+			return providers.DefaultTurnSessions(profile, provider, providers.Options{})
+		},
+		ProbeProviderHealth: deps.probeProviderHealth,
+		UserAgent:           userAgent(),
 		PrepareRunCompletionWarning: func() {
 			scratchBaseline = scratchFileSnapshot(workspaceRoot)
 		},
@@ -959,6 +994,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			return scratchFileWarning(workspaceRoot, scratchBaseline)
 		},
 		Registry:           registry,
+		AwaitToolReadiness: awaitToolReadiness,
 		SessionStore:       sessionStore,
 		PeerService:        peerService,
 		SandboxStore:       sandboxStore,

internal/cli/app_test.go

@@ -17,6 +17,7 @@ import (
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/tui"
 	"github.com/Gitlawb/zero/internal/update"
@@ -211,12 +212,12 @@ func TestTUIStartupSuppressesWarningForUnconfiguredDefaultServer(t *testing.T) {
 		},
 		resolveMCPConfig: func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
-				"firecrawl": config.DefaultMCPServers()["firecrawl"],
+				"exa": config.DefaultMCPServers()["exa"],
 			}}, nil
 		},
 		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
 			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{
-				{Name: "firecrawl", Err: errors.New("returned HTTP 401"), UnconfiguredDefault: true},
+				{Name: "exa", Err: errors.New("returned HTTP 401"), UnconfiguredDefault: true},
 			}}, nil
 		},
 		runTUI: func(ctx context.Context, options tui.Options) int {
@@ -227,7 +228,7 @@ func TestTUIStartupSuppressesWarningForUnconfiguredDefaultServer(t *testing.T) {
 	if exitCode != 0 {
 		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
 	}
-	if strings.Contains(stderr.String(), "firecrawl") {
+	if strings.Contains(stderr.String(), "exa") {
 		t.Fatalf("stderr = %q, want no warning for an unconfigured default server", stderr.String())
 	}
 }
@@ -486,6 +487,74 @@ func TestRunNoArgsLaunchesTUIWithMCPState(t *testing.T) {
 	}
 }
 
+func TestRunNoArgsPaintsBeforeOptionalDefaultMCPIsReady(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	userConfigPath := filepath.Join(t.TempDir(), "zero", "config.json")
+	permissionStore, err := mcp.NewPermissionStore(mcp.StoreOptions{FilePath: filepath.Join(t.TempDir(), "mcp-permissions.json")})
+	if err != nil {
+		t.Fatalf("NewPermissionStore() error = %v", err)
+	}
+	tokenStore, err := mcp.NewTokenStore(mcp.TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "mcp-oauth.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	started := make(chan struct{})
+	release := make(chan struct{})
+	configResult := make(chan error, 1)
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd:          func() (string, error) { return cwd, nil },
+		userConfigPath: func() (string, error) { return userConfigPath, nil },
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{MaxTurns: 8}, nil
+		},
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: config.DefaultMCPServers()}, nil
+		},
+		newMCPStore:      func() (*mcp.PermissionStore, error) { return permissionStore, nil },
+		newMCPTokenStore: func() (*mcp.TokenStore, error) { return tokenStore, nil },
+		registerMCPTools: func(_ context.Context, registry *tools.Registry, cfg config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			var configErr error
+			if _, ok := cfg.Servers["exa"]; !ok || len(cfg.Servers) != 1 {
+				configErr = fmt.Errorf("optional MCP config = %#v, want only exa", cfg.Servers)
+			}
+			configResult <- configErr
+			close(started)
+			<-release
+			registry.Register(cliFakeMCPRegistryTool{})
+			return noopMCPRuntime{}, nil
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			select {
+			case <-started:
+			case <-time.After(time.Second):
+				t.Fatal("optional MCP initialization did not start")
+			}
+			if options.AwaitToolReadiness == nil {
+				t.Fatal("TUI did not receive optional tool readiness barrier")
+			}
+			if _, ok := options.Registry.Get("mcp_docs_lookup"); ok {
+				t.Fatal("optional MCP tool was published before registration completed")
+			}
+			close(release)
+			options.AwaitToolReadiness(ctx)
+			if _, ok := options.Registry.Get("mcp_docs_lookup"); !ok {
+				t.Fatal("optional MCP tool was not visible after readiness completed")
+			}
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	if configErr := <-configResult; configErr != nil {
+		t.Fatal(configErr)
+	}
+}
+
 func TestTUIMCPCommandUsesLastGoodConfigOnRefreshError(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
@@ -565,6 +634,7 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 		t.Fatalf("NewTokenStore() error = %v", err)
 	}
 	runtimeClosed := false
+	secret := "sk-proj-" + strings.Repeat("a", 23) + "0"
 
 	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
 		getwd: func() (string, error) {
@@ -591,7 +661,7 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 			return closeFunc(func() error {
 				runtimeClosed = true
 				return nil
-			}), errors.New("register mcp tools failed")
+			}), fmt.Errorf("register mcp tools failed: %s", secret)
 		},
 		runTUI: func(ctx context.Context, options tui.Options) int {
 			t.Fatal("TUI should not launch after MCP registration fails")
@@ -605,6 +675,12 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 	if !strings.Contains(stderr.String(), "register mcp tools failed") {
 		t.Fatalf("stderr missing registration error: %s", stderr.String())
 	}
+	if strings.Contains(stderr.String(), secret) {
+		t.Fatalf("stderr leaked MCP registration secret: %s", stderr.String())
+	}
+	if !strings.Contains(stderr.String(), redaction.RedactedSecret) {
+		t.Fatalf("stderr missing redaction marker: %s", stderr.String())
+	}
 	if !runtimeClosed {
 		t.Fatal("partial MCP runtime was not closed after registration error")
 	}

internal/cli/exec.go

@@ -939,6 +939,9 @@ func registerToolSearchIfEligible(registry *tools.Registry, deferThreshold int,
 	if deferThreshold <= 0 {
 		return
 	}
+	if _, registered := registry.Get(tools.ToolSearchToolName); registered {
+		return
+	}
 	if deferredEligibleCount(registry, permissionMode, enabledTools, disabledTools) < deferThreshold {
 		return
 	}

internal/cli/exec_writer.go

@@ -267,23 +267,29 @@ func streamJSONPermissionEventType(event agent.PermissionEvent) streamjson.Event
 func (writer *execEventWriter) usage(usage agent.Usage) {
 	if writer.format == execOutputJSON {
 		writer.writeJSON(map[string]any{
-			"type":              "usage",
-			"prompt_tokens":     usage.PromptTokens,
-			"completion_tokens": usage.CompletionTokens,
-			"total_tokens":      usage.TotalTokens(),
+			"type":                "usage",
+			"prompt_tokens":       usage.PromptTokens,
+			"completion_tokens":   usage.CompletionTokens,
+			"cached_input_tokens": usage.CachedInputTokens,
+			"cache_write_tokens":  usage.CacheWriteTokens,
+			"total_tokens":        usage.TotalTokens(),
 		})
 		return
 	}
 	if writer.format == execOutputStreamJSON {
 		promptTokens := usage.EffectiveInputTokens()
 		completionTokens := usage.EffectiveOutputTokens()
+		cachedInputTokens := usage.CachedInputTokens
+		cacheWriteTokens := usage.CacheWriteTokens
 		totalTokens := usage.TotalTokens()
 		writer.writeStreamJSON(streamjson.Event{
-			Type:             streamjson.EventUsage,
-			RunID:            writer.runID,
-			PromptTokens:     &promptTokens,
-			CompletionTokens: &completionTokens,
-			TotalTokens:      &totalTokens,
+			Type:              streamjson.EventUsage,
+			RunID:             writer.runID,
+			PromptTokens:      &promptTokens,
+			CompletionTokens:  &completionTokens,
+			CachedInputTokens: &cachedInputTokens,
+			CacheWriteTokens:  &cacheWriteTokens,
+			TotalTokens:       &totalTokens,
 		})
 	}
 }

internal/cli/exec_writer_test.go

@@ -52,3 +52,39 @@ func TestExecWriterPropagatesToolResultTruncation(t *testing.T) {
 		})
 	}
 }
+
+func TestExecWriterPropagatesCacheUsage(t *testing.T) {
+	for _, format := range []execOutputFormat{execOutputJSON, execOutputStreamJSON} {
+		t.Run(string(format), func(t *testing.T) {
+			var stdout bytes.Buffer
+			writer := execEventWriter{
+				stdout:       &stdout,
+				format:       format,
+				runID:        "run_usage",
+				streamedText: &strings.Builder{},
+			}
+			writer.usage(agent.Usage{
+				InputTokens:       100,
+				OutputTokens:      20,
+				PromptTokens:      100,
+				CompletionTokens:  20,
+				CachedInputTokens: 60,
+				CacheWriteTokens:  10,
+			})
+			if writer.err != nil {
+				t.Fatalf("usage: %v", writer.err)
+			}
+			var payload map[string]any
+			if err := json.Unmarshal(bytes.TrimSpace(stdout.Bytes()), &payload); err != nil {
+				t.Fatalf("decode output %q: %v", stdout.String(), err)
+			}
+			cachedKey, writeKey := "cached_input_tokens", "cache_write_tokens"
+			if format == execOutputStreamJSON {
+				cachedKey, writeKey = "cachedInputTokens", "cacheWriteTokens"
+			}
+			if payload[cachedKey] != float64(60) || payload[writeKey] != float64(10) {
+				t.Fatalf("cache usage missing from payload: %#v", payload)
+			}
+		})
+	}
+}

internal/cli/mcp_default_test.go

@@ -9,15 +9,15 @@ import (
 	"github.com/Gitlawb/zero/internal/config"
 )
 
-// `zero mcp disable firecrawl` must work even though firecrawl is a built-in
+// `zero mcp disable exa` must work even though exa is a built-in
 // default that is not written to the user's config file until overridden.
-func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
+func TestRunMCPDisableSeededExaDefault(t *testing.T) {
 	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
-	// A config with no firecrawl entry — the default lives in code, not the file.
+	// A config with no Exa entry — the default lives in code, not the file.
 	writeMCPCommandRawConfig(t, configPath, `{"activeProvider":"fast"}`)
 
 	var out, errBuf bytes.Buffer
-	code := runWithDeps([]string{"mcp", "disable", "firecrawl", "--json"}, &out, &errBuf, appDeps{
+	code := runWithDeps([]string{"mcp", "disable", "exa", "--json"}, &out, &errBuf, appDeps{
 		userConfigPath: func() (string, error) { return configPath, nil },
 	})
 	if code != exitSuccess {
@@ -31,8 +31,8 @@ func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
 	if err := json.Unmarshal(out.Bytes(), &payload); err != nil {
 		t.Fatalf("decode disable JSON: %v\n%s", err, out.String())
 	}
-	if payload.ServerName != "firecrawl" || !payload.Disabled || !payload.Changed {
-		t.Fatalf("disable payload = %#v, want firecrawl disabled+changed", payload)
+	if payload.ServerName != "exa" || !payload.Disabled || !payload.Changed {
+		t.Fatalf("disable payload = %#v, want exa disabled+changed", payload)
 	}
 
 	// End-to-end: resolving that config now turns the default off.
@@ -40,7 +40,7 @@ func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	if !cfg.Servers["firecrawl"].Disabled {
-		t.Fatal("expected `mcp disable firecrawl` to turn the seeded default off in the resolved config")
+	if !cfg.Servers["exa"].Disabled {
+		t.Fatal("expected `mcp disable exa` to turn the seeded default off in the resolved config")
 	}
 }

internal/cli/mcp_startup.go

@@ -0,0 +1,147 @@
+package cli
+
+import (
+	"context"
+	"sync"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+const (
+	optionalMCPPromptGrace = time.Second
+	optionalMCPCloseGrace  = 250 * time.Millisecond
+)
+
+// splitMCPStartupConfig keeps explicitly configured servers on the
+// prompt-critical path. Only unchanged built-in defaults may initialize after
+// first paint; their failure was already intentionally silent.
+func splitMCPStartupConfig(cfg config.MCPConfig) (critical config.MCPConfig, optional config.MCPConfig) {
+	critical.Servers = map[string]config.MCPServerConfig{}
+	optional.Servers = map[string]config.MCPServerConfig{}
+	for name, server := range cfg.Servers {
+		if server.Disabled {
+			continue
+		}
+		if config.IsUnconfiguredDefault(name, server) {
+			optional.Servers[name] = server
+			continue
+		}
+		critical.Servers[name] = server
+	}
+	return critical, optional
+}
+
+type optionalMCPStartup struct {
+	cancel context.CancelFunc
+	done   chan struct{}
+
+	mu      sync.Mutex
+	runtime mcpToolRuntime
+
+	closeOnce        sync.Once
+	runtimeCloseOnce sync.Once
+	closeErr         error
+}
+
+func startOptionalMCP(
+	parent context.Context,
+	registry *tools.Registry,
+	cfg config.MCPConfig,
+	options mcp.RegisterOptions,
+	register func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error),
+	onReady func(),
+) *optionalMCPStartup {
+	ctx, cancel := context.WithCancel(parent)
+	startup := &optionalMCPStartup{cancel: cancel, done: make(chan struct{})}
+	go func() {
+		runtime, err := register(ctx, registry, cfg, options)
+		if err == nil && onReady != nil {
+			onReady()
+		}
+		startup.mu.Lock()
+		startup.runtime = runtime
+		startup.mu.Unlock()
+		close(startup.done)
+	}()
+	return startup
+}
+
+// Await gives optional tools one bounded opportunity to join the next turn's
+// immutable registry snapshot. A timeout leaves this turn on the existing core
+// tool set; a later turn observes the completed generation.
+func (startup *optionalMCPStartup) Await(ctx context.Context, timeout time.Duration) bool {
+	if startup == nil {
+		return true
+	}
+	select {
+	case <-startup.done:
+		return true
+	default:
+	}
+	if timeout <= 0 {
+		select {
+		case <-startup.done:
+			return true
+		case <-ctx.Done():
+			return false
+		}
+	}
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	select {
+	case <-startup.done:
+		return true
+	case <-timer.C:
+		return false
+	case <-ctx.Done():
+		return false
+	}
+}
+
+func (startup *optionalMCPStartup) Close() error {
+	if startup == nil {
+		return nil
+	}
+	startup.closeOnce.Do(func() {
+		startup.cancel()
+		reaped := make(chan struct{})
+		go func() {
+			defer close(reaped)
+			<-startup.done
+			startup.closeRuntime()
+		}()
+		timer := time.NewTimer(optionalMCPCloseGrace)
+		defer timer.Stop()
+		select {
+		case <-reaped:
+		case <-timer.C:
+		}
+	})
+	startup.mu.Lock()
+	defer startup.mu.Unlock()
+	return startup.closeErr
+}
+
+func (startup *optionalMCPStartup) closeRuntime() {
+	startup.runtimeCloseOnce.Do(func() {
+		startup.mu.Lock()
+		runtime := startup.runtime
+		startup.mu.Unlock()
+		if runtime == nil {
+			return
+		}
+		err := runtime.Close()
+		startup.mu.Lock()
+		if err != nil && startup.closeErr == nil {
+			startup.closeErr = err
+		}
+		startup.mu.Unlock()
+	})
+}
+
+func (startup *optionalMCPStartup) Skipped() []mcp.SkippedServer {
+	return nil
+}

internal/cli/mcp_startup_test.go

@@ -0,0 +1,187 @@
+package cli
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+type closeTrackingMCPRuntime struct {
+	closed chan struct{}
+	err    error
+}
+
+func (runtime *closeTrackingMCPRuntime) Close() error {
+	close(runtime.closed)
+	return runtime.err
+}
+
+func (*closeTrackingMCPRuntime) Skipped() []mcp.SkippedServer { return nil }
+
+func TestSplitMCPStartupConfigDefersOnlyUnconfiguredDefaults(t *testing.T) {
+	defaults := config.DefaultMCPServers()
+	cfg := config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"exa":      defaults["exa"],
+		"docs":     {Type: "http", URL: "https://docs.example.test/mcp"},
+		"disabled": {Type: "http", URL: "https://disabled.example.test/mcp", Disabled: true},
+	}}
+
+	critical, optional := splitMCPStartupConfig(cfg)
+	if _, ok := critical.Servers["docs"]; !ok || len(critical.Servers) != 1 {
+		t.Fatalf("critical servers = %#v, want only docs", critical.Servers)
+	}
+	if _, ok := optional.Servers["exa"]; !ok || len(optional.Servers) != 1 {
+		t.Fatalf("optional servers = %#v, want only exa", optional.Servers)
+	}
+}
+
+func TestOptionalMCPAwaitTimesOutThenObservesPublishedTools(t *testing.T) {
+	registry := tools.NewRegistry()
+	release := make(chan struct{})
+	started := make(chan struct{})
+	startup := startOptionalMCP(
+		context.Background(),
+		registry,
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(ctx context.Context, registry *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			select {
+			case <-release:
+				registry.RegisterBatch([]tools.Tool{tools.NewScopedReadFileTool(t.TempDir(), nil)})
+				return noopMCPRuntime{}, nil
+			case <-ctx.Done():
+				return noopMCPRuntime{}, ctx.Err()
+			}
+		},
+		nil,
+	)
+	defer func() { _ = startup.Close() }()
+	<-started
+
+	if startup.Await(t.Context(), time.Millisecond) {
+		t.Fatal("optional startup unexpectedly completed before release")
+	}
+	if got := len(registry.All()); got != 0 {
+		t.Fatalf("registry published tools before readiness: %d", got)
+	}
+	close(release)
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("optional startup did not complete after release")
+	}
+	if _, ok := registry.Get("read_file"); !ok {
+		t.Fatal("completed optional startup did not publish its tool batch")
+	}
+}
+
+func TestOptionalMCPFailedRegistrationSkipsReadinessCallback(t *testing.T) {
+	readyCalls := 0
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return noopMCPRuntime{}, errors.New("connect failed")
+		},
+		func() { readyCalls++ },
+	)
+	defer func() { _ = startup.Close() }()
+
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("failed optional startup did not settle")
+	}
+	if readyCalls != 0 {
+		t.Fatalf("readiness callback ran after failed registration: %d", readyCalls)
+	}
+}
+
+func TestOptionalMCPCloseCancelsBlockedStartup(t *testing.T) {
+	started := make(chan struct{})
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(ctx context.Context, _ *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			<-ctx.Done()
+			return noopMCPRuntime{}, nil
+		},
+		nil,
+	)
+	<-started
+	if err := startup.Close(); err != nil {
+		t.Fatalf("Close() error = %v", err)
+	}
+}
+
+func TestOptionalMCPCloseReturnsRuntimeErrorWithinGrace(t *testing.T) {
+	closeErr := errors.New("close failed")
+	runtime := &closeTrackingMCPRuntime{closed: make(chan struct{}), err: closeErr}
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return runtime, nil
+		},
+		nil,
+	)
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("optional startup did not settle")
+	}
+	if err := startup.Close(); !errors.Is(err, closeErr) {
+		t.Fatalf("Close() error = %v, want %v", err, closeErr)
+	}
+}
+
+func TestOptionalMCPCloseBoundsRegistrationThatIgnoresCancellation(t *testing.T) {
+	started := make(chan struct{})
+	release := make(chan struct{})
+	released := false
+	defer func() {
+		if !released {
+			close(release)
+		}
+	}()
+	runtime := &closeTrackingMCPRuntime{closed: make(chan struct{})}
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			<-release
+			return runtime, nil
+		},
+		nil,
+	)
+	<-started
+
+	closeResult := make(chan error, 1)
+	go func() { closeResult <- startup.Close() }()
+	select {
+	case err := <-closeResult:
+		if err != nil {
+			t.Fatalf("Close() error = %v", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Close blocked on registration after cancellation")
+	}
+
+	close(release)
+	released = true
+	select {
+	case <-runtime.closed:
+	case <-time.After(time.Second):
+		t.Fatal("late optional MCP runtime was not closed")
+	}
+}

internal/config/mcp_defaults.go

@@ -5,22 +5,20 @@ import (
 	"strings"
 )
 
-// DefaultMCPServers returns the MCP servers Zero ships ENABLED by default so web
-// search and scraping work out of the box with no setup and no API key. They are
-// seeded before user/project config is merged (see ResolveMCP), so a user can
-// override any field — for example point firecrawl at a self-hosted instance, or
-// add an API-key header to lift the free-tier limit — or disable it entirely with
+// DefaultMCPServers returns the MCP servers Zero ships ENABLED by default so
+// web search and page fetching work out of the box with no setup and no API
+// key. They are seeded before user/project config is merged (see ResolveMCP),
+// so a user can override any field — for example add an API-key header to lift
+// Exa's anonymous rate limit — or disable it entirely with
 // `zero mcp disable <name>` (which writes `"disabled": true`).
 //
-// Keyless Firecrawl routes requests through firecrawl.dev (1,000 free credits per
-// month, no account). Self-host Firecrawl (AGPL-3.0) for unlimited and private
-// use. Zero only calls it over the network, so Firecrawl's license never reaches
-// into Zero's own code.
+// Exa's hosted MCP server works anonymously with rate limits. Users can add an
+// Exa API key for higher limits.
 func DefaultMCPServers() map[string]MCPServerConfig {
 	return map[string]MCPServerConfig{
-		"firecrawl": {
+		"exa": {
 			Type: "http",
-			URL:  "https://mcp.firecrawl.dev/v2/mcp",
+			URL:  "https://mcp.exa.ai/mcp",
 		},
 	}
 }
@@ -35,13 +33,13 @@ func IsDefaultMCPServer(name string) bool {
 
 // IsUnconfiguredDefault reports whether server is one of Zero's built-in
 // defaults that the user never wrote an entry for in their config — i.e. it is
-// running with whatever Zero ships (e.g. keyless Firecrawl, no credentials).
+// running with whatever Zero ships (e.g. keyless Exa, no credentials).
 //
 // Both conditions below must hold:
 //   - !server.configured: the user's JSON never declared an object for this
 //     server key at all (set by MCPServerConfig.UnmarshalJSON only when it
 //     actually ran for this key). Any explicit action — including a
-//     disable/enable toggle like `zero mcp enable firecrawl` that leaves the
+//     disable/enable toggle like `zero mcp enable exa` that leaves the
 //     resolved value unchanged — sets configured, so it always counts as
 //     user-configured, even though the value comparison below could not tell
 //     the difference on its own.
@@ -58,3 +56,90 @@ func IsUnconfiguredDefault(name string, server MCPServerConfig) bool {
 	def, ok := DefaultMCPServers()[strings.TrimSpace(name)]
 	return ok && !server.configured && reflect.DeepEqual(def, server)
 }
+
+// retiredDefaultMCPServer describes a built-in default Zero no longer ships.
+type retiredDefaultMCPServer struct {
+	// successor is the default that replaced it.
+	successor string
+	// shipped is the value Zero used to seed for this name. A user entry
+	// written while the default was still shipped may name only an override
+	// (a header, say) and rely on these fields for the rest, so they have to
+	// outlive the default itself.
+	shipped MCPServerConfig
+}
+
+// retiredDefaultMCPServers maps a retired built-in default to what replaced it
+// and to the value Zero used to ship for it.
+var retiredDefaultMCPServers = map[string]retiredDefaultMCPServer{
+	"firecrawl": {
+		successor: "exa",
+		shipped: MCPServerConfig{
+			Type: "http",
+			URL:  "https://mcp.firecrawl.dev/v2/mcp",
+		},
+	},
+}
+
+// migrateRetiredDefaultMCPServers keeps an upgrade across a default-provider
+// swap from changing what the user's config means. It handles the two ways a
+// rename can break a config written against the old default:
+//
+//   - A user who ran `zero mcp disable <old default>` made an explicit choice
+//     not to open that outbound connection. Without carrying that decision to
+//     the successor, the upgrade re-opens it under the new name — and because
+//     the replacement looks like an untouched default, even the startup
+//     warning stays quiet (see IsUnconfiguredDefault and issue #552).
+//   - A user who customized the old default could name only the field they
+//     were changing, because the seeded default supplied the transport. Once
+//     the default stops being seeded that entry resolves to a server with no
+//     type, url, or command, which fails NormalizeConfig and takes the whole
+//     startup down rather than just that one server.
+//
+// It runs immediately after the user layer merges (see ResolveMCP), which
+// scopes it to user-level decisions — the only scope whose disable is sticky.
+// The carried-over disable applies only when the user never declared the
+// successor themselves: an explicit `exa` entry wins whether it enables or
+// disables. Because the disable is recorded as a user-level decision, the
+// lower-trust project layer cannot lift it, while `zero mcp enable exa` still
+// can (the CLI override scope merges with canReenable=true).
+func migrateRetiredDefaultMCPServers(cfg *MCPConfig) {
+	for name, retired := range retiredDefaultMCPServers {
+		entry, ok := cfg.Servers[name]
+		if !ok {
+			continue
+		}
+		if inheritsRetiredTransport(entry) {
+			// Re-seed what Zero used to ship underneath the user's fields, so a
+			// partial entry still resolves to the complete server it named when
+			// it was written. Their own values still win — this only fills the
+			// gap the retired default used to cover.
+			entry = mergeMCPServer(retired.shipped, entry, true)
+			cfg.Servers[name] = entry
+		}
+		if !entry.disabledSet || !entry.Disabled {
+			continue
+		}
+		replacement, ok := cfg.Servers[retired.successor]
+		if !ok || replacement.configured {
+			continue
+		}
+		replacement.Disabled = true
+		replacement.disabledSet = true
+		cfg.Servers[retired.successor] = replacement
+	}
+}
+
+// inheritsRetiredTransport reports whether entry named no transport of its own
+// — no type, url, command, or args — and so relied entirely on the retired
+// default to supply one.
+//
+// Anything else is a complete definition the user owns (a self-hosted url, a
+// stdio command) and must be left alone: grafting the retired default's http
+// type onto an entry that names a command produces a server that is rejected
+// outright, turning a working config into a startup failure.
+func inheritsRetiredTransport(entry MCPServerConfig) bool {
+	return strings.TrimSpace(entry.Type) == "" &&
+		strings.TrimSpace(entry.URL) == "" &&
+		strings.TrimSpace(entry.Command) == "" &&
+		len(entry.Args) == 0
+}

internal/config/mcp_defaults_test.go

@@ -7,10 +7,10 @@ import (
 )
 
 func TestIsDefaultMCPServer(t *testing.T) {
-	if !IsDefaultMCPServer("firecrawl") {
-		t.Fatal("firecrawl should be a built-in default")
+	if !IsDefaultMCPServer("exa") {
+		t.Fatal("exa should be a built-in default")
 	}
-	if IsDefaultMCPServer("  firecrawl  ") == false {
+	if IsDefaultMCPServer("  exa  ") == false {
 		t.Fatal("IsDefaultMCPServer should trim whitespace")
 	}
 	if IsDefaultMCPServer("not-a-default") {
@@ -18,45 +18,45 @@ func TestIsDefaultMCPServer(t *testing.T) {
 	}
 }
 
-func TestResolveMCPSeedsEnabledFirecrawlDefault(t *testing.T) {
+func TestResolveMCPSeedsEnabledExaDefault(t *testing.T) {
 	cfg, err := ResolveMCP(ResolveOptions{})
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	firecrawl, ok := cfg.Servers["firecrawl"]
+	exa, ok := cfg.Servers["exa"]
 	if !ok {
-		t.Fatal("expected the firecrawl default to be seeded with no user config")
+		t.Fatal("expected the exa default to be seeded with no user config")
 	}
-	if firecrawl.Type != "http" || firecrawl.URL != "https://mcp.firecrawl.dev/v2/mcp" {
-		t.Fatalf("unexpected firecrawl default: %#v", firecrawl)
+	if exa.Type != "http" || exa.URL != "https://mcp.exa.ai/mcp" {
+		t.Fatalf("unexpected exa default: %#v", exa)
 	}
-	if firecrawl.Disabled {
-		t.Fatal("the firecrawl default must be enabled out of the box")
+	if exa.Disabled {
+		t.Fatal("the exa default must be enabled out of the box")
 	}
 }
 
 func TestResolveMCPUserCanDisableDefault(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "config.json")
-	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"exa":{"disabled":true}}}}`), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	if !cfg.Servers["firecrawl"].Disabled {
+	if !cfg.Servers["exa"].Disabled {
 		t.Fatal("a user must be able to disable the default by writing over it")
 	}
 }
 
 func TestIsUnconfiguredDefault(t *testing.T) {
-	if !IsUnconfiguredDefault("firecrawl", DefaultMCPServers()["firecrawl"]) {
-		t.Fatal("an untouched firecrawl default should be reported as unconfigured")
+	if !IsUnconfiguredDefault("exa", DefaultMCPServers()["exa"]) {
+		t.Fatal("an untouched exa default should be reported as unconfigured")
 	}
-	if IsUnconfiguredDefault("firecrawl", MCPServerConfig{Type: "http", URL: "http://localhost:3002/mcp"}) {
+	if IsUnconfiguredDefault("exa", MCPServerConfig{Type: "http", URL: "https://example.com/mcp"}) {
 		t.Fatal("a server overriding the default URL is no longer unconfigured")
 	}
-	if IsUnconfiguredDefault("firecrawl", MCPServerConfig{Type: "http", URL: "https://mcp.firecrawl.dev/v2/mcp", Auth: "bearer"}) {
+	if IsUnconfiguredDefault("exa", MCPServerConfig{Type: "http", URL: "https://mcp.exa.ai/mcp", Auth: "bearer"}) {
 		t.Fatal("a server with credentials added is no longer unconfigured")
 	}
 	if IsUnconfiguredDefault("not-a-default", MCPServerConfig{}) {
@@ -65,57 +65,173 @@ func TestIsUnconfiguredDefault(t *testing.T) {
 }
 
 func TestResolveMCPExplicitReenableIsNotUnconfiguredDefault(t *testing.T) {
-	// `zero mcp enable firecrawl` after a prior disable writes {"disabled":false}
+	// `zero mcp enable exa` after a prior disable writes {"disabled":false}
 	// explicitly. The resolved value is identical to the untouched default (both
 	// enabled, no credentials), but the user DID take an explicit action here, so
 	// IsUnconfiguredDefault must not treat it as untouched (issue #563 review).
 	path := filepath.Join(t.TempDir(), "config.json")
-	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":false}}}}`), 0o600); err != nil {
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"exa":{"disabled":false}}}}`), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	firecrawl := cfg.Servers["firecrawl"]
-	if firecrawl.Disabled {
-		t.Fatalf("explicit re-enable should leave the server enabled: %#v", firecrawl)
+	exa := cfg.Servers["exa"]
+	if exa.Disabled {
+		t.Fatalf("explicit re-enable should leave the server enabled: %#v", exa)
 	}
-	if IsUnconfiguredDefault("firecrawl", firecrawl) {
+	if IsUnconfiguredDefault("exa", exa) {
 		t.Fatal("an explicit enable/disable toggle must count as user-configured, even though the resolved value matches the default")
 	}
 }
 
 func TestResolveMCPExplicitRedeclareOfDefaultValuesIsNotUnconfiguredDefault(t *testing.T) {
-	// A user who copies firecrawl's exact default type/url into their config
+	// A user who copies Exa's exact default type/url into their config
 	// (e.g. from an example file, planning to add credentials later) produces a
-	// resolved value byte-identical to DefaultMCPServers()["firecrawl"] — the
+	// resolved value byte-identical to DefaultMCPServers()["exa"] — the
 	// same trap TestResolveMCPExplicitReenableIsNotUnconfiguredDefault covers for
 	// the disabled toggle. IsUnconfiguredDefault must still treat this as
 	// user-configured because the user's JSON declared an entry for it, even
 	// though a plain resolved-value comparison could not tell the difference.
 	path := filepath.Join(t.TempDir(), "config.json")
-	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"type":"http","url":"https://mcp.firecrawl.dev/v2/mcp"}}}}`), 0o600); err != nil {
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"exa":{"type":"http","url":"https://mcp.exa.ai/mcp"}}}}`), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	firecrawl := cfg.Servers["firecrawl"]
-	want := DefaultMCPServers()["firecrawl"]
-	if firecrawl.Type != want.Type || firecrawl.URL != want.URL || firecrawl.Disabled != want.Disabled {
-		t.Fatalf("expected the resolved value to match the default's fields exactly: %#v", firecrawl)
+	exa := cfg.Servers["exa"]
+	want := DefaultMCPServers()["exa"]
+	if exa.Type != want.Type || exa.URL != want.URL || exa.Disabled != want.Disabled {
+		t.Fatalf("expected the resolved value to match the default's fields exactly: %#v", exa)
 	}
-	if IsUnconfiguredDefault("firecrawl", firecrawl) {
+	if IsUnconfiguredDefault("exa", exa) {
 		t.Fatal("redeclaring the default's exact values is still an explicit user configuration, not an untouched default")
 	}
 }
 
 func TestResolveMCPUserCanOverrideDefaultURLKeepingOtherFields(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "config.json")
-	// Point firecrawl at a self-hosted instance; the default's Type must survive.
-	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"url":"http://localhost:3002/mcp"}}}}`), 0o600); err != nil {
+	// Point Exa at a proxy; the default's Type must survive.
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"exa":{"url":"https://example.com/mcp"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	exa := cfg.Servers["exa"]
+	if exa.URL != "https://example.com/mcp" {
+		t.Fatalf("user override of the default URL did not apply: %#v", exa)
+	}
+	if exa.Type != "http" {
+		t.Fatalf("override should keep the default's other fields (type), got %#v", exa)
+	}
+}
+
+func TestResolveMCPCarriesLegacyDefaultDisableToSuccessor(t *testing.T) {
+	// Upgrade path: the user disabled the firecrawl default Zero used to ship.
+	// Swapping the default to exa must not re-open an outbound connection they
+	// explicitly switched off.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	if !cfg.Servers["exa"].Disabled {
+		t.Fatalf("a prior disable of the retired default must carry to its replacement: %#v", cfg.Servers["exa"])
+	}
+}
+
+func TestResolveMCPExplicitSuccessorBeatsLegacyDefaultDisable(t *testing.T) {
+	// The user disabled the old default but has since explicitly configured exa.
+	// Their newer, explicit choice wins — the migration must not override it.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true},"exa":{"disabled":false}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	exa := cfg.Servers["exa"]
+	if exa.Disabled {
+		t.Fatalf("an explicit exa entry must survive the legacy-disable migration: %#v", exa)
+	}
+	if IsUnconfiguredDefault("exa", exa) {
+		t.Fatal("an explicitly configured exa is not an untouched default")
+	}
+}
+
+func TestResolveMCPLegacyDefaultLeftEnabledDoesNotDisableSuccessor(t *testing.T) {
+	// A user who configured firecrawl without disabling it (e.g. added a header)
+	// never opted out of the default search server, so exa stays enabled.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"headers":{"Authorization":"Bearer k"}}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	if cfg.Servers["exa"].Disabled {
+		t.Fatal("only an explicit disable of the retired default should carry over")
+	}
+}
+
+func TestResolveMCPLegacyDefaultDisableIsLiftableByOverride(t *testing.T) {
+	// The carried-over disable is a user-level decision, not a permanent one:
+	// `zero mcp enable exa` merges through the CLI override scope, which is the
+	// one layer allowed to re-enable.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	enabled := MCPServerConfig{Disabled: false, disabledSet: true}
+	cfg, err := ResolveMCP(ResolveOptions{
+		UserConfigPath: path,
+		Overrides:      Overrides{MCP: MCPConfig{Servers: map[string]MCPServerConfig{"exa": enabled}}},
+	})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	if cfg.Servers["exa"].Disabled {
+		t.Fatalf("an explicit enable must lift the carried-over disable: %#v", cfg.Servers["exa"])
+	}
+}
+
+func TestResolveMCPLegacyDisableSurvivesProjectReenable(t *testing.T) {
+	// The project layer is lower-trust and must not lift the carried-over
+	// user-level disable — the same guard a direct `zero mcp disable exa` gets.
+	dir := t.TempDir()
+	userPath := filepath.Join(dir, "config.json")
+	if err := os.WriteFile(userPath, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	projectPath := filepath.Join(dir, "project.json")
+	if err := os.WriteFile(projectPath, []byte(`{"mcp":{"servers":{"exa":{"disabled":false}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	if !cfg.Servers["exa"].Disabled {
+		t.Fatal("project config must not re-enable a server the user disabled")
+	}
+}
+
+func TestResolveMCPKeepsRetiredDefaultTransportForPartialEntry(t *testing.T) {
+	// A user who added an API key to the firecrawl default wrote only the header
+	// — the seeded default supplied the http type and URL. Retiring the default
+	// must not strip those out from under them and leave an unusable server.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"headers":{"Authorization":"Bearer k"}}}}}`), 0o600); err != nil {
 		t.Fatal(err)
 	}
 	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
@@ -123,10 +239,55 @@ func TestResolveMCPUserCanOverrideDefaultURLKeepingOtherFields(t *testing.T) {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
 	firecrawl := cfg.Servers["firecrawl"]
-	if firecrawl.URL != "http://localhost:3002/mcp" {
-		t.Fatalf("user override of the default URL did not apply: %#v", firecrawl)
+	if firecrawl.Type != "http" || firecrawl.URL != "https://mcp.firecrawl.dev/v2/mcp" {
+		t.Fatalf("a partial entry must keep the retired default's transport: %#v", firecrawl)
 	}
-	if firecrawl.Type != "http" {
-		t.Fatalf("override should keep the default's other fields (type), got %#v", firecrawl)
+	if firecrawl.Headers["Authorization"] != "Bearer k" {
+		t.Fatalf("the user's own field must still win: %#v", firecrawl)
+	}
+}
+
+func TestResolveMCPDoesNotGraftRetiredDefaultOntoOwnTransport(t *testing.T) {
+	// A self-hosted firecrawl over stdio is a complete definition the user owns.
+	// Re-seeding the retired http default under it would produce an http server
+	// carrying a command, which NormalizeConfig rejects outright.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"command":"firecrawl-mcp"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	firecrawl := cfg.Servers["firecrawl"]
+	if firecrawl.Type != "" || firecrawl.URL != "" {
+		t.Fatalf("an entry naming its own transport must be left alone: %#v", firecrawl)
+	}
+	if firecrawl.Command != "firecrawl-mcp" {
+		t.Fatalf("the user's command must survive: %#v", firecrawl)
+	}
+}
+
+func TestResolveMCPDisabledRetiredEntryStaysReEnableable(t *testing.T) {
+	// The disable carries to exa, and the retired entry itself keeps a usable
+	// transport so a later `zero mcp enable firecrawl` does not resolve to a
+	// server with no type, url, or command.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	firecrawl := cfg.Servers["firecrawl"]
+	if !firecrawl.Disabled {
+		t.Fatalf("the retired entry must stay disabled: %#v", firecrawl)
+	}
+	if firecrawl.Type != "http" || firecrawl.URL != "https://mcp.firecrawl.dev/v2/mcp" {
+		t.Fatalf("a disabled retired entry should still carry a usable transport: %#v", firecrawl)
+	}
+	if !cfg.Servers["exa"].Disabled {
+		t.Fatal("the disable must still carry to the successor")
 	}
 }

internal/config/resolve_mcp_exclude_project_test.go

@@ -34,7 +34,7 @@ func TestResolveMCPExcludeProjectDropsProjectServers(t *testing.T) {
 		if _, ok := cfg.Servers["user-srv"]; !ok {
 			t.Fatal("the user server must always be present")
 		}
-		if _, ok := cfg.Servers["firecrawl"]; !ok {
+		if _, ok := cfg.Servers["exa"]; !ok {
 			t.Fatal("the built-in default must always be present")
 		}
 	})
@@ -54,7 +54,7 @@ func TestResolveMCPExcludeProjectDropsProjectServers(t *testing.T) {
 		if _, ok := cfg.Servers["user-srv"]; !ok {
 			t.Fatal("the user server must survive when the project layer is dropped")
 		}
-		if _, ok := cfg.Servers["firecrawl"]; !ok {
+		if _, ok := cfg.Servers["exa"]; !ok {
 			t.Fatal("the built-in default must survive when the project layer is dropped")
 		}
 	})

internal/config/resolver.go

@@ -180,7 +180,7 @@ func Resolve(options ResolveOptions) (ResolvedConfig, error) {
 }
 
 func ResolveMCP(options ResolveOptions) (MCPConfig, error) {
-	// Seed Zero's built-in default MCP servers (e.g. keyless Firecrawl for free,
+	// Seed Zero's built-in default MCP servers (e.g. keyless Exa for free,
 	// no-setup web search/scrape) BEFORE merging user/project config, so the user
 	// can override any field or disable a default by writing over it.
 	cfg := FileConfig{MCP: MCPConfig{Servers: DefaultMCPServers()}}
@@ -194,6 +194,10 @@ func ResolveMCP(options ResolveOptions) (MCPConfig, error) {
 		// server the user disabled (a user-level disable is sticky, but the user
 		// scope itself may lift it).
 		mergeMCPConfig(&cfg.MCP, fileConfig.MCP, true)
+		// Reconcile config written against a default Zero has since retired,
+		// here while the user layer is the only one merged — so a carried-over
+		// disable counts as a user-level decision the project layer cannot lift.
+		migrateRetiredDefaultMCPServers(&cfg.MCP)
 	}
 	// Drop the project layer when the workspace is untrusted, so a cloned repo's
 	// ./.zero/config.json cannot register (and spawn) MCP servers. Fail-closed:

internal/config/types.go

@@ -465,7 +465,7 @@ type MCPServerConfig struct {
 	// fields it set or what values they hold. A built-in default seeded by
 	// DefaultMCPServers() is never unmarshaled from JSON, so it starts false;
 	// any explicit entry in the user/project file — even one that happens to
-	// repeat a default's exact field values (e.g. re-declaring firecrawl's
+	// repeat a default's exact field values (e.g. re-declaring Exa's
 	// default URL) — sets it true. IsUnconfiguredDefault checks this alongside
 	// a resolved-value comparison, so redeclaring default values verbatim still
 	// counts as user-configured.

internal/execution/contracts.go

@@ -9,7 +9,16 @@ import (
 	"strings"
 )
 
-const PolicyVersion = 1
+// PolicyVersion is the version of the approval policy that recorded grants were
+// evaluated under. Bump it whenever a change narrows what may be approved: the
+// grant store then backs the file up, keeps deny grants, invalidates prior
+// approvals, and tells the user once, instead of silently honoring approvals
+// that the current rules would have refused.
+//
+// v2 tightened command-prefix validation so a launcher spelled with a version
+// or an executable extension ("python3.11", "python.exe") can no longer be
+// approved as an ordinary command.
+const PolicyVersion = 2
 
 type Origin string
 

internal/mcp/config.go

@@ -35,7 +35,7 @@ type Server struct {
 	// credential lookup uses it to avoid reusing legacy user tokens by name.
 	ProjectConfigured bool
 	// UnconfiguredDefault is true when this server is one of Zero's built-in
-	// defaults (e.g. keyless Firecrawl) that the user never touched in their
+	// defaults (e.g. keyless Exa) that the user never touched in their
 	// config — no credentials, no overrides. Callers use it to avoid warning
 	// loudly when a server nobody configured fails to connect.
 	UnconfiguredDefault bool

internal/mcp/config_test.go

@@ -1,6 +1,8 @@
 package mcp
 
 import (
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 
@@ -90,7 +92,7 @@ func TestNormalizeConfigValidatesTransportBoundaries(t *testing.T) {
 
 func TestNormalizeConfigFlagsUnconfiguredDefault(t *testing.T) {
 	cfg := config.MCPConfig{Servers: map[string]config.MCPServerConfig{
-		"firecrawl": config.DefaultMCPServers()["firecrawl"],
+		"exa": config.DefaultMCPServers()["exa"],
 		"web": {
 			Type: "http",
 			URL:  "https://example.com/mcp",
@@ -107,8 +109,8 @@ func TestNormalizeConfigFlagsUnconfiguredDefault(t *testing.T) {
 		byName[server.Name] = server
 	}
 
-	if !byName["firecrawl"].UnconfiguredDefault {
-		t.Fatal("an untouched firecrawl default should be flagged UnconfiguredDefault")
+	if !byName["exa"].UnconfiguredDefault {
+		t.Fatal("an untouched exa default should be flagged UnconfiguredDefault")
 	}
 	if byName["web"].UnconfiguredDefault {
 		t.Fatal("a server the user configured must not be flagged UnconfiguredDefault")
@@ -240,3 +242,116 @@ func TestCopyStringMapTrimsKeysAndPreservesValues(t *testing.T) {
 		t.Fatalf("copied[TOKEN] = %q, want value preserved verbatim", copied["TOKEN"])
 	}
 }
+
+// Startup behavior for the firecrawl -> exa default migration: a user who
+// disabled the retired default must end up with no Exa server to connect to,
+// not merely a disabled entry in the resolved config.
+func TestNormalizeConfigSkipsExaAfterLegacyFirecrawlDisable(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":true}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.ResolveMCP(config.ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	servers, err := NormalizeConfig(cfg)
+	if err != nil {
+		t.Fatalf("NormalizeConfig: %v", err)
+	}
+	for _, server := range servers {
+		if server.Name == "exa" {
+			t.Fatalf("exa must not be started after the user disabled the default it replaced: %#v", server)
+		}
+	}
+}
+
+// The same path with no legacy disable still starts Exa, so the migration
+// cannot quietly switch off the default for everyone else.
+func TestNormalizeConfigStartsExaWithoutLegacyDisable(t *testing.T) {
+	cfg, err := config.ResolveMCP(config.ResolveOptions{})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	servers, err := NormalizeConfig(cfg)
+	if err != nil {
+		t.Fatalf("NormalizeConfig: %v", err)
+	}
+	for _, server := range servers {
+		if server.Name == "exa" {
+			if !server.UnconfiguredDefault {
+				t.Fatalf("an untouched exa default should be flagged unconfigured: %#v", server)
+			}
+			return
+		}
+	}
+	t.Fatal("expected the exa default to be started out of the box")
+}
+
+// The failure path CodeRabbit flagged: a firecrawl entry that named only a
+// header relied on the retired default for its transport. Resolving it must
+// still produce a complete server — an incomplete one fails NormalizeConfig,
+// which takes down every other server's startup with it, not just this one.
+func TestNormalizeConfigResolvesPartialRetiredDefaultEntry(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"headers":{"Authorization":"Bearer k"}}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.ResolveMCP(config.ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	servers, err := NormalizeConfig(cfg)
+	if err != nil {
+		t.Fatalf("a header-only legacy entry must not break startup: %v", err)
+	}
+	var firecrawl *Server
+	var sawExa bool
+	for i := range servers {
+		switch servers[i].Name {
+		case "firecrawl":
+			firecrawl = &servers[i]
+		case "exa":
+			sawExa = true
+		}
+	}
+	if firecrawl == nil {
+		t.Fatalf("expected the customized firecrawl server to survive: %#v", servers)
+	}
+	if firecrawl.Type != ServerTypeHTTP || firecrawl.URL != "https://mcp.firecrawl.dev/v2/mcp" {
+		t.Fatalf("firecrawl lost the transport its entry relied on: %#v", *firecrawl)
+	}
+	if firecrawl.Headers["Authorization"] != "Bearer k" {
+		t.Fatalf("firecrawl lost the user's header: %#v", *firecrawl)
+	}
+	if !sawExa {
+		t.Fatal("the new exa default should still start alongside it")
+	}
+}
+
+// The mirror case: an entry that names its own stdio transport must not have
+// the retired http default grafted onto it, which would be rejected outright.
+func TestNormalizeConfigLeavesRetiredEntryWithOwnTransportAlone(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"command":"firecrawl-mcp"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := config.ResolveMCP(config.ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	servers, err := NormalizeConfig(cfg)
+	if err != nil {
+		t.Fatalf("a self-hosted stdio entry must keep working: %v", err)
+	}
+	for _, server := range servers {
+		if server.Name != "firecrawl" {
+			continue
+		}
+		if server.Type != ServerTypeStdio || server.Command != "firecrawl-mcp" {
+			t.Fatalf("the user's own stdio transport was corrupted: %#v", server)
+		}
+		return
+	}
+	t.Fatalf("expected the self-hosted firecrawl server to survive: %#v", servers)
+}

internal/mcp/registry.go

@@ -176,9 +176,11 @@ func RegisterTools(ctx context.Context, registry *tools.Registry, cfg config.MCP
 			staged = append(staged, tool)
 		}
 	}
-	for _, tool := range staged {
-		registry.Register(tool)
+	registered := make([]tools.Tool, 0, len(staged))
+	for index := range staged {
+		registered = append(registered, staged[index])
 	}
+	registry.RegisterBatch(registered)
 	return runtime, nil
 }
 

internal/mcp/registry_test.go

@@ -70,6 +70,62 @@ func TestRegisterToolsAddsPromptGatedMCPTools(t *testing.T) {
 	}
 }
 
+func TestRegisterToolsPublishesServerToolsInOneGeneration(t *testing.T) {
+	registry := tools.NewRegistry()
+	before := registry.Snapshot().Generation
+	client := &fakeToolClient{listed: []RemoteTool{
+		{Name: "lookup", Description: "Lookup documentation"},
+		{Name: "search", Description: "Search documentation"},
+	}}
+	runtime, err := RegisterTools(context.Background(), registry, config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"docs": {Type: "stdio", Command: "docs-mcp"},
+	}}, RegisterOptions{ClientFactory: func(context.Context, Server) (ToolClient, error) {
+		return client, nil
+	}})
+	if err != nil {
+		t.Fatalf("RegisterTools() error = %v", err)
+	}
+	defer runtime.Close()
+
+	after := registry.Snapshot()
+	if after.Generation != before+1 {
+		t.Fatalf("registry generation = %d, want %d", after.Generation, before+1)
+	}
+	for _, name := range []string{"mcp_docs_lookup", "mcp_docs_search"} {
+		if _, ok := registry.Get(name); !ok {
+			t.Fatalf("expected %s to be published in the batch", name)
+		}
+	}
+}
+
+func TestRegisterToolsPublishesNoneWhenServerToolValidationFails(t *testing.T) {
+	registry := tools.NewRegistry()
+	before := registry.Snapshot().Generation
+	client := &fakeToolClient{listed: []RemoteTool{
+		{Name: "lookup", Description: "Lookup documentation"},
+		{Name: "", Description: "Invalid nameless tool"},
+	}}
+	runtime, err := RegisterTools(context.Background(), registry, config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"docs": {Type: "stdio", Command: "docs-mcp"},
+	}}, RegisterOptions{ClientFactory: func(context.Context, Server) (ToolClient, error) {
+		return client, nil
+	}})
+	if err != nil {
+		t.Fatalf("RegisterTools() error = %v", err)
+	}
+	defer runtime.Close()
+
+	if got := registry.Snapshot().Generation; got != before {
+		t.Fatalf("registry generation = %d, want unchanged %d", got, before)
+	}
+	if _, ok := registry.Get("mcp_docs_lookup"); ok {
+		t.Fatal("valid prefix of rejected server batch was published")
+	}
+	if skipped := runtime.Skipped(); len(skipped) != 1 || skipped[0].Name != "docs" || skipped[0].Err == nil {
+		t.Fatalf("Skipped() = %#v, want one validation failure for docs", skipped)
+	}
+}
+
 func TestRegisterToolsMarksPersistentlyApprovedToolsAllow(t *testing.T) {
 	store, err := NewPermissionStore(StoreOptions{
 		FilePath: filepath.Join(t.TempDir(), "permissions.json"),
@@ -153,16 +209,16 @@ func TestRegisterToolsSkipsUnreachableServerAndKeepsOthers(t *testing.T) {
 }
 
 func TestRegisterToolsFlagsUnconfiguredDefaultOnSkip(t *testing.T) {
-	// firecrawl is seeded by config.DefaultMCPServers with no credentials; when a
-	// user never configures it, a connect failure (e.g. the real HTTP 401 from
-	// issue #552) must be recorded as UnconfiguredDefault so callers can avoid
+	// Exa is seeded by config.DefaultMCPServers with no credentials; when a
+	// user never configures it, a connect failure must be recorded as
+	// UnconfiguredDefault so callers can avoid
 	// warning about a server the user never asked for. "custom" is configured by
 	// the user and must NOT be flagged even though it also fails to connect.
 	registry := tools.NewRegistry()
 
 	runtime, err := RegisterTools(context.Background(), registry, config.MCPConfig{Servers: map[string]config.MCPServerConfig{
-		"firecrawl": config.DefaultMCPServers()["firecrawl"],
-		"custom":    {Type: "stdio", Command: "custom-mcp"},
+		"exa":    config.DefaultMCPServers()["exa"],
+		"custom": {Type: "stdio", Command: "custom-mcp"},
 	}}, RegisterOptions{
 		ClientFactory: func(_ context.Context, server Server) (ToolClient, error) {
 			return nil, errors.New(server.Name + " connect failed")
@@ -177,12 +233,12 @@ func TestRegisterToolsFlagsUnconfiguredDefaultOnSkip(t *testing.T) {
 	for _, skipped := range runtime.Skipped() {
 		byName[skipped.Name] = skipped
 	}
-	firecrawl, ok := byName["firecrawl"]
+	exa, ok := byName["exa"]
 	if !ok {
-		t.Fatalf("Skipped() = %#v, want an entry for firecrawl", runtime.Skipped())
+		t.Fatalf("Skipped() = %#v, want an entry for exa", runtime.Skipped())
 	}
-	if !firecrawl.UnconfiguredDefault {
-		t.Fatalf("Skipped()[firecrawl] = %#v, want UnconfiguredDefault", firecrawl)
+	if !exa.UnconfiguredDefault {
+		t.Fatalf("Skipped()[exa] = %#v, want UnconfiguredDefault", exa)
 	}
 	custom, ok := byName["custom"]
 	if !ok {

internal/modelregistry/catalog.go

@@ -231,6 +231,12 @@ func reasoningEffortsForModelName(name string) []ReasoningEffort {
 		// the gateway's translation concern — unknown fields are ignored, so the
 		// worst case is a silent no-op rather than a 400.
 		return []ReasoningEffort{ReasoningEffortLow, ReasoningEffortMedium, ReasoningEffortHigh}
+	case strings.Contains(n, "deepseek"):
+		// DeepSeek V4 (deepseek-v4-flash / deepseek-v4-pro) exposes thinking-mode
+		// effort control through the OpenAI-compatible reasoning_effort parameter.
+		// The API accepts low/high/max and maps medium/xhigh onto high, so only
+		// the three distinct tiers are advertised.
+		return []ReasoningEffort{ReasoningEffortLow, ReasoningEffortHigh, ReasoningEffortMax}
 	default:
 		return nil
 	}

internal/modelregistry/effort_fallback_test.go

@@ -20,6 +20,10 @@ func TestReasoningEffortsFallbackForGPT5AndOSeries(t *testing.T) {
 		{"hy3", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortMedium, ReasoningEffortHigh}},
 		{"hunyuan-t1", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortMedium, ReasoningEffortHigh}},
 		{"openai/o3-mini", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortMedium, ReasoningEffortHigh}}, // github-style vendor/ id
+		{"deepseek-v4-flash", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortHigh, ReasoningEffortMax}},
+		{"deepseek-v4-pro", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortHigh, ReasoningEffortMax}},
+		{"deepseek-reasoner", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortHigh, ReasoningEffortMax}},
+		{"deepseek-chat", []ReasoningEffort{ReasoningEffortLow, ReasoningEffortHigh, ReasoningEffortMax}},
 		{"gpt-4.1", nil}, // non-reasoning, registered: stays empty
 		{"gpt-4o-mini", nil},
 		{"ollama/llama3.1", nil},

internal/perfbench/turn_bench.go

@@ -191,6 +191,7 @@ type TurnBenchResult struct {
 type TurnBenchTotals struct {
 	InputTokens       int64 `json:"inputTokens"`
 	CachedInputTokens int64 `json:"cachedInputTokens"`
+	CacheWriteTokens  int64 `json:"cacheWriteTokens"`
 	OutputTokens      int64 `json:"outputTokens"`
 	ModelRequests     int64 `json:"modelRequests"`
 	ToolCalls         int64 `json:"toolCalls"`
@@ -479,6 +480,7 @@ func topLatencySources(perSpan map[string]SpanStats, top int) []LatencySource {
 func aggregateTotals(totals *TurnBenchTotals, tr *trace.TurnTrace) {
 	totals.InputTokens += tr.Counter(trace.CounterInputTokens)
 	totals.CachedInputTokens += tr.Counter(trace.CounterCachedInputTokens)
+	totals.CacheWriteTokens += tr.Counter(trace.CounterCacheWriteTokens)
 	totals.OutputTokens += tr.Counter(trace.CounterOutputTokens)
 	totals.ModelRequests += tr.Counter(trace.CounterModelRequests)
 	totals.ToolCalls += tr.Counter(trace.CounterToolCalls)
@@ -531,8 +533,9 @@ func FormatTurnBenchSummary(result TurnBenchResult) string {
 			lines = append(lines, fmt.Sprintf("  %-18s %10s  %5.1f%%", src.Span, FormatMetric(src.TotalMs, "ms"), src.Share*100))
 		}
 	}
-	lines = append(lines, fmt.Sprintf("totals: in=%d (cached %d) out=%d | requests=%d tools=%d retries=%d reconnects=%d compactions=%d",
-		result.Totals.InputTokens, result.Totals.CachedInputTokens, result.Totals.OutputTokens,
+	lines = append(lines, fmt.Sprintf("totals: in=%d (cache-read %d, cache-write %d, uncached %d) out=%d | requests=%d tools=%d retries=%d reconnects=%d compactions=%d",
+		result.Totals.InputTokens, result.Totals.CachedInputTokens, result.Totals.CacheWriteTokens,
+		uncachedInputTokens(result.Totals), result.Totals.OutputTokens,
 		result.Totals.ModelRequests, result.Totals.ToolCalls, result.Totals.Retries,
 		result.Totals.Reconnects, result.Totals.Compactions))
 	for _, class := range sortedClasses(result.PerClass) {
@@ -545,6 +548,23 @@ func FormatTurnBenchSummary(result TurnBenchResult) string {
 	return strings.Join(lines, "\n")
 }
 
+func uncachedInputTokens(totals TurnBenchTotals) int64 {
+	remaining := totals.InputTokens
+	if remaining <= 0 {
+		return 0
+	}
+	for _, cached := range []int64{totals.CachedInputTokens, totals.CacheWriteTokens} {
+		if cached <= 0 {
+			continue
+		}
+		if cached >= remaining {
+			return 0
+		}
+		remaining -= cached
+	}
+	return remaining
+}
+
 // classTier returns the oracle tier label a class was classified into, so the
 // per-class roll-up states explicitly which oracle (if any) its "passed" count
 // is measured against.

internal/perfbench/turn_bench_test.go

@@ -74,6 +74,7 @@ func TestRunTurnBenchAggregation(t *testing.T) {
 		"t3": cannedTrace(150, 20, 1500),
 		"t4": cannedTrace(300, 10, 1000),
 	}
+	canned["t1"].Counters = append(canned["t1"].Counters, trace.Counter{Name: trace.CounterCacheWriteTokens, Value: 125})
 	cfg := TurnBenchConfig{
 		Model:      "fake-model",
 		Iterations: 1,
@@ -155,6 +156,9 @@ func TestRunTurnBenchAggregation(t *testing.T) {
 	if result.Totals.InputTokens != 5500 {
 		t.Fatalf("inputTokens = %d, want 5500", result.Totals.InputTokens)
 	}
+	if result.Totals.CacheWriteTokens != 125 {
+		t.Fatalf("cacheWriteTokens = %d, want 125", result.Totals.CacheWriteTokens)
+	}
 	if result.Totals.OutputTokens != 2750 {
 		t.Fatalf("outputTokens = %d, want 2750", result.Totals.OutputTokens)
 	}
@@ -334,6 +338,35 @@ func TestFormatTurnBenchSummaryNamesTopSources(t *testing.T) {
 	}
 }
 
+func TestUncachedInputTokensClampsInconsistentTotals(t *testing.T) {
+	tests := []struct {
+		name   string
+		totals TurnBenchTotals
+		want   int64
+	}{
+		{name: "normal", totals: TurnBenchTotals{InputTokens: 100, CachedInputTokens: 30, CacheWriteTokens: 20}, want: 50},
+		{name: "cached exceeds input", totals: TurnBenchTotals{InputTokens: 100, CachedInputTokens: 120}, want: 0},
+		{name: "write exceeds remainder", totals: TurnBenchTotals{InputTokens: 100, CachedInputTokens: 20, CacheWriteTokens: 90}, want: 0},
+		{name: "overflowing inconsistent counters", totals: TurnBenchTotals{CachedInputTokens: 1<<63 - 1, CacheWriteTokens: 1<<63 - 1}, want: 0},
+		{name: "overflowing counters with input", totals: TurnBenchTotals{InputTokens: 100, CachedInputTokens: 1<<63 - 1, CacheWriteTokens: 1<<63 - 1}, want: 0},
+		{name: "negative cache counters ignored", totals: TurnBenchTotals{InputTokens: 100, CachedInputTokens: -20, CacheWriteTokens: -10}, want: 100},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if got := uncachedInputTokens(test.totals); got != test.want {
+				t.Fatalf("uncachedInputTokens(%+v) = %d, want %d", test.totals, got, test.want)
+			}
+		})
+	}
+	summary := FormatTurnBenchSummary(TurnBenchResult{Totals: TurnBenchTotals{
+		CachedInputTokens: 1<<63 - 1,
+		CacheWriteTokens:  1<<63 - 1,
+	}})
+	if !strings.Contains(summary, "uncached 0") {
+		t.Fatalf("summary did not clamp inconsistent cache totals:\n%s", summary)
+	}
+}
+
 func TestLoadBaselineManifest(t *testing.T) {
 	path := filepath.Join("manifests", "baseline.json")
 	set, err := LoadTaskSet(path)

internal/providermodeldiscovery/discovery.go

@@ -539,7 +539,8 @@ func mergeLiveModels(provider providercatalog.Descriptor, liveModels []Model, ca
 	hasCatalog := len(byID) > 0
 	// Aggregators publish the live list as the source of truth. Keep live-only
 	// ids even when a remote catalog also loaded, instead of intersecting.
-	preferLive := providermodelcatalog.PublicLiveCatalog(provider.ID) ||
+	preferLive := provider.Local ||
+		providermodelcatalog.PublicLiveCatalog(provider.ID) ||
 		providercatalog.NormalizeID(provider.ID) == "chatgpt"
 	result := make([]Model, 0, len(liveModels))
 	for _, live := range liveModels {

internal/providermodeldiscovery/discovery_test.go

@@ -226,6 +226,29 @@ func TestMergeLiveModelsUsesLiveDefaultsAndReasoningCapability(t *testing.T) {
 	}
 }
 
+func TestMergeLocalModelsKeepsLiveOnlyEntries(t *testing.T) {
+	models := mergeLiveModels(
+		providercatalog.Descriptor{ID: "lmstudio", Local: true},
+		[]Model{
+			{ID: "qwen3-coder-30b", ContextWindow: 32768},
+			{ID: "text-embedding-local"},
+			{ID: "qwen2.5-coder-32b-instruct", Description: "live description"},
+		},
+		[]Model{{ID: "qwen2.5-coder-32b-instruct", Description: "catalog description", ToolCall: true}},
+	)
+
+	if got, want := strings.Join(modelIDs(models), ","), "qwen3-coder-30b,qwen2.5-coder-32b-instruct"; got != want {
+		t.Fatalf("models = %q, want %q", got, want)
+	}
+	for _, model := range models {
+		if model.ID == "qwen2.5-coder-32b-instruct" {
+			if model.Description != "catalog description" || !model.ToolCall {
+				t.Fatalf("merged catalog metadata = %#v", model)
+			}
+		}
+	}
+}
+
 func TestDiscoverCatalogOpenGatewayUsesLiveListWithoutKey(t *testing.T) {
 	// Catalog and live endpoints return distinct payloads so the merge must keep
 	// live-only ids that are absent from the remote catalog response.

internal/providers/openai/codex_responses.go

@@ -62,6 +62,7 @@ const (
 	responsesEventReasoningDelta    = "response.reasoning_summary_text.delta"
 	responsesEventOutputTextDone    = "response.output_text.done"
 	responsesEventFunctionArgsDelta = "response.function_call_arguments.delta"
+	responsesEventCustomInputDelta  = "response.custom_tool_call_input.delta"
 	responsesEventContentPartDone   = "response.content_part.done"
 	responsesEventOutputItemDone    = "response.output_item.done"
 	responsesEventCompleted         = "response.completed"
@@ -70,17 +71,37 @@ const (
 	responsesEventIncomplete        = "response.incomplete"
 )
 
+const applyPatchLarkGrammar = `start: begin_patch hunk+ end_patch
+begin_patch: "*** Begin Patch" LF
+end_patch: "*** End Patch" LF?
+hunk: add_hunk | delete_hunk | update_hunk
+add_hunk: "*** Add File: " filename LF add_line+
+delete_hunk: "*** Delete File: " filename LF
+update_hunk: "*** Update File: " filename LF change_move? change?
+filename: /(.+)/
+add_line: "+" /(.*)/ LF -> line
+change_move: "*** Move to: " filename LF
+change: (change_context | change_line)+ eof_line?
+change_context: ("@@" | "@@ " /(.+)/) LF
+change_line: ("+" | "-" | " ") /(.*)/ LF
+eof_line: "*** End of File" LF
+%import common.LF`
+
 // responsesRequest is the wire shape POSTed to {baseURL}/responses.
 type responsesRequest struct {
-	Model           string              `json:"model"`
-	Instructions    string              `json:"instructions"`
-	Input           []inputItem         `json:"input"`
-	Stream          bool                `json:"stream"`
-	Store           bool                `json:"store"`
-	MaxOutputTokens int                 `json:"max_output_tokens,omitempty"`
-	Tools           []responsesTool     `json:"tools,omitempty"`
-	Reasoning       *responsesReasoning `json:"reasoning,omitempty"`
-	ServiceTier     string              `json:"service_tier,omitempty"`
+	Type               string              `json:"type,omitempty"`
+	Model              string              `json:"model"`
+	Instructions       string              `json:"instructions"`
+	PreviousResponseID string              `json:"previous_response_id,omitempty"`
+	Input              []inputItem         `json:"input"`
+	Stream             bool                `json:"stream"`
+	Store              bool                `json:"store"`
+	ParallelToolCalls  bool                `json:"parallel_tool_calls"`
+	MaxOutputTokens    int                 `json:"max_output_tokens,omitempty"`
+	Tools              []responsesTool     `json:"tools,omitempty"`
+	Reasoning          *responsesReasoning `json:"reasoning,omitempty"`
+	ServiceTier        string              `json:"service_tier,omitempty"`
+	PromptCacheKey     string              `json:"prompt_cache_key,omitempty"`
 }
 
 // responsesReasoning carries the reasoning controls for the Responses API. The
@@ -105,6 +126,7 @@ type inputItem struct {
 	CallID    string `json:"call_id,omitempty"`
 	Name      string `json:"name,omitempty"`
 	Arguments string `json:"arguments,omitempty"`
+	Input     string `json:"input,omitempty"`
 	// function_call_output fields
 	Output any `json:"output,omitempty"`
 }
@@ -118,10 +140,11 @@ type contentItem struct {
 }
 
 type responsesTool struct {
-	Type        string         `json:"type"`
-	Name        string         `json:"name"`
-	Description string         `json:"description,omitempty"`
-	Parameters  map[string]any `json:"parameters,omitempty"`
+	Type        string                            `json:"type"`
+	Name        string                            `json:"name"`
+	Description string                            `json:"description,omitempty"`
+	Parameters  map[string]any                    `json:"parameters,omitempty"`
+	Format      *zeroruntime.ToolDefinitionFormat `json:"format,omitempty"`
 }
 
 // responsesEvent is the decoded form of one Responses SSE data payload.
@@ -132,6 +155,7 @@ type responsesEvent struct {
 	Delta string `json:"delta,omitempty"`
 	// item payloads (response.output_item.added / done)
 	ItemID string `json:"item_id,omitempty"`
+	CallID string `json:"call_id,omitempty"`
 	// OutputIndex is a *int so a real 0 (the first output) is distinguishable from
 	// "absent" — a plain int defaulting to 0 dropped a no-item_id call's args (M1).
 	OutputIndex *int         `json:"output_index,omitempty"`
@@ -150,6 +174,7 @@ type itemPayload struct {
 	CallID    string `json:"call_id,omitempty"`
 	Name      string `json:"name,omitempty"`
 	Arguments string `json:"arguments,omitempty"`
+	Input     string `json:"input,omitempty"`
 	Status    string `json:"status,omitempty"`
 }
 
@@ -173,7 +198,8 @@ type outputDetails struct {
 }
 
 type inputDetails struct {
-	CachedTokens int `json:"cached_tokens"`
+	CachedTokens     int `json:"cached_tokens"`
+	CacheWriteTokens int `json:"cache_write_tokens"`
 }
 
 type errorPayload struct {
@@ -187,8 +213,10 @@ type errorPayload struct {
 // response.output_item.done.
 type toolCallBuilder struct {
 	ID        string
+	CallID    string
 	Name      string
 	Arguments strings.Builder
+	Freeform  bool
 	started   bool
 	ended     bool
 }
@@ -197,15 +225,57 @@ type toolCallBuilder struct {
 // event has been emitted. It is the Responses-API equivalent of the
 // chat-completions toolState in provider.go.
 type responsesState struct {
-	toolCalls map[string]*toolCallBuilder
-	usage     *usagePayload
-	done      bool
+	toolCalls  map[string]*toolCallBuilder
+	toolOrder  []string
+	text       strings.Builder
+	responseID string
+	usage      *usagePayload
+	emitted    bool
+	done       bool
 }
 
 func newResponsesState() *responsesState {
 	return &responsesState{toolCalls: map[string]*toolCallBuilder{}}
 }
 
+func (state *responsesState) outputInputItems() []inputItem {
+	items := make([]inputItem, 0, len(state.toolOrder)+1)
+	if state.text.Len() > 0 {
+		items = append(items, inputItem{
+			Type:    "message",
+			Role:    "assistant",
+			Content: []contentItem{{Type: "output_text", Text: state.text.String()}},
+		})
+	}
+	for _, key := range state.toolOrder {
+		builder := state.toolCalls[key]
+		if builder == nil || builder.ID == "" || builder.Name == "" {
+			continue
+		}
+		callID := builder.CallID
+		if callID == "" {
+			callID = builder.ID
+		}
+		itemType := "function_call"
+		if builder.Freeform {
+			itemType = "custom_tool_call"
+		}
+		item := inputItem{
+			Type:   itemType,
+			ID:     builder.ID,
+			CallID: callID,
+			Name:   builder.Name,
+		}
+		if builder.Freeform {
+			item.Input = builder.Arguments.String()
+		} else {
+			item.Arguments = builder.Arguments.String()
+		}
+		items = append(items, item)
+	}
+	return items
+}
+
 // buildResponsesRequest converts a runtime CompletionRequest into the
 // Responses API wire format. System messages fold into a top-level
 // `instructions` field; everything else (user/assistant turns, tool
@@ -215,11 +285,14 @@ func (p *CodexProvider) buildResponsesRequest(request zeroruntime.CompletionRequ
 		return nil, errors.New("codex provider: model is required")
 	}
 	req := &responsesRequest{
-		Model:           p.inner.model,
-		Stream:          true,
-		MaxOutputTokens: p.inner.maxTokens,
+		Model:             p.inner.model,
+		Stream:            true,
+		ParallelToolCalls: true,
+		MaxOutputTokens:   p.inner.maxTokens,
+		PromptCacheKey:    request.PromptCacheKey,
 	}
 	instructions := []string{}
+	customCallIDs := map[string]string{}
 	for _, msg := range request.Messages {
 		switch msg.Role {
 		case zeroruntime.MessageRoleSystem:
@@ -234,13 +307,31 @@ func (p *CodexProvider) buildResponsesRequest(request zeroruntime.CompletionRequ
 				return nil, err
 			}
 			req.Input = append(req.Input, items...)
+			for _, call := range msg.ToolCalls {
+				if call.Freeform && call.ID != "" {
+					callID := call.ProviderCallID
+					if callID == "" {
+						callID = call.ID
+					}
+					customCallIDs[call.ID] = callID
+				}
+			}
 		case zeroruntime.MessageRoleTool:
 			if msg.ToolCallID == "" {
 				return nil, fmt.Errorf("codex provider: tool message missing tool call id")
 			}
+			outputType := "function_call_output"
+			callID := msg.ToolCallProviderID
+			if callID == "" {
+				callID = msg.ToolCallID
+			}
+			if customID, custom := customCallIDs[msg.ToolCallID]; custom {
+				outputType = "custom_tool_call_output"
+				callID = customID
+			}
 			req.Input = append(req.Input, inputItem{
-				Type:   "function_call_output",
-				CallID: msg.ToolCallID,
+				Type:   outputType,
+				CallID: callID,
 				Output: msg.Content,
 			})
 		default:
@@ -249,12 +340,25 @@ func (p *CodexProvider) buildResponsesRequest(request zeroruntime.CompletionRequ
 	}
 	req.Instructions = strings.Join(instructions, "\n\n")
 	for _, tool := range request.Tools {
-		req.Tools = append(req.Tools, responsesTool{
-			Type:        "function",
-			Name:        tool.Name,
-			Description: tool.Description,
-			Parameters:  tool.Parameters,
-		})
+		definition := tool
+		if definition.Name == "apply_patch" && definition.Type == zeroruntime.ToolDefinitionFreeform {
+			definition.Description = "The apply_patch tool edits files from a raw structured patch. Use absolute file paths in patch headers to target a granted extra write root. Do not wrap the patch in JSON."
+			definition.Format = &zeroruntime.ToolDefinitionFormat{Type: "grammar", Syntax: "lark", Definition: applyPatchLarkGrammar}
+		}
+		wireTool := responsesTool{
+			Type:        string(definition.Type),
+			Name:        definition.Name,
+			Description: definition.Description,
+			Parameters:  definition.Parameters,
+			Format:      definition.Format,
+		}
+		if wireTool.Type == "" {
+			wireTool.Type = string(zeroruntime.ToolDefinitionFunction)
+		}
+		if wireTool.Type == string(zeroruntime.ToolDefinitionFreeform) {
+			wireTool.Parameters = nil
+		}
+		req.Tools = append(req.Tools, wireTool)
 	}
 	// Codex / o-series reasoning models take the effort tier nested under
 	// `reasoning` on the Responses API (the chat-completions `reasoning_effort`
@@ -312,13 +416,26 @@ func (p *CodexProvider) assistantInputItems(msg zeroruntime.Message) ([]inputIte
 		if tc.Name == "" {
 			return nil, errors.New("codex provider: assistant tool call missing name")
 		}
-		items = append(items, inputItem{
-			Type:      "function_call",
-			ID:        tc.ID,
-			CallID:    tc.ID,
-			Name:      tc.Name,
-			Arguments: tc.Arguments,
-		})
+		callID := tc.ProviderCallID
+		if callID == "" {
+			callID = tc.ID
+		}
+		itemType := "function_call"
+		if tc.Freeform {
+			itemType = "custom_tool_call"
+		}
+		item := inputItem{
+			Type:   itemType,
+			ID:     tc.ID,
+			CallID: callID,
+			Name:   tc.Name,
+		}
+		if tc.Freeform {
+			item.Input = tc.Arguments
+		} else {
+			item.Arguments = tc.Arguments
+		}
+		items = append(items, item)
 	}
 	return items, nil
 }
@@ -479,6 +596,18 @@ func (p *CodexProvider) emitResponsesEvent(
 		state.done = true
 		return false
 	}
+	return p.emitParsedResponsesEvent(ctx, &event, state, events)
+}
+
+// emitParsedResponsesEvent converts one already-decoded Responses event into
+// zero or more runtime events. WebSocket sessions use this path to avoid
+// decoding each frame twice; the SSE path retains emitResponsesEvent above.
+func (p *CodexProvider) emitParsedResponsesEvent(
+	ctx context.Context,
+	event *responsesEvent,
+	state *responsesState,
+	events chan<- zeroruntime.StreamEvent,
+) bool {
 	if event.Type == "" {
 		// A payload that parses as JSON but carries no `type` field is a
 		// malformed Responses event — the type is the discriminator that
@@ -500,12 +629,17 @@ func (p *CodexProvider) emitResponsesEvent(
 		// the runtime. response.incomplete is folded into a length finish at
 		// scan-end (state.done stays false so the wrapper emits the
 		// StreamEventDone with FinishReasonLength).
+		if event.Response != nil && event.Response.ID != "" {
+			state.responseID = event.Response.ID
+		}
 		return true
 	case responsesEventOutputItemAdded:
-		p.handleOutputItemAdded(ctx, &event, state, events)
+		p.handleOutputItemAdded(ctx, event, state, events)
 		return true
 	case responsesEventOutputTextDelta:
 		if event.Delta != "" {
+			state.emitted = true
+			state.text.WriteString(event.Delta)
 			providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 				Type:    zeroruntime.StreamEventText,
 				Content: event.Delta,
@@ -517,6 +651,7 @@ func (p *CodexProvider) emitResponsesEvent(
 		// phase shows progress (and keeps the activity clock fresh) instead of
 		// looking like a hang. Requested via reasoning.summary="auto".
 		if event.Delta != "" {
+			state.emitted = true
 			providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 				Type:    zeroruntime.StreamEventReasoning,
 				Content: event.Delta,
@@ -524,13 +659,16 @@ func (p *CodexProvider) emitResponsesEvent(
 		}
 		return true
 	case responsesEventFunctionArgsDelta:
-		p.handleFunctionArgsDelta(ctx, &event, state, events)
+		p.handleToolInputDelta(ctx, event, state, events, false)
+		return true
+	case responsesEventCustomInputDelta:
+		p.handleToolInputDelta(ctx, event, state, events, true)
 		return true
 	case responsesEventOutputItemDone:
-		p.handleOutputItemDone(ctx, &event, state, events)
+		p.handleOutputItemDone(ctx, event, state, events)
 		return true
 	case responsesEventCompleted, responsesEventFailed:
-		return p.handleTerminalResponse(ctx, &event, state, events)
+		return p.handleTerminalResponse(ctx, event, state, events)
 	case responsesEventError:
 		message := event.Message
 		if event.Code != "" {
@@ -570,22 +708,38 @@ func (p *CodexProvider) handleOutputItemAdded(
 		return
 	}
 	switch event.Item.Type {
-	case "function_call":
+	case "function_call", "custom_tool_call":
 		key := p.toolCallKey(event)
 		if key == "" {
 			return
 		}
-		builder := &toolCallBuilder{
-			ID:   key,
-			Name: event.Item.Name,
+		builder, exists := state.toolCalls[key]
+		if !exists {
+			builder = &toolCallBuilder{ID: key, Freeform: event.Item.Type == "custom_tool_call"}
+			state.toolCalls[key] = builder
+			state.toolOrder = append(state.toolOrder, key)
+		}
+		if event.Item.CallID != "" {
+			builder.CallID = event.Item.CallID
+		}
+		if event.Item.Name != "" {
+			builder.Name = event.Item.Name
+		}
+		builder.Freeform = builder.Freeform || event.Item.Type == "custom_tool_call"
+		if builder.Freeform && event.Item.Input != "" && builder.Arguments.Len() == 0 {
+			builder.Arguments.WriteString(event.Item.Input)
+		}
+		state.emitted = true
+		if !builder.started {
+			providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
+				Type:               zeroruntime.StreamEventToolCallStart,
+				ToolCallID:         key,
+				ToolCallProviderID: builder.CallID,
+				ToolName:           builder.Name,
+				ToolCallFreeform:   builder.Freeform,
+			})
+			builder.started = true
 		}
-		state.toolCalls[key] = builder
-		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
-			Type:       zeroruntime.StreamEventToolCallStart,
-			ToolCallID: key,
-			ToolName:   event.Item.Name,
-		})
-		builder.started = true
 	}
 }
 
@@ -593,11 +747,12 @@ func (p *CodexProvider) handleOutputItemAdded(
 // to the in-flight builder and emits StreamEventToolCallDelta. The
 // Codex backend uses either `item_id` or `output_index` to attribute
 // the delta; we honor both.
-func (p *CodexProvider) handleFunctionArgsDelta(
+func (p *CodexProvider) handleToolInputDelta(
 	ctx context.Context,
 	event *responsesEvent,
 	state *responsesState,
 	events chan<- zeroruntime.StreamEvent,
+	freeform bool,
 ) {
 	key := p.toolCallKey(event)
 	if key == "" {
@@ -610,10 +765,16 @@ func (p *CodexProvider) handleFunctionArgsDelta(
 		// an intermediate proxy). Treat the delta as the start of a new
 		// tool call with no known name yet — the eventual
 		// response.output_item.done carries the final name.
-		builder = &toolCallBuilder{ID: key}
+		builder = &toolCallBuilder{ID: key, Freeform: freeform}
 		state.toolCalls[key] = builder
+		state.toolOrder = append(state.toolOrder, key)
+	}
+	builder.Freeform = builder.Freeform || freeform
+	if event.CallID != "" && builder.CallID == "" {
+		builder.CallID = event.CallID
 	}
 	builder.Arguments.WriteString(event.Delta)
+	state.emitted = true
 	providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 		Type:              zeroruntime.StreamEventToolCallDelta,
 		ToolCallID:        key,
@@ -632,7 +793,7 @@ func (p *CodexProvider) handleOutputItemDone(
 	state *responsesState,
 	events chan<- zeroruntime.StreamEvent,
 ) {
-	if event.Item == nil || event.Item.Type != "function_call" {
+	if event.Item == nil || (event.Item.Type != "function_call" && event.Item.Type != "custom_tool_call") {
 		return
 	}
 	key := p.toolCallKey(event)
@@ -641,20 +802,31 @@ func (p *CodexProvider) handleOutputItemDone(
 	}
 	builder, ok := state.toolCalls[key]
 	if !ok {
-		builder = &toolCallBuilder{ID: key}
+		builder = &toolCallBuilder{ID: key, Freeform: event.Item.Type == "custom_tool_call"}
 		state.toolCalls[key] = builder
+		state.toolOrder = append(state.toolOrder, key)
 	}
 	if event.Item.Name != "" {
 		builder.Name = event.Item.Name
 	}
-	if event.Item.Arguments != "" && builder.Arguments.Len() == 0 {
-		builder.Arguments.WriteString(event.Item.Arguments)
+	if event.Item.CallID != "" {
+		builder.CallID = event.Item.CallID
+	}
+	builder.Freeform = builder.Freeform || event.Item.Type == "custom_tool_call"
+	finalInput := event.Item.Arguments
+	if builder.Freeform {
+		finalInput = event.Item.Input
+	}
+	if finalInput != "" && builder.Arguments.Len() == 0 {
+		builder.Arguments.WriteString(finalInput)
 	}
 	if !builder.started {
 		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
-			Type:       zeroruntime.StreamEventToolCallStart,
-			ToolCallID: key,
-			ToolName:   builder.Name,
+			Type:               zeroruntime.StreamEventToolCallStart,
+			ToolCallID:         key,
+			ToolCallProviderID: builder.CallID,
+			ToolName:           builder.Name,
+			ToolCallFreeform:   builder.Freeform,
 		})
 		builder.started = true
 	}
@@ -664,6 +836,7 @@ func (p *CodexProvider) handleOutputItemDone(
 		ToolName:   builder.Name,
 	})
 	builder.ended = true
+	state.emitted = true
 }
 
 // handleTerminalResponse emits the final usage + done event for a
@@ -693,6 +866,7 @@ func (p *CodexProvider) handleTerminalResponse(
 		return false
 	}
 	state.usage = event.Response.Usage
+	state.responseID = event.Response.ID
 	if event.Response.Usage != nil {
 		usage := zeroruntime.Usage{
 			InputTokens:  event.Response.Usage.InputTokens,
@@ -700,6 +874,7 @@ func (p *CodexProvider) handleTerminalResponse(
 		}
 		if event.Response.Usage.InputTokensDetails != nil {
 			usage.CachedInputTokens = event.Response.Usage.InputTokensDetails.CachedTokens
+			usage.CacheWriteTokens = event.Response.Usage.InputTokensDetails.CacheWriteTokens
 		}
 		if event.Response.Usage.OutputTokensDetails != nil {
 			usage.ReasoningTokens = event.Response.Usage.OutputTokensDetails.ReasoningTokens
@@ -735,7 +910,7 @@ func (p *CodexProvider) handleTerminalResponse(
 		state.done = true
 		return false
 	}
-	providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone})
+	providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone, ResponseID: event.Response.ID})
 	state.done = true
 	return false
 }
@@ -754,6 +929,9 @@ func (p *CodexProvider) toolCallKey(event *responsesEvent) string {
 	if event.Item != nil && event.Item.CallID != "" {
 		return event.Item.CallID
 	}
+	if event.CallID != "" {
+		return event.CallID
+	}
 	if event.OutputIndex != nil {
 		// output_index is present (a *int distinguishes a real 0 — the first output
 		// in the response — from "absent"), so key on it even when 0 instead of

internal/providers/openai/codex_session.go

@@ -0,0 +1,421 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/url"
+	"reflect"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/coder/websocket"
+
+	"github.com/Gitlawb/zero/internal/providers/providerio"
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+const (
+	responsesWebSocketBetaHeader = "OpenAI-Beta"
+	responsesWebSocketBetaValue  = "responses_websockets=2026-02-06"
+	responsesWebSocketRequest    = "response.create"
+	previousResponseNotFoundCode = "previous_response_not_found"
+	webSocketConnectionLimitCode = "websocket_connection_limit_reached"
+	responsesWebSocketReadLimit  = 16 << 20
+)
+
+var errCodexTurnSessionBusy = errors.New("codex turn session already has an active stream")
+
+// NewCodexTurnSessionProvider enables one Responses WebSocket connection for
+// each agent run. The session falls back to the existing HTTP/SSE provider when
+// the endpoint cannot upgrade or loses its response chain.
+func NewCodexTurnSessionProvider(provider *CodexProvider, caps zeroruntime.ProviderCapabilities) zeroruntime.TurnSessionProvider {
+	return &codexTurnSessionProvider{provider: provider, caps: caps}
+}
+
+type codexTurnSessionProvider struct {
+	provider *CodexProvider
+	caps     zeroruntime.ProviderCapabilities
+}
+
+func (provider *codexTurnSessionProvider) OpenTurnSession(ctx context.Context) (zeroruntime.TurnSession, error) {
+	if provider == nil || provider.provider == nil || provider.provider.inner == nil {
+		return nil, errors.New("codex turn session requires a provider")
+	}
+	sessionCtx, cancel := context.WithCancel(ctx)
+	return &codexTurnSession{
+		ctx:         sessionCtx,
+		cancel:      cancel,
+		provider:    provider.provider,
+		prewarmDone: make(chan struct{}),
+	}, nil
+}
+
+func (provider *codexTurnSessionProvider) Capabilities() zeroruntime.ProviderCapabilities {
+	return provider.caps
+}
+
+type codexTurnSession struct {
+	ctx      context.Context
+	cancel   context.CancelFunc
+	provider *CodexProvider
+
+	mu             sync.Mutex
+	connection     *websocket.Conn
+	webSocketOff   bool
+	lastRequest    *responsesRequest
+	lastResponseID string
+	lastOutput     []inputItem
+	streamActive   bool
+
+	prewarmOnce sync.Once
+	prewarmDone chan struct{}
+	closeOnce   sync.Once
+}
+
+// Prewarm starts the WebSocket handshake without blocking prompt assembly. The
+// first Stream waits for this bounded attempt; failure simply selects HTTP/SSE.
+func (session *codexTurnSession) Prewarm(ctx context.Context) error {
+	session.prewarmOnce.Do(func() {
+		go func() {
+			defer close(session.prewarmDone)
+			prewarmCtx, cancel := context.WithTimeout(session.ctx, prewarmTimeout)
+			defer cancel()
+			recorder := trace.FromContext(ctx)
+			span := recorder.Span(trace.SpanProviderPrewarm)
+			defer span.End()
+			recorder.Counter(trace.CounterPrewarmAttempts, 1)
+			connection, err := session.provider.dialResponsesWebSocket(prewarmCtx)
+			session.mu.Lock()
+			defer session.mu.Unlock()
+			if err != nil {
+				session.webSocketOff = true
+				return
+			}
+			session.connection = connection
+		}()
+	})
+	return nil
+}
+
+func (session *codexTurnSession) Stream(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	_ = session.Prewarm(ctx)
+	select {
+	case <-session.prewarmDone:
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+
+	fullRequest, err := session.provider.buildResponsesRequest(request)
+	if err != nil {
+		return nil, err
+	}
+
+	session.mu.Lock()
+	if session.streamActive {
+		session.mu.Unlock()
+		return nil, errCodexTurnSessionBusy
+	}
+	if session.webSocketOff || session.connection == nil {
+		session.mu.Unlock()
+		trace.FromContext(ctx).Counter(trace.CounterResponsesHTTPFallback, 1)
+		return session.provider.StreamCompletion(ctx, request)
+	}
+	session.streamActive = true
+	connection := session.connection
+	wireRequest := *fullRequest
+	wireRequest.Type = responsesWebSocketRequest
+	if delta, ok := responseInputDelta(session.lastRequest, session.lastOutput, fullRequest); ok && session.lastResponseID != "" {
+		wireRequest.PreviousResponseID = session.lastResponseID
+		wireRequest.Input = delta
+		trace.FromContext(ctx).Counter(trace.CounterResponseChainReused, 1)
+	} else if session.lastRequest != nil {
+		trace.FromContext(ctx).Counter(trace.CounterResponseChainReset, 1)
+		session.lastRequest = nil
+		session.lastResponseID = ""
+		session.lastOutput = nil
+	}
+	session.mu.Unlock()
+
+	body, err := json.Marshal(wireRequest)
+	if err != nil {
+		session.releaseStream()
+		return nil, fmt.Errorf("encode codex websocket request: %w", err)
+	}
+	events := make(chan zeroruntime.StreamEvent, 16)
+	go func() {
+		defer close(events)
+		defer session.releaseStream()
+		session.streamWebSocket(ctx, connection, body, fullRequest, request, events)
+	}()
+	return events, nil
+}
+
+func (session *codexTurnSession) releaseStream() {
+	session.mu.Lock()
+	session.streamActive = false
+	session.mu.Unlock()
+}
+
+func (session *codexTurnSession) Compact(context.Context, zeroruntime.CompletionRequest) ([]zeroruntime.Message, error) {
+	return nil, zeroruntime.ErrCompactionUnsupported
+}
+
+func (session *codexTurnSession) Close() error {
+	session.closeOnce.Do(func() {
+		session.cancel()
+		session.prewarmOnce.Do(func() { close(session.prewarmDone) })
+		select {
+		case <-session.prewarmDone:
+		case <-time.After(prewarmTimeout):
+		}
+		session.mu.Lock()
+		connection := session.connection
+		session.connection = nil
+		session.webSocketOff = true
+		session.mu.Unlock()
+		if connection != nil {
+			_ = connection.CloseNow()
+		}
+	})
+	return nil
+}
+
+func responseInputDelta(previous *responsesRequest, output []inputItem, current *responsesRequest) ([]inputItem, bool) {
+	if previous == nil || current == nil || !responsesRequestPropertiesEqual(previous, current) {
+		return nil, false
+	}
+	prefix := make([]inputItem, 0, len(previous.Input)+len(output))
+	prefix = append(prefix, previous.Input...)
+	prefix = append(prefix, output...)
+	if len(current.Input) <= len(prefix) || !reflect.DeepEqual(current.Input[:len(prefix)], prefix) {
+		return nil, false
+	}
+	return append([]inputItem(nil), current.Input[len(prefix):]...), true
+}
+
+func responsesRequestPropertiesEqual(left, right *responsesRequest) bool {
+	return left.Model == right.Model &&
+		left.Instructions == right.Instructions &&
+		left.Stream == right.Stream &&
+		left.Store == right.Store &&
+		left.ParallelToolCalls == right.ParallelToolCalls &&
+		left.MaxOutputTokens == right.MaxOutputTokens &&
+		reflect.DeepEqual(left.Tools, right.Tools) &&
+		reflect.DeepEqual(left.Reasoning, right.Reasoning) &&
+		left.ServiceTier == right.ServiceTier &&
+		left.PromptCacheKey == right.PromptCacheKey
+}
+
+func (session *codexTurnSession) streamWebSocket(
+	ctx context.Context,
+	connection *websocket.Conn,
+	body []byte,
+	fullRequest *responsesRequest,
+	runtimeRequest zeroruntime.CompletionRequest,
+	events chan<- zeroruntime.StreamEvent,
+) {
+	if err := connection.Write(ctx, websocket.MessageText, body); err != nil {
+		session.disableWebSocket(connection)
+		session.forwardHTTP(ctx, runtimeRequest, events)
+		return
+	}
+
+	state := newResponsesState()
+	for {
+		readCtx := ctx
+		cancel := func() {}
+		idleTimeout := session.provider.inner.streamIdleTimeout
+		if idleTimeout > 0 {
+			readCtx, cancel = context.WithTimeout(ctx, idleTimeout)
+		}
+		messageType, data, err := connection.Read(readCtx)
+		readTimedOut := errors.Is(readCtx.Err(), context.DeadlineExceeded) && ctx.Err() == nil
+		cancel()
+		if err != nil {
+			session.disableWebSocket(connection)
+			if ctx.Err() != nil {
+				return
+			}
+			if !state.emitted {
+				session.forwardHTTP(ctx, runtimeRequest, events)
+				return
+			}
+			errorMessage := err.Error()
+			if readTimedOut {
+				errorMessage = providerio.StreamTimeoutMessage(providerio.ErrStreamIdle, idleTimeout)
+			}
+			providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
+				Type:  zeroruntime.StreamEventError,
+				Error: session.provider.redact("provider stream error: " + errorMessage),
+			})
+			return
+		}
+		if messageType != websocket.MessageText {
+			continue
+		}
+
+		var responseEvent responsesEvent
+		decodeErr := json.Unmarshal(data, &responseEvent)
+		if decodeErr == nil {
+			if (responseEvent.Code == previousResponseNotFoundCode || responseEvent.Code == webSocketConnectionLimitCode) &&
+				!state.emitted {
+				session.disableWebSocket(connection)
+				session.forwardHTTP(ctx, runtimeRequest, events)
+				return
+			}
+			if responseEvent.Type == responsesEventIncomplete {
+				session.resetResponseChain(connection)
+				providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
+					Type:         zeroruntime.StreamEventDone,
+					FinishReason: zeroruntime.FinishReasonLength,
+				})
+				return
+			}
+		}
+
+		keepReading := false
+		if decodeErr == nil {
+			keepReading = session.provider.emitParsedResponsesEvent(ctx, &responseEvent, state, events)
+		} else {
+			keepReading = session.provider.emitResponsesEvent(ctx, string(data), state, events)
+		}
+		if keepReading {
+			continue
+		}
+		if responseEvent.Type == responsesEventCompleted && responseEvent.Response != nil &&
+			responseEvent.Response.Error == nil && responseEvent.Response.Status != "failed" && state.responseID != "" {
+			session.mu.Lock()
+			if session.connection == connection && !session.webSocketOff {
+				requestCopy := *fullRequest
+				requestCopy.Input = append([]inputItem(nil), fullRequest.Input...)
+				session.lastRequest = &requestCopy
+				session.lastResponseID = state.responseID
+				session.lastOutput = state.outputInputItems()
+			}
+			session.mu.Unlock()
+		}
+		return
+	}
+}
+
+func (session *codexTurnSession) resetResponseChain(connection *websocket.Conn) {
+	session.mu.Lock()
+	defer session.mu.Unlock()
+	if session.connection != connection {
+		return
+	}
+	session.lastRequest = nil
+	session.lastResponseID = ""
+	session.lastOutput = nil
+}
+
+func (session *codexTurnSession) forwardHTTP(ctx context.Context, request zeroruntime.CompletionRequest, events chan<- zeroruntime.StreamEvent) {
+	trace.FromContext(ctx).Counter(trace.CounterResponsesHTTPFallback, 1)
+	stream, err := session.provider.StreamCompletion(ctx, request)
+	if err != nil {
+		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
+			Type:  zeroruntime.StreamEventError,
+			Error: session.provider.redact(err.Error()),
+		})
+		return
+	}
+	for event := range stream {
+		providerio.SendEvent(ctx, events, event)
+	}
+}
+
+func (session *codexTurnSession) disableWebSocket(connection *websocket.Conn) {
+	session.mu.Lock()
+	if session.connection == connection {
+		session.connection = nil
+		session.webSocketOff = true
+		session.lastRequest = nil
+		session.lastResponseID = ""
+		session.lastOutput = nil
+	}
+	session.mu.Unlock()
+	_ = connection.CloseNow()
+}
+
+func (provider *CodexProvider) dialResponsesWebSocket(ctx context.Context) (*websocket.Conn, error) {
+	endpoint, err := responsesWebSocketURL(provider.inner.endpoint)
+	if err != nil {
+		return nil, err
+	}
+	for authTry := 0; authTry < 2; authTry++ {
+		headers, err := provider.responsesWebSocketHeaders(ctx, authTry == 1)
+		if err != nil {
+			return nil, err
+		}
+		connection, response, err := websocket.Dial(ctx, endpoint, &websocket.DialOptions{
+			HTTPClient:      provider.inner.httpClient,
+			HTTPHeader:      headers,
+			CompressionMode: websocket.CompressionNoContextTakeover,
+		})
+		if err == nil {
+			connection.SetReadLimit(responsesWebSocketReadLimit)
+			return connection, nil
+		}
+		if response == nil || response.StatusCode != http.StatusUnauthorized || provider.inner.oauthResolver == nil || authTry == 1 {
+			return nil, err
+		}
+	}
+	return nil, errors.New("codex websocket authentication failed")
+}
+
+func (provider *CodexProvider) responsesWebSocketHeaders(ctx context.Context, forceRefresh bool) (http.Header, error) {
+	request, err := http.NewRequestWithContext(ctx, http.MethodGet, provider.inner.endpoint, nil)
+	if err != nil {
+		return nil, err
+	}
+	base := providerio.AuthHeaders{
+		APIKey:            provider.inner.apiKey,
+		DefaultAuthHeader: "Authorization",
+		DefaultAuthScheme: "Bearer",
+		AuthHeader:        provider.inner.authHeader,
+		AuthScheme:        provider.inner.authScheme,
+		AuthHeaderValue:   provider.inner.authHeaderValue,
+		CustomHeaders:     provider.inner.customHeaders,
+	}
+	if resolver := provider.inner.oauthResolver; resolver != nil {
+		header, value, ok, err := resolver(ctx, forceRefresh)
+		if err != nil {
+			return nil, err
+		}
+		if ok {
+			providerio.ApplyAuthHeaders(request, providerio.AuthHeaders{CustomHeaders: provider.inner.customHeaders})
+			if strings.TrimSpace(header) == "" {
+				header = "Authorization"
+			}
+			request.Header.Set(header, value)
+		} else {
+			providerio.ApplyAuthHeaders(request, base)
+		}
+	} else {
+		providerio.ApplyAuthHeaders(request, base)
+	}
+	provider.injectCodexHeaders(request)
+	request.Header.Set(responsesWebSocketBetaHeader, responsesWebSocketBetaValue)
+	return request.Header.Clone(), nil
+}
+
+func responsesWebSocketURL(endpoint string) (string, error) {
+	parsed, err := url.Parse(strings.TrimSpace(endpoint))
+	if err != nil {
+		return "", err
+	}
+	switch parsed.Scheme {
+	case "http":
+		parsed.Scheme = "ws"
+	case "https":
+		parsed.Scheme = "wss"
+	default:
+		return "", fmt.Errorf("unsupported Responses websocket URL scheme %q", parsed.Scheme)
+	}
+	return parsed.String(), nil
+}

internal/providers/openai/codex_session_test.go

@@ -0,0 +1,590 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/coder/websocket"
+
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestCodexTurnSessionChainsOnlyAppendOnlyRequests(t *testing.T) {
+	var mu sync.Mutex
+	requests := []map[string]any{}
+	betaHeader := ""
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		if !strings.EqualFold(request.Header.Get("Upgrade"), "websocket") {
+			http.Error(writer, "websocket required", http.StatusUpgradeRequired)
+			return
+		}
+		mu.Lock()
+		betaHeader = request.Header.Get(responsesWebSocketBetaHeader)
+		mu.Unlock()
+		connection, err := websocket.Accept(writer, request, nil)
+		if err != nil {
+			return
+		}
+		defer connection.CloseNow()
+		for index := 0; index < 3; index++ {
+			_, body, err := connection.Read(request.Context())
+			if err != nil {
+				return
+			}
+			var payload map[string]any
+			if err := json.Unmarshal(body, &payload); err != nil {
+				return
+			}
+			mu.Lock()
+			requests = append(requests, payload)
+			mu.Unlock()
+			switch index {
+			case 0:
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.created","response":{"id":"resp-1","status":"in_progress"}}`,
+					`{"type":"response.output_item.added","item_id":"call-1","item":{"type":"function_call","id":"call-1","call_id":"call-1","name":"read_file"}}`,
+					`{"type":"response.function_call_arguments.delta","item_id":"call-1","delta":"{}"}`,
+					`{"type":"response.output_item.done","item_id":"call-1","item":{"type":"function_call","id":"call-1","call_id":"call-1","name":"read_file","arguments":"{}"}}`,
+					`{"type":"response.completed","response":{"id":"resp-1","status":"completed"}}`,
+				)
+			case 1:
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.output_text.delta","delta":"done"}`,
+					`{"type":"response.completed","response":{"id":"resp-2","status":"completed"}}`,
+				)
+			case 2:
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.completed","response":{"id":"resp-3","status":"completed"}}`,
+				)
+			}
+		}
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	// A disabled stream watchdog must use the parent context rather than creating
+	// an immediately-expired read deadline.
+	provider.inner.streamIdleTimeout = 0
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	recorder := trace.NewRecorder("session-1", "run-1", "test")
+	recorder.Start()
+	streamContext := trace.WithContext(t.Context(), recorder)
+
+	base := zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{
+			{Role: zeroruntime.MessageRoleSystem, Content: "Use tools carefully."},
+			{Role: zeroruntime.MessageRoleUser, Content: "Read the file."},
+		},
+		Tools:          []zeroruntime.ToolDefinition{{Name: "read_file", Parameters: map[string]any{"type": "object"}}},
+		PromptCacheKey: "session-1",
+	}
+	first := collectCodexSessionEvents(t, session, base)
+	if !hasToolCallEnd(first, "call-1") {
+		t.Fatalf("first stream did not produce the tool call: %#v", first)
+	}
+
+	second := base
+	second.Messages = append(append([]zeroruntime.Message(nil), base.Messages...),
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: "call-1", Name: "read_file", Arguments: "{}"}}},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleTool, ToolCallID: "call-1", Content: "contents"},
+	)
+	secondEvents := collectCodexSessionEventsContext(t, streamContext, session, second)
+	if got := joinedText(secondEvents); got != "done" {
+		t.Fatalf("second stream text = %q, want done", got)
+	}
+
+	third := second
+	third.ReasoningEffort = "high"
+	third.Messages = append(append([]zeroruntime.Message(nil), second.Messages...),
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "done"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "Continue."},
+	)
+	collectCodexSessionEventsContext(t, streamContext, session, third)
+	finishedTrace := recorder.Finish()
+	if got := finishedTrace.Counter(trace.CounterResponseChainReused); got != 1 {
+		t.Fatalf("response-chain reuse count = %d, want 1", got)
+	}
+	if got := finishedTrace.Counter(trace.CounterResponseChainReset); got != 1 {
+		t.Fatalf("response-chain reset count = %d, want 1", got)
+	}
+
+	mu.Lock()
+	defer mu.Unlock()
+	if len(requests) != 3 {
+		t.Fatalf("websocket requests = %d, want 3", len(requests))
+	}
+	if betaHeader != responsesWebSocketBetaValue {
+		t.Fatalf("OpenAI-Beta = %q, want %q", betaHeader, responsesWebSocketBetaValue)
+	}
+	if requests[0]["type"] != responsesWebSocketRequest || requests[0]["previous_response_id"] != nil {
+		t.Fatalf("first request chaining fields = %#v", requests[0])
+	}
+	if requests[0]["parallel_tool_calls"] != true || requests[0]["prompt_cache_key"] != "session-1" {
+		t.Fatalf("first request stable fields = %#v", requests[0])
+	}
+	if requests[1]["previous_response_id"] != "resp-1" {
+		t.Fatalf("second previous_response_id = %#v", requests[1]["previous_response_id"])
+	}
+	secondInput, _ := requests[1]["input"].([]any)
+	if len(secondInput) != 1 || secondInput[0].(map[string]any)["type"] != "function_call_output" {
+		t.Fatalf("second incremental input = %#v", requests[1]["input"])
+	}
+	if requests[2]["previous_response_id"] != nil {
+		t.Fatalf("incompatible request reused response chain: %#v", requests[2])
+	}
+	thirdInput, _ := requests[2]["input"].([]any)
+	if len(thirdInput) <= 1 {
+		t.Fatalf("incompatible request did not send full input: %#v", thirdInput)
+	}
+}
+
+func TestCodexTurnSessionFallsBackWhenPreviousResponseIsMissing(t *testing.T) {
+	var mu sync.Mutex
+	webSocketRequests := 0
+	httpRequests := 0
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		if strings.EqualFold(request.Header.Get("Upgrade"), "websocket") {
+			connection, err := websocket.Accept(writer, request, nil)
+			if err != nil {
+				return
+			}
+			defer connection.CloseNow()
+			for index := 0; index < 2; index++ {
+				if _, _, err := connection.Read(request.Context()); err != nil {
+					return
+				}
+				mu.Lock()
+				webSocketRequests++
+				mu.Unlock()
+				if index == 0 {
+					writeWebSocketEvents(request.Context(), connection,
+						`{"type":"response.output_item.added","item_id":"call-1","item":{"type":"function_call","id":"call-1","call_id":"call-1","name":"read_file"}}`,
+						`{"type":"response.output_item.done","item_id":"call-1","item":{"type":"function_call","id":"call-1","call_id":"call-1","name":"read_file","arguments":"{}"}}`,
+						`{"type":"response.completed","response":{"id":"resp-1","status":"completed"}}`,
+					)
+					continue
+				}
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.error","code":"previous_response_not_found","message":"missing"}`,
+				)
+			}
+			return
+		}
+
+		mu.Lock()
+		httpRequests++
+		mu.Unlock()
+		writer.Header().Set("Content-Type", "text/event-stream")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.output_text.delta\",\"delta\":\"fallback\"}\n\n")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-http\",\"status\":\"completed\"}}\n\n")
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	base := zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Read it."}},
+		Tools:    []zeroruntime.ToolDefinition{{Name: "read_file", Parameters: map[string]any{"type": "object"}}},
+	}
+	collectCodexSessionEvents(t, session, base)
+	second := base
+	second.Messages = append(append([]zeroruntime.Message(nil), base.Messages...),
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: "call-1", Name: "read_file", Arguments: "{}"}}},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleTool, ToolCallID: "call-1", Content: "contents"},
+	)
+	events := collectCodexSessionEvents(t, session, second)
+	if got := joinedText(events); got != "fallback" {
+		t.Fatalf("fallback stream text = %q, want fallback; events=%#v", got, events)
+	}
+	for _, event := range events {
+		if event.Type == zeroruntime.StreamEventError {
+			t.Fatalf("response-chain recovery leaked an error: %#v", event)
+		}
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if webSocketRequests != 2 || httpRequests != 1 {
+		t.Fatalf("requests = websocket:%d http:%d, want 2/1", webSocketRequests, httpRequests)
+	}
+}
+
+func TestCodexTurnSessionFallsBackWhenSocketClosesBeforeOutput(t *testing.T) {
+	var mu sync.Mutex
+	httpRequests := 0
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		if strings.EqualFold(request.Header.Get("Upgrade"), "websocket") {
+			connection, err := websocket.Accept(writer, request, nil)
+			if err != nil {
+				return
+			}
+			if _, _, err := connection.Read(request.Context()); err == nil {
+				_ = connection.CloseNow()
+			}
+			return
+		}
+		mu.Lock()
+		httpRequests++
+		mu.Unlock()
+		writer.Header().Set("Content-Type", "text/event-stream")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.output_text.delta\",\"delta\":\"recovered\"}\n\n")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-http\",\"status\":\"completed\"}}\n\n")
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	recorder := trace.NewRecorder("session-fallback", "run-fallback", "test")
+	recorder.Start()
+	events := collectCodexSessionEventsContext(t, trace.WithContext(t.Context(), recorder), session, zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Hello"}},
+	})
+	if got := joinedText(events); got != "recovered" {
+		t.Fatalf("fallback text = %q, want recovered", got)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if httpRequests != 1 {
+		t.Fatalf("HTTP fallback requests = %d, want 1", httpRequests)
+	}
+	if got := recorder.Finish().Counter(trace.CounterResponsesHTTPFallback); got != 1 {
+		t.Fatalf("responses HTTP fallback counter = %d, want 1", got)
+	}
+}
+
+func TestCodexTurnSessionKeepsWebSocketAfterIncompleteResponse(t *testing.T) {
+	var mu sync.Mutex
+	webSocketRequests := 0
+	httpRequests := 0
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		if strings.EqualFold(request.Header.Get("Upgrade"), "websocket") {
+			connection, err := websocket.Accept(writer, request, nil)
+			if err != nil {
+				return
+			}
+			defer connection.CloseNow()
+			for index := 0; index < 2; index++ {
+				if _, _, err := connection.Read(request.Context()); err != nil {
+					return
+				}
+				mu.Lock()
+				webSocketRequests++
+				mu.Unlock()
+				if index == 0 {
+					writeWebSocketEvents(request.Context(), connection,
+						`{"type":"response.incomplete","response":{"id":"resp-1","status":"incomplete"}}`,
+					)
+					continue
+				}
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.output_text.delta","delta":"websocket"}`,
+					`{"type":"response.completed","response":{"id":"resp-2","status":"completed"}}`,
+				)
+			}
+			return
+		}
+
+		mu.Lock()
+		httpRequests++
+		mu.Unlock()
+		writer.Header().Set("Content-Type", "text/event-stream")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.output_text.delta\",\"delta\":\"fallback\"}\n\n")
+		_, _ = fmt.Fprint(writer, "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-http\",\"status\":\"completed\"}}\n\n")
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	request := zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Continue."}},
+	}
+	first := collectCodexSessionEvents(t, session, request)
+	if len(first) != 1 || first[0].Type != zeroruntime.StreamEventDone || first[0].FinishReason != zeroruntime.FinishReasonLength {
+		t.Fatalf("incomplete response events = %#v, want length completion", first)
+	}
+	second := collectCodexSessionEvents(t, session, request)
+	if got := joinedText(second); got != "websocket" {
+		t.Fatalf("second turn text = %q, want websocket; events=%#v", got, second)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if webSocketRequests != 2 || httpRequests != 0 {
+		t.Fatalf("requests = websocket:%d http:%d, want 2/0", webSocketRequests, httpRequests)
+	}
+}
+
+func TestCodexTurnSessionDoesNotReplayAfterVisibleOutput(t *testing.T) {
+	var mu sync.Mutex
+	httpRequests := 0
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		if strings.EqualFold(request.Header.Get("Upgrade"), "websocket") {
+			connection, err := websocket.Accept(writer, request, nil)
+			if err != nil {
+				return
+			}
+			if _, _, err := connection.Read(request.Context()); err == nil {
+				writeWebSocketEvents(request.Context(), connection,
+					`{"type":"response.output_text.delta","delta":"visible"}`,
+				)
+			}
+			_ = connection.Close(websocket.StatusInternalError, "interrupted")
+			return
+		}
+		mu.Lock()
+		httpRequests++
+		mu.Unlock()
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	stream, err := session.Stream(t.Context(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Hello"}},
+	})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	sawText, sawError := false, false
+	for event := range stream {
+		sawText = sawText || event.Type == zeroruntime.StreamEventText
+		sawError = sawError || event.Type == zeroruntime.StreamEventError
+	}
+	if !sawText || !sawError {
+		t.Fatalf("visible interrupted stream = text:%v error:%v", sawText, sawError)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if httpRequests != 0 {
+		t.Fatalf("visible interrupted stream was replayed over HTTP %d time(s)", httpRequests)
+	}
+}
+
+func TestCodexTurnSessionReportsWebSocketIdleTimeout(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		connection, err := websocket.Accept(writer, request, nil)
+		if err != nil {
+			return
+		}
+		defer connection.CloseNow()
+		if _, _, err := connection.Read(request.Context()); err != nil {
+			return
+		}
+		writeWebSocketEvents(request.Context(), connection,
+			`{"type":"response.output_text.delta","delta":"visible"}`,
+		)
+		<-request.Context().Done()
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	provider.inner.streamIdleTimeout = 200 * time.Millisecond
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	stream, err := session.Stream(t.Context(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Hello"}},
+	})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+
+	var streamError string
+	for event := range stream {
+		if event.Type == zeroruntime.StreamEventError {
+			streamError = event.Error
+			break
+		}
+	}
+	if !strings.Contains(streamError, "idle timeout after 200ms") {
+		t.Fatalf("stream error = %q, want idle-timeout detail", streamError)
+	}
+}
+
+func TestCodexTurnSessionRejectsOverlappingStreams(t *testing.T) {
+	started := make(chan struct{})
+	release := make(chan struct{})
+	var releaseOnce sync.Once
+	releaseStream := func() { releaseOnce.Do(func() { close(release) }) }
+	t.Cleanup(releaseStream)
+	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
+		connection, err := websocket.Accept(writer, request, nil)
+		if err != nil {
+			return
+		}
+		defer connection.CloseNow()
+		if _, _, err := connection.Read(request.Context()); err != nil {
+			return
+		}
+		close(started)
+		<-release
+		writeWebSocketEvents(request.Context(), connection,
+			`{"type":"response.completed","response":{"id":"resp-1","status":"completed"}}`,
+		)
+	}))
+	defer server.Close()
+
+	provider := newCodexSessionTestProvider(t, server)
+	session := openCodexSession(t, provider)
+	defer session.Close()
+	request := zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Hello"}},
+	}
+	first, err := session.Stream(t.Context(), request)
+	if err != nil {
+		t.Fatalf("first Stream: %v", err)
+	}
+	select {
+	case <-started:
+	case <-time.After(time.Second):
+		t.Fatal("first WebSocket stream did not start")
+	}
+	if _, err := session.Stream(t.Context(), request); !errors.Is(err, errCodexTurnSessionBusy) {
+		t.Fatalf("overlapping Stream error = %v, want %v", err, errCodexTurnSessionBusy)
+	}
+	releaseStream()
+	for range first {
+	}
+}
+
+func TestCodexTurnSessionHTTPFallbackRedactsSetupError(t *testing.T) {
+	const secret = "sk-secret-fallback"
+	provider, err := NewCodexProvider(CodexOptions{Options: Options{
+		APIKey:  secret,
+		BaseURL: "https://chatgpt.example/backend-api/codex",
+		Model:   "gpt-test",
+	}})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	session := &codexTurnSession{provider: provider}
+	events := make(chan zeroruntime.StreamEvent, 1)
+	session.forwardHTTP(t.Context(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRole(secret), Content: "invalid"}},
+	}, events)
+
+	select {
+	case event := <-events:
+		if event.Type != zeroruntime.StreamEventError {
+			t.Fatalf("fallback event type = %s, want error", event.Type)
+		}
+		if strings.Contains(event.Error, secret) {
+			t.Fatalf("fallback error leaked credential: %q", event.Error)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("fallback did not emit setup error")
+	}
+}
+
+func TestCodexTurnSessionCloseDoesNotStartPrewarm(t *testing.T) {
+	provider, err := NewCodexProvider(CodexOptions{Options: Options{
+		APIKey:  "test-token",
+		BaseURL: "https://chatgpt.example/backend-api/codex",
+		Model:   "gpt-test",
+	}})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	session, err := NewCodexTurnSessionProvider(provider, zeroruntime.ProviderCapabilities{}).OpenTurnSession(t.Context())
+	if err != nil {
+		t.Fatalf("OpenTurnSession: %v", err)
+	}
+	done := make(chan error, 1)
+	go func() { done <- session.Close() }()
+	select {
+	case err := <-done:
+		if err != nil {
+			t.Fatalf("Close() error = %v", err)
+		}
+	case <-time.After(200 * time.Millisecond):
+		t.Fatal("Close blocked without a prewarm attempt")
+	}
+}
+
+func newCodexSessionTestProvider(t *testing.T, server *httptest.Server) *CodexProvider {
+	t.Helper()
+	provider, err := NewCodexProvider(CodexOptions{
+		Options: Options{
+			APIKey:     "test-token",
+			BaseURL:    server.URL,
+			Model:      "gpt-test",
+			HTTPClient: server.Client(),
+		},
+		AccountID: "account-test",
+	})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	return provider
+}
+
+func openCodexSession(t *testing.T, provider *CodexProvider) zeroruntime.TurnSession {
+	t.Helper()
+	session, err := NewCodexTurnSessionProvider(provider, zeroruntime.ProviderCapabilities{}).OpenTurnSession(t.Context())
+	if err != nil {
+		t.Fatalf("OpenTurnSession: %v", err)
+	}
+	if err := session.Prewarm(t.Context()); err != nil {
+		t.Fatalf("Prewarm: %v", err)
+	}
+	return session
+}
+
+func collectCodexSessionEvents(t *testing.T, session zeroruntime.TurnSession, request zeroruntime.CompletionRequest) []zeroruntime.StreamEvent {
+	t.Helper()
+	return collectCodexSessionEventsContext(t, t.Context(), session, request)
+}
+
+func collectCodexSessionEventsContext(t *testing.T, ctx context.Context, session zeroruntime.TurnSession, request zeroruntime.CompletionRequest) []zeroruntime.StreamEvent {
+	t.Helper()
+	stream, err := session.Stream(ctx, request)
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	events := []zeroruntime.StreamEvent{}
+	for event := range stream {
+		events = append(events, event)
+		if event.Type == zeroruntime.StreamEventError {
+			t.Fatalf("stream error: %s", event.Error)
+		}
+	}
+	return events
+}
+
+func writeWebSocketEvents(ctx context.Context, connection *websocket.Conn, events ...string) {
+	for _, event := range events {
+		if err := connection.Write(ctx, websocket.MessageText, []byte(event)); err != nil {
+			return
+		}
+	}
+}
+
+func hasToolCallEnd(events []zeroruntime.StreamEvent, id string) bool {
+	for _, event := range events {
+		if event.Type == zeroruntime.StreamEventToolCallEnd && event.ToolCallID == id {
+			return true
+		}
+	}
+	return false
+}
+
+func joinedText(events []zeroruntime.StreamEvent) string {
+	var builder strings.Builder
+	for _, event := range events {
+		if event.Type == zeroruntime.StreamEventText {
+			builder.WriteString(event.Content)
+		}
+	}
+	return builder.String()
+}

internal/providers/openai/codex_terminal_test.go

@@ -2,6 +2,8 @@ package openai
 
 import (
 	"context"
+	"reflect"
+	"strings"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -25,6 +27,126 @@ func TestToolCallKeyOutputIndexZero(t *testing.T) {
 	}
 }
 
+func TestResponsesStatePreservesDistinctCallID(t *testing.T) {
+	provider := &CodexProvider{}
+	state := newResponsesState()
+	events := make(chan zeroruntime.StreamEvent, 4)
+	provider.handleOutputItemAdded(context.Background(), &responsesEvent{
+		ItemID: "item-1",
+		Item: &itemPayload{
+			Type:   "function_call",
+			ID:     "item-1",
+			CallID: "call-1",
+			Name:   "read_file",
+		},
+	}, state, events)
+	items := state.outputInputItems()
+	if len(items) != 1 || items[0].ID != "item-1" || items[0].CallID != "call-1" {
+		t.Fatalf("captured function call = %#v, want item id item-1 and call id call-1", items)
+	}
+}
+
+func TestCodexCustomApplyPatchStreamProducesFreeformCall(t *testing.T) {
+	provider := &CodexProvider{}
+	state := newResponsesState()
+	events := make(chan zeroruntime.StreamEvent, 16)
+	patch := "*** Begin Patch\n*** Add File: hello.txt\n+hello\n*** End Patch\n"
+	for _, payload := range []string{
+		`{"type":"response.output_item.added","item_id":"item-1","item":{"type":"custom_tool_call","call_id":"call-1","name":"apply_patch"}}`,
+		`{"type":"response.custom_tool_call_input.delta","item_id":"item-1","call_id":"call-1","delta":"*** Begin Patch\n*** Add File: hello.txt\n"}`,
+		`{"type":"response.custom_tool_call_input.delta","item_id":"item-1","call_id":"call-1","delta":"+hello\n*** End Patch\n"}`,
+		`{"type":"response.output_item.done","item_id":"item-1","item":{"type":"custom_tool_call","call_id":"call-1","name":"apply_patch","input":"*** Begin Patch\n*** Add File: hello.txt\n+hello\n*** End Patch\n"}}`,
+		`{"type":"response.completed","response":{"id":"resp-1","status":"completed"}}`,
+	} {
+		if keepReading := provider.emitResponsesEvent(context.Background(), payload, state, events); !keepReading && !state.done {
+			t.Fatalf("stream stopped before terminal event for %s", payload)
+		}
+	}
+	close(events)
+	collected := zeroruntime.CollectStream(context.Background(), events)
+	if len(collected.ToolCalls) != 1 {
+		t.Fatalf("tool calls = %#v, want one", collected.ToolCalls)
+	}
+	call := collected.ToolCalls[0]
+	if call.ID != "item-1" || call.ProviderCallID != "call-1" || call.Name != "apply_patch" || !call.Freeform || call.Arguments != patch {
+		t.Fatalf("custom apply_patch call = %#v", call)
+	}
+}
+
+func TestCodexApplyPatchUsesFreeformWireContract(t *testing.T) {
+	provider, err := NewCodexProvider(CodexOptions{Options: Options{
+		APIKey:  "test-token",
+		BaseURL: "https://chatgpt.example/backend-api/codex",
+		Model:   "gpt-test",
+	}})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	patch := "*** Begin Patch\n*** Add File: hello.txt\n+hello\n*** End Patch\n"
+	request, err := provider.buildResponsesRequest(zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{
+			{Role: zeroruntime.MessageRoleUser, Content: "Add the file."},
+			{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{
+				ID: "item-1", ProviderCallID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+			}}},
+			{Role: zeroruntime.MessageRoleTool, ToolCallID: "item-1", ToolCallProviderID: "call-1", Content: "Done!"},
+		},
+		Tools: []zeroruntime.ToolDefinition{{
+			Name: "apply_patch", Description: "Apply a patch.", Type: zeroruntime.ToolDefinitionFreeform,
+		}},
+	})
+	if err != nil {
+		t.Fatalf("buildResponsesRequest: %v", err)
+	}
+	if len(request.Tools) != 1 || request.Tools[0].Type != string(zeroruntime.ToolDefinitionFreeform) ||
+		request.Tools[0].Format == nil || request.Tools[0].Format.Syntax != "lark" || request.Tools[0].Parameters != nil {
+		t.Fatalf("apply_patch tool wire shape = %#v", request.Tools)
+	}
+	if !strings.Contains(request.Tools[0].Format.Definition, "*** Begin Patch") {
+		t.Fatalf("apply_patch grammar = %q", request.Tools[0].Format.Definition)
+	}
+	if !strings.Contains(request.Tools[0].Description, "absolute file paths") || !strings.Contains(request.Tools[0].Description, "granted extra write root") {
+		t.Fatalf("apply_patch description does not explain extra-root paths: %q", request.Tools[0].Description)
+	}
+	if len(request.Input) != 3 || request.Input[1].Type != "custom_tool_call" || request.Input[1].ID != "item-1" ||
+		request.Input[1].CallID != "call-1" || request.Input[1].Input != patch ||
+		request.Input[2].Type != "custom_tool_call_output" || request.Input[2].CallID != "call-1" {
+		t.Fatalf("custom apply_patch replay input = %#v", request.Input)
+	}
+}
+
+func TestCodexPreservesCustomApplyPatchFunction(t *testing.T) {
+	provider, err := NewCodexProvider(CodexOptions{Options: Options{
+		APIKey:  "test-token",
+		BaseURL: "https://chatgpt.example/backend-api/codex",
+		Model:   "gpt-test",
+	}})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	parameters := map[string]any{
+		"type":       "object",
+		"properties": map[string]any{"value": map[string]any{"type": "string"}},
+	}
+	request, err := provider.buildResponsesRequest(zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "Call the custom tool."}},
+		Tools: []zeroruntime.ToolDefinition{{
+			Name: "apply_patch", Description: "Custom integration tool.", Parameters: parameters,
+		}},
+	})
+	if err != nil {
+		t.Fatalf("buildResponsesRequest: %v", err)
+	}
+	if len(request.Tools) != 1 {
+		t.Fatalf("tools = %#v, want one", request.Tools)
+	}
+	tool := request.Tools[0]
+	if tool.Type != string(zeroruntime.ToolDefinitionFunction) || tool.Format != nil ||
+		tool.Description != "Custom integration tool." || !reflect.DeepEqual(tool.Parameters, parameters) {
+		t.Fatalf("custom apply_patch function was rewritten: %#v", tool)
+	}
+}
+
 func TestHandleTerminalResponseNilPayload(t *testing.T) {
 	p := &CodexProvider{}
 
@@ -90,11 +212,14 @@ func TestHandleTerminalResponseFailedPayloadWithoutError(t *testing.T) {
 	// A response.completed with a payload and no error remains a clean done.
 	ok := make(chan zeroruntime.StreamEvent, 8)
 	p.handleTerminalResponse(context.Background(),
-		&responsesEvent{Type: responsesEventCompleted, Response: &responsePayload{Status: "completed"}}, &responsesState{}, ok)
+		&responsesEvent{Type: responsesEventCompleted, Response: &responsePayload{ID: "resp-1", Status: "completed"}}, &responsesState{}, ok)
 	close(ok)
 	for ev := range ok {
 		if ev.Type == zeroruntime.StreamEventError {
 			t.Error("response.completed with a non-error payload must NOT emit an error")
 		}
+		if ev.Type == zeroruntime.StreamEventDone && ev.ResponseID != "resp-1" {
+			t.Errorf("done response id = %q, want resp-1", ev.ResponseID)
+		}
 	}
 }

internal/providers/openai/codex_test.go

@@ -771,7 +771,7 @@ func TestCodexProviderParsesResponsesTextDeltas(t *testing.T) {
 		`{"type":"response.output_text.delta","delta":"hello "}`,
 		`{"type":"response.output_text.delta","delta":"world"}`,
 		`{"type":"response.output_text.delta","delta":"!"}`,
-		`{"type":"response.completed","response":{"id":"resp-1","status":"completed","usage":{"input_tokens":7,"output_tokens":3,"output_tokens_details":{"reasoning_tokens":1},"input_tokens_details":{"cached_tokens":2}}}}`,
+		`{"type":"response.completed","response":{"id":"resp-1","status":"completed","usage":{"input_tokens":7,"output_tokens":3,"output_tokens_details":{"reasoning_tokens":1},"input_tokens_details":{"cached_tokens":2,"cache_write_tokens":1}}}}`,
 	)
 	defer srv.Close()
 
@@ -809,8 +809,8 @@ func TestCodexProviderParsesResponsesTextDeltas(t *testing.T) {
 	if usage == nil {
 		t.Fatal("expected a usage event, got none")
 	}
-	if usage.InputTokens != 7 || usage.OutputTokens != 3 || usage.CachedInputTokens != 2 || usage.ReasoningTokens != 1 {
-		t.Fatalf("usage = %+v, want input=7 output=3 cached=2 reasoning=1", *usage)
+	if usage.InputTokens != 7 || usage.OutputTokens != 3 || usage.CachedInputTokens != 2 || usage.CacheWriteTokens != 1 || usage.ReasoningTokens != 1 {
+		t.Fatalf("usage = %+v, want input=7 output=3 cached=2 cache-write=1 reasoning=1", *usage)
 	}
 	if !gotDone {
 		t.Fatal("expected a done event, got none")

internal/providers/openai/provider.go

@@ -357,6 +357,7 @@ func (provider *Provider) emitChunk(
 				PromptTokens:      chunk.Usage.PromptTokens,
 				CompletionTokens:  chunk.Usage.CompletionTokens,
 				CachedInputTokens: chunk.Usage.PromptTokensDetails.CachedTokens,
+				CacheWriteTokens:  chunk.Usage.PromptTokensDetails.CacheWriteTokens,
 				ReasoningTokens:   chunk.Usage.CompletionTokensDetails.ReasoningTokens,
 			},
 		})
@@ -486,7 +487,7 @@ func (provider *Provider) openAIRequest(request zeroruntime.CompletionRequest) c
 				Function: toolFunction{
 					Name:        tool.Name,
 					Description: tool.Description,
-					Parameters:  tool.Parameters,
+					Parameters:  normalizeToolParameters(tool.Parameters),
 				},
 			})
 		}
@@ -494,6 +495,21 @@ func (provider *Provider) openAIRequest(request zeroruntime.CompletionRequest) c
 	return mapped
 }
 
+// normalizeToolParameters keeps schemas accepted by strict OpenAI-compatible
+// servers such as LM Studio. No-argument tools still need an object-valued
+// parameters.properties field instead of an omitted field.
+func normalizeToolParameters(parameters map[string]any) map[string]any {
+	normalized := make(map[string]any, len(parameters)+1)
+	for key, value := range parameters {
+		normalized[key] = value
+	}
+	properties, ok := normalized["properties"].(map[string]any)
+	if !ok || properties == nil {
+		normalized["properties"] = map[string]any{}
+	}
+	return normalized
+}
+
 // promptCacheKeyDisabled reports whether the ZERO_DISABLE_PROMPT_CACHE_KEY
 // kill switch is set to a truthy value. "0" and "false" (any case) are
 // no-ops, matching how ZERO_FORMAT_ON_WRITE parses boolean flags, so an

internal/providers/openai/provider_test.go

@@ -3,6 +3,7 @@ package openai
 import (
 	"context"
 	"encoding/json"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"reflect"
@@ -105,6 +106,72 @@ func TestStreamCompletionPostsChatCompletionRequest(t *testing.T) {
 	if tool["type"] != "function" {
 		t.Fatalf("unexpected tool wrapper: %#v", tool)
 	}
+	functionDefinition := tool["function"].(map[string]any)
+	parameters := functionDefinition["parameters"].(map[string]any)
+	properties, ok := parameters["properties"].(map[string]any)
+	if !ok || len(properties) != 0 {
+		t.Fatalf("parameters.properties = %#v, want empty object", parameters["properties"])
+	}
+}
+
+// TestStreamCompletionSerializesTypedNilPropertiesAsEmptyObject locks in the
+// normalizeToolParameters fix: a tool whose parameters carry a typed-nil
+// properties map must serialize "properties":{} on the wire, never
+// "properties":null, because strict OpenAI-compatible servers (LM Studio)
+// reject the null form.
+func TestStreamCompletionSerializesTypedNilPropertiesAsEmptyObject(t *testing.T) {
+	var gotBody map[string]any
+	var gotRaw []byte
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, err := io.ReadAll(r.Body)
+		if err != nil {
+			t.Fatalf("read request body: %v", err)
+		}
+		gotRaw = raw
+		if err := json.Unmarshal(raw, &gotBody); err != nil {
+			t.Fatalf("decode request body: %v", err)
+		}
+		writeSSE(w, `{"choices":[]}`)
+		writeSSE(w, `[DONE]`)
+	}))
+	defer server.Close()
+
+	provider, err := New(Options{
+		APIKey:  "sk-secret",
+		BaseURL: server.URL + "/",
+		Model:   "gpt-test",
+	})
+	if err != nil {
+		t.Fatalf("New returned error: %v", err)
+	}
+
+	stream, err := provider.StreamCompletion(context.Background(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "hi"}},
+		Tools: []zeroruntime.ToolDefinition{{
+			Name:        "no_args",
+			Description: "Takes no arguments",
+			Parameters:  map[string]any{"type": "object", "properties": map[string]any(nil)},
+		}},
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion returned error: %v", err)
+	}
+	drain(stream)
+
+	if !strings.Contains(string(gotRaw), `"properties":{}`) {
+		t.Fatalf("request body must serialize properties as an empty object, got: %s", gotRaw)
+	}
+	if strings.Contains(string(gotRaw), `"properties":null`) {
+		t.Fatalf("request body must not serialize properties as null, got: %s", gotRaw)
+	}
+	tools := gotBody["tools"].([]any)
+	tool := tools[0].(map[string]any)
+	functionDefinition := tool["function"].(map[string]any)
+	parameters := functionDefinition["parameters"].(map[string]any)
+	properties, ok := parameters["properties"].(map[string]any)
+	if !ok || len(properties) != 0 {
+		t.Fatalf("parameters.properties = %#v, want empty object", parameters["properties"])
+	}
 }
 
 func TestNewRequiresModelButNotAPIKey(t *testing.T) {
@@ -193,14 +260,14 @@ func TestStreamCompletionAppliesCustomAuthAndHeaders(t *testing.T) {
 func TestStreamCompletionEmitsTextUsageAndDone(t *testing.T) {
 	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"choices":[{"delta":{"content":"hello "}}]}`)
-		writeSSE(w, `{"choices":[{"delta":{"content":"zero"}}],"usage":{"prompt_tokens":12,"completion_tokens":5,"prompt_tokens_details":{"cached_tokens":3}}}`)
+		writeSSE(w, `{"choices":[{"delta":{"content":"zero"}}],"usage":{"prompt_tokens":12,"completion_tokens":5,"prompt_tokens_details":{"cached_tokens":3,"cache_write_tokens":2}}}`)
 		writeSSE(w, `[DONE]`)
 	})
 
 	events := collectProviderEvents(t, provider)
 	assertEvent(t, events[0], zeroruntime.StreamEventText, "hello ")
 	assertEvent(t, events[1], zeroruntime.StreamEventText, "zero")
-	if events[2].Type != zeroruntime.StreamEventUsage || events[2].Usage.PromptTokens != 12 || events[2].Usage.CompletionTokens != 5 || events[2].Usage.CachedInputTokens != 3 {
+	if events[2].Type != zeroruntime.StreamEventUsage || events[2].Usage.PromptTokens != 12 || events[2].Usage.CompletionTokens != 5 || events[2].Usage.CachedInputTokens != 3 || events[2].Usage.CacheWriteTokens != 2 {
 		t.Fatalf("unexpected usage event: %#v", events[2])
 	}
 	if events[3].Type != zeroruntime.StreamEventDone {
@@ -1479,3 +1546,66 @@ func TestOpenAIRequestPreservesCacheablePrefixAcrossTurns(t *testing.T) {
 		t.Fatalf("wire prompt cache key must remain stable: first=%#v second=%#v", first["prompt_cache_key"], second["prompt_cache_key"])
 	}
 }
+
+// TestStreamCompletionSerializesNonMapPropertiesAsEmptyObject locks in the
+// normalizeToolParameters type-assert branch: when a tool's parameters carry a
+// `properties` field whose value is a non-map (e.g., an empty slice), it must
+// be normalized to "properties":{} on the wire, never "properties":[] or
+// "properties":null.
+func TestStreamCompletionSerializesNonMapPropertiesAsEmptyObject(t *testing.T) {
+	var gotBody map[string]any
+	var gotRaw []byte
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, err := io.ReadAll(r.Body)
+		if err != nil {
+			t.Fatalf("read request body: %v", err)
+		}
+		gotRaw = raw
+		if err := json.Unmarshal(raw, &gotBody); err != nil {
+			t.Fatalf("decode request body: %v", err)
+		}
+		writeSSE(w, `{"choices":[]}`)
+		writeSSE(w, `[DONE]`)
+	}))
+	defer server.Close()
+
+	provider, err := New(Options{
+		APIKey:  "sk-secret",
+		BaseURL: server.URL + "/",
+		Model:   "gpt-test",
+	})
+	if err != nil {
+		t.Fatalf("New returned error: %v", err)
+	}
+
+	stream, err := provider.StreamCompletion(context.Background(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "hi"}},
+		Tools: []zeroruntime.ToolDefinition{{
+			Name:        "no_args",
+			Description: "Takes no arguments",
+			Parameters:  map[string]any{"type": "object", "properties": []any{}},
+		}},
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion returned error: %v", err)
+	}
+	drain(stream)
+
+	if !strings.Contains(string(gotRaw), `"properties":{}`) {
+		t.Fatalf("request body must serialize properties as an empty object, got: %s", gotRaw)
+	}
+	if strings.Contains(string(gotRaw), `"properties":null`) {
+		t.Fatalf("request body must not serialize properties as null, got: %s", gotRaw)
+	}
+	if strings.Contains(string(gotRaw), `"properties":[]`) {
+		t.Fatalf("request body must not serialize properties as an empty array, got: %s", gotRaw)
+	}
+	tools := gotBody["tools"].([]any)
+	tool := tools[0].(map[string]any)
+	functionDefinition := tool["function"].(map[string]any)
+	parameters := functionDefinition["parameters"].(map[string]any)
+	properties, ok := parameters["properties"].(map[string]any)
+	if !ok || len(properties) != 0 {
+		t.Fatalf("parameters.properties = %#v, want empty object", parameters["properties"])
+	}
+}

internal/providers/openai/session.go

@@ -195,7 +195,12 @@ func (s *turnSession) computeFingerprint(request zeroruntime.CompletionRequest)
 func wireToolsDigest(tools []zeroruntime.ToolDefinition) string {
 	rendered := make([]string, 0, len(tools))
 	for _, tool := range tools {
-		schema, err := json.Marshal(tool.Parameters)
+		wireShape := struct {
+			Type       zeroruntime.ToolDefinitionType    `json:"type,omitempty"`
+			Parameters map[string]any                    `json:"parameters,omitempty"`
+			Format     *zeroruntime.ToolDefinitionFormat `json:"format,omitempty"`
+		}{Type: tool.Type, Parameters: tool.Parameters, Format: tool.Format}
+		schema, err := json.Marshal(wireShape)
 		if err != nil {
 			rendered = append(rendered, tool.Name+"\n"+tool.Description+"\n__non_json:"+tool.Name)
 			continue

internal/providers/openai/types.go

@@ -109,7 +109,8 @@ type usage struct {
 }
 
 type promptTokenDetails struct {
-	CachedTokens int `json:"cached_tokens"`
+	CachedTokens     int `json:"cached_tokens"`
+	CacheWriteTokens int `json:"cache_write_tokens"`
 }
 
 type completionTokenDetails struct {

internal/providers/turn_session.go

@@ -15,9 +15,27 @@ import (
 // value (e.g. "1") to enable — the same boolean idiom as ZERO_FORMAT_ON_WRITE.
 const openaiTurnSessionEnv = "ZERO_OPENAI_TURN_SESSION"
 
+// ChatGPT Responses sessions are enabled by default for their native provider.
+// Setting this to 0/false/off restores the stateless HTTP/SSE transport.
+const chatGPTTurnSessionEnv = "ZERO_CHATGPT_TURN_SESSION"
+
 func openaiTurnSessionEnabled() bool {
 	value := strings.TrimSpace(os.Getenv(openaiTurnSessionEnv))
-	return value != "" && value != "0" && !strings.EqualFold(value, "false")
+	return value != "" && !explicitEnvFalse(value)
+}
+
+func chatGPTTurnSessionEnabled() bool {
+	value := strings.TrimSpace(os.Getenv(chatGPTTurnSessionEnv))
+	return value == "" || !explicitEnvFalse(value)
+}
+
+func explicitEnvFalse(value string) bool {
+	switch strings.ToLower(strings.TrimSpace(value)) {
+	case "0", "false", "off":
+		return true
+	default:
+		return false
+	}
 }
 
 // OptimizedTurnSessions returns the gated optimized TurnSessionProvider for an
@@ -25,20 +43,33 @@ func openaiTurnSessionEnabled() bool {
 // is ineligible. Callers leave agent Options.TurnSessionProvider nil on false,
 // so the loop's default wrap keeps today's behavior byte-identical.
 //
-// Eligibility is deliberately narrow: the resolved provider kind must be
-// official OpenAI (NOT openai-compatible gateways — the constructor merges the
-// two, so this branches on the resolved kind explicitly), not the ChatGPT
-// Codex catalog, and the provider value must be the concrete *openai.Provider
-// (a fake, a Codex provider, or nil falls back to the default path safely).
-// No base-URL check on top of the kind: prewarm and fingerprint telemetry are
-// harmless against any host, and kind==openai mirrors the existing
-// official-OpenAI precedent used for prompt_cache_key.
+// Eligibility is deliberately narrow: native ChatGPT Responses uses its
+// concrete Codex provider and defaults on with a kill switch; official OpenAI
+// chat-completions keeps its existing opt-in prewarm session. Compatible
+// gateways and foreign provider values always retain the default adapter.
 func OptimizedTurnSessions(profile config.ProviderProfile, provider zeroruntime.Provider, options Options) (zeroruntime.TurnSessionProvider, bool) {
-	if !openaiTurnSessionEnabled() {
+	resolved, err := resolveProfile(profile, options)
+	if err != nil {
 		return nil, false
 	}
-	resolved, err := resolveProfile(profile, options)
-	if err != nil || resolved.providerKind != config.ProviderKindOpenAI || isCodexCatalog(profile, resolved) {
+	if isCodexCatalog(profile, resolved) {
+		if !chatGPTTurnSessionEnabled() {
+			return nil, false
+		}
+		concrete, ok := provider.(*openai.CodexProvider)
+		if !ok {
+			return nil, false
+		}
+		caps, err := resolveCapabilities(profile, options)
+		if err != nil {
+			return nil, false
+		}
+		return openai.NewCodexTurnSessionProvider(concrete, caps), true
+	}
+	if resolved.providerKind != config.ProviderKindOpenAI {
+		return nil, false
+	}
+	if !openaiTurnSessionEnabled() {
 		return nil, false
 	}
 	concrete, ok := provider.(*openai.Provider)

internal/providers/turn_session_gate_test.go

@@ -5,6 +5,7 @@ import (
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providers/openai"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -45,7 +46,7 @@ func TestOptimizedTurnSessionsDefaultOff(t *testing.T) {
 func TestOptimizedTurnSessionsFalseyValues(t *testing.T) {
 	profile := openaiEligibleProfile()
 	provider := buildGateProvider(t, profile)
-	for _, value := range []string{"0", "false", "FALSE", "False", " ", ""} {
+	for _, value := range []string{"0", "false", "FALSE", "False", "off", "OFF", " ", ""} {
 		t.Setenv(openaiTurnSessionEnv, value)
 		if _, ok := OptimizedTurnSessions(profile, provider, Options{}); ok {
 			t.Fatalf("gate enabled for falsey value %q", value)
@@ -59,6 +60,21 @@ func TestOptimizedTurnSessionsFalseyValues(t *testing.T) {
 	}
 }
 
+func TestChatGPTTurnSessionGateValues(t *testing.T) {
+	for _, value := range []string{"0", "false", "FALSE", "off", "OFF"} {
+		t.Setenv(chatGPTTurnSessionEnv, value)
+		if chatGPTTurnSessionEnabled() {
+			t.Fatalf("ChatGPT gate enabled for falsey value %q", value)
+		}
+	}
+	for _, value := range []string{"", " ", "1", "true", "on"} {
+		t.Setenv(chatGPTTurnSessionEnv, value)
+		if !chatGPTTurnSessionEnabled() {
+			t.Fatalf("ChatGPT gate disabled for default/truthy value %q", value)
+		}
+	}
+}
+
 func TestOptimizedTurnSessionsOpenAIEligible(t *testing.T) {
 	t.Setenv(openaiTurnSessionEnv, "1")
 	profile := openaiEligibleProfile()
@@ -88,13 +104,64 @@ func TestOptimizedTurnSessionsRejectsCompatible(t *testing.T) {
 }
 
 func TestOptimizedTurnSessionsRejectsCodexCatalog(t *testing.T) {
-	t.Setenv(openaiTurnSessionEnv, "1")
+	t.Setenv(chatGPTTurnSessionEnv, "0")
+	profile := openaiEligibleProfile()
+	profile.Name = "chatgpt"
+	profile.CatalogID = "chatgpt"
+	if _, ok := OptimizedTurnSessions(profile, buildChatGPTGateProvider(t, profile), Options{}); ok {
+		t.Fatal("the ChatGPT Codex catalog must honor the turn-session kill switch")
+	}
+}
+
+func TestOptimizedTurnSessionsEnablesChatGPTResponsesByDefault(t *testing.T) {
+	t.Setenv(chatGPTTurnSessionEnv, "")
+	profile := openaiEligibleProfile()
+	profile.Name = "chatgpt"
+	profile.CatalogID = "chatgpt"
+	provider := buildChatGPTGateProvider(t, profile)
+
+	turnSessions, ok := OptimizedTurnSessions(profile, provider, Options{})
+	if !ok || turnSessions == nil {
+		t.Fatal("ChatGPT Responses provider did not receive its default turn session")
+	}
+	if got := turnSessions.Capabilities().Model; got != profile.Model {
+		t.Fatalf("Capabilities().Model = %q, want %q", got, profile.Model)
+	}
+}
+
+func TestOptimizedTurnSessionsRecognizesCompatibleLabeledChatGPTProfile(t *testing.T) {
+	t.Setenv(chatGPTTurnSessionEnv, "")
+	profile := openaiEligibleProfile()
+	profile.Name = "chatgpt"
+	profile.CatalogID = "chatgpt"
+	profile.ProviderKind = config.ProviderKindOpenAICompatible
+	if turnSessions, ok := OptimizedTurnSessions(profile, buildChatGPTGateProvider(t, profile), Options{}); !ok || turnSessions == nil {
+		t.Fatal("ChatGPT catalog profile must use its native session regardless of the stored compatibility label")
+	}
+}
+
+func buildChatGPTGateProvider(t *testing.T, profile config.ProviderProfile) *openai.CodexProvider {
+	t.Helper()
+	provider, err := openai.NewCodexProvider(openai.CodexOptions{
+		Options: openai.Options{
+			APIKey:  "test-token",
+			BaseURL: "https://chatgpt.example/backend-api/codex",
+			Model:   profile.Model,
+		},
+		AccountID: "account-test",
+	})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	return provider
+}
+
+func TestOptimizedTurnSessionsRejectsNonCodexProviderForChatGPT(t *testing.T) {
+	t.Setenv(chatGPTTurnSessionEnv, "1")
 	profile := openaiEligibleProfile()
 	profile.CatalogID = "chatgpt"
-	// Build the provider from the base profile: the point under test is the
-	// catalog check, not Codex construction (which needs OAuth wiring).
 	if _, ok := OptimizedTurnSessions(profile, buildGateProvider(t, openaiEligibleProfile()), Options{}); ok {
-		t.Fatal("the ChatGPT Codex catalog must not get the optimized session")
+		t.Fatal("a ChatGPT catalog profile with a non-Codex provider must use the default session")
 	}
 }
 

internal/sandbox/apply_patch_paths_test.go

@@ -0,0 +1,96 @@
+package sandbox
+
+import (
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestApplyPatchPathBlockOnlyRejectsRelativeTraversal(t *testing.T) {
+	root := t.TempDir()
+	inside := filepath.Join(root, "main.js")
+	structured := func(path string) string {
+		return strings.Join([]string{"*** Begin Patch", "*** Update File: " + path, "@@", "-a", "+b", "*** End Patch"}, "\n")
+	}
+	for name, patch := range map[string]string{
+		"absolute inside workspace": structured(inside),
+		"relative":                  structured("main.js"),
+		"decorated markers":         "*** Begin Patch ***\n*** Update File: main.js\n@@\n-a\n+b\n*** End Patch ***",
+		"no-space marker":           "***Begin Patch\n*** Update File: main.js\n@@\n-a\n+b\n***End Patch",
+	} {
+		request := Request{ToolName: "apply_patch", WorkspaceRoot: root, SideEffect: SideEffectWrite, Args: map[string]any{"patch": patch}}
+		if block := applyPatchPathBlock(request); block != nil {
+			t.Fatalf("%s: unexpected block %+v", name, block)
+		}
+	}
+	for _, path := range []string{"../escape.js", ".."} {
+		for name, patch := range map[string]string{"canonical": structured(path), "no-space": "***Begin Patch\n*** Update File: " + path + "\n@@\n-a\n+b\n***End Patch"} {
+			request := Request{ToolName: "apply_patch", WorkspaceRoot: root, SideEffect: SideEffectWrite, Args: map[string]any{"patch": patch}}
+			block := applyPatchPathBlock(request)
+			if block == nil || block.Code != BlockOutsideWorkspace {
+				t.Fatalf("%s %q must be blocked as traversal, got %+v", name, path, block)
+			}
+		}
+	}
+}
+
+// Every marker spelling the tool applies must be classified as structured at
+// the sandbox boundary too; otherwise the boundary scans the patch as a unified
+// diff, extracts no targets, and validates nothing (fail-open).
+func TestStructuredPatchClassifierMatchesToolSpellings(t *testing.T) {
+	for _, header := range []string{"*** Begin Patch", "*** Begin Patch ***", "***Begin Patch", "  *** Begin Patch  ", "\ufeff*** Begin Patch"} {
+		patch := header + "\n*** Update File: main.js\n@@\n-a\n+b\n*** End Patch"
+		if !IsStructuredPatch(patch) {
+			t.Fatalf("%q must classify as a structured patch", header)
+		}
+		if paths := applyPatchPaths(patch); len(paths) != 1 || paths[0] != "main.js" {
+			t.Fatalf("%q: sandbox must extract the structured target, got %v", header, paths)
+		}
+	}
+	for _, header := range []string{"--- a/x", "Begin Patch", "*** Begin Patchwork", "*** Update File: x"} {
+		if IsStructuredPatch(header + "\n-a\n+b") {
+			t.Fatalf("%q must not classify as a structured patch", header)
+		}
+	}
+	if StructuredPatchMarker("*** End Patch ***") != "end" || StructuredPatchMarker("***End Patch") != "end" {
+		t.Fatal("decorated end markers must classify as end")
+	}
+}
+
+func TestApplyPatchRequestPathsCarryAbsolutePathsToScopeValidation(t *testing.T) {
+	root := t.TempDir()
+	// NewScope also grants the system temp dir, so a sibling t.TempDir() is
+	// legitimately in scope; pick a path under the filesystem root instead.
+	outside, err := filepath.Abs(filepath.Join(string(filepath.Separator), "zero-outside-workspace-test", "escape.js"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	scope, err := NewScope(root, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	inside := filepath.Join(root, "main.js")
+	structured := func(header, footer, path string) string {
+		return strings.Join([]string{header, "*** Update File: " + path, "@@", "-a", "+b", footer}, "\n")
+	}
+	for name, spelling := range map[string][2]string{"canonical": {"*** Begin Patch", "*** End Patch"}, "no-space": {"***Begin Patch", "***End Patch"}} {
+		// Failure path: the exact path the boundary parsed must be denied by the
+		// scope. structuredPatchHeaderPaths normalises separators to "/", so the
+		// parsed form is compared in slash form and then validated as-is.
+		paths := applyPatchRequestPaths(map[string]any{"patch": structured(spelling[0], spelling[1], outside)})
+		if len(paths) != 1 || paths[0] != filepath.ToSlash(outside) {
+			t.Fatalf("%s: absolute patch path must reach scope validation unchanged, got %v", name, paths)
+		}
+		if block := scope.validate(paths[0]); block == nil || block.Code != BlockOutsideWorkspace {
+			t.Fatalf("%s: scope must deny the parsed outside path %q, got %+v", name, paths[0], block)
+		}
+		// Success path: the parsed inside path must be accepted by the scope.
+		paths = applyPatchRequestPaths(map[string]any{"patch": structured(spelling[0], spelling[1], inside)})
+		if len(paths) != 1 || paths[0] != filepath.ToSlash(inside) {
+			t.Fatalf("%s: inside patch path must reach scope validation unchanged, got %v", name, paths)
+		}
+		if block := scope.validate(paths[0]); block != nil {
+			t.Fatalf("%s: scope must accept the parsed inside path %q, got %+v", name, paths[0], block)
+		}
+	}
+}

internal/sandbox/command_prefix.go

@@ -148,8 +148,10 @@ func unsafeCommandPrefix(prefix []string) bool {
 			return true
 		}
 	}
+	normalized := append([]string(nil), prefix...)
+	normalized[0] = normalizeLauncherName(normalized[0])
 	for _, banned := range bannedCommandPrefixSuggestions {
-		if sameStringSlice(prefix, banned) {
+		if sameStringSlice(prefix, banned) || sameStringSlice(normalized, banned) {
 			return true
 		}
 	}
@@ -165,25 +167,118 @@ func unsafeCommandPrefixPart(part string) bool {
 }
 
 func unsafeCommandPrefixLauncher(program string) bool {
-	program = strings.ToLower(strings.TrimSpace(program))
-	if strings.ContainsAny(program, `/\`) {
+	// A path is already refused. ":" and "~" are refused for the same reason:
+	// on Windows they introduce names that resolve to a different file than the
+	// one matched here, and no list can keep up with them. "PYTHON~1.EXE" is
+	// python.exe under an 8.3 short name, and 8.3 truncates the stem, so
+	// "POWERS~1.EXE" cannot be recognized by name at all; "python.exe::$DATA"
+	// names the same executable through its default stream. Refusing the shape
+	// costs a permission prompt on names no real command uses.
+	if strings.ContainsAny(program, `/\:~`) {
 		return true
 	}
-	switch program {
-	case "bash", "sh", "zsh", "/bin/bash", "/bin/zsh",
-		"pwsh", "powershell", "powershell.exe",
-		"env", "sudo", "doas", "su", "run0", "osascript",
+	name := normalizeLauncherName(program)
+	return bannedLauncherName(name) || bannedLauncherName(strings.TrimRight(name, "0123456789"))
+}
+
+func bannedLauncherName(name string) bool {
+	switch name {
+	case "bash", "sh", "zsh", "dash", "ash", "ksh", "csh", "tcsh", "fish", "busybox",
+		"pwsh", "powershell", "cmd", "wsl",
+		"env", "sudo", "sudoedit", "doas", "su", "run0", "osascript",
 		"command", "eval", "exec", "time",
 		"find", "xargs", "timeout", "nice", "nohup", "watch", "setsid", "stdbuf", "ionice",
-		"ssh", "make", "npm", "npx",
-		"python", "python3", "py", "pythonw", "pyw", "pypy", "pypy3",
-		"node", "perl", "ruby", "php", "lua", "deno", "bun":
+		"ssh", "make", "npm", "npx", "pnpm", "yarn", "bunx",
+		"python", "py", "pythonw", "pyw", "pypy", "uv", "uvx",
+		"node", "nodejs", "perl", "ruby", "php", "lua", "deno", "bun":
 		return true
 	default:
 		return false
 	}
 }
 
+// normalizeLauncherName reduces a program name to the launcher it actually runs
+// so the list above cannot be stepped around by spelling the same interpreter
+// differently: "python3.11", "python.exe" and "PYTHON3" all reach it as a name
+// the list matches. internal/agent's commandName already normalizes case and
+// Windows executable extensions on the allow side; this side matched raw, so
+// every versioned or .exe-suffixed launcher validated as an ordinary command
+// and persisted as a prefix grant that auto-allows every later invocation.
+//
+// A name is only ever narrowed toward an existing entry, so the failure
+// direction is an extra permission prompt, never a silent grant.
+func normalizeLauncherName(program string) string {
+	name := strings.ToLower(strings.TrimSpace(program))
+	// Windows discards trailing dots and spaces when resolving a filename, so
+	// "python." starts python.exe. Drop them before anything matches on the
+	// result, on every platform: the deny side must not depend on which OS is
+	// reading the grant, since the grants file travels with a synced home
+	// directory.
+	name = strings.TrimRight(name, ". ")
+	for _, suffix := range []string{".exe", ".cmd", ".bat", ".com", ".ps1"} {
+		if strings.HasSuffix(name, suffix) {
+			name = strings.TrimSuffix(name, suffix)
+			name = strings.TrimRight(name, ". ")
+			break
+		}
+	}
+	return trimLauncherVersion(trimLauncherABISuffix(trimLauncherBuildSuffix(name)))
+}
+
+// trimLauncherBuildSuffix drops the build-channel suffixes distributions and
+// upstreams append to an otherwise unchanged interpreter, so "python3.11-dbg",
+// "bash-static" and "pwsh-preview" are the launchers they say they are. The set
+// is closed on purpose: a general "-word" strip would also swallow
+// "node-gyp", "python3-config" and "ruby-lsp", which are ordinary tools a user
+// should still be able to approve.
+func trimLauncherBuildSuffix(name string) string {
+	for _, suffix := range []string{"-dbg", "-debug", "-static", "-nightly", "-preview", "-beta"} {
+		if strings.HasSuffix(name, suffix) {
+			return strings.TrimSuffix(name, suffix)
+		}
+	}
+	return name
+}
+
+// trimLauncherABISuffix drops CPython's ABI flags from a versioned interpreter
+// name, so "python3.7m" and "python3.6dm" reduce to their version, and the
+// free-threaded builds "python3.13t" and "python3.13td" reduce to theirs. The
+// flags are only removed when a digit sits underneath them, which leaves
+// ordinary names that happen to end in those letters alone: "sha256sum" keeps
+// its "sum", "zstd" its "td", "cat" its "t".
+func trimLauncherABISuffix(name string) string {
+	trimmed := strings.TrimRight(name, "dmut")
+	if trimmed == name || trimmed == "" {
+		return name
+	}
+	if last := trimmed[len(trimmed)-1]; last < '0' || last > '9' {
+		return name
+	}
+	return trimmed
+}
+
+// trimLauncherVersion drops a trailing "<separator><digits>" version from a
+// program name, so "python3.11" becomes "python3" and "python2.7" becomes
+// "python2". Digits without a separator are part of the name, which keeps
+// "base64", "7z" and "sha256sum" intact; the caller handles the "python3" shape
+// by also testing the digit-stripped form.
+func trimLauncherVersion(name string) string {
+	for {
+		trimmed := strings.TrimRight(name, "0123456789")
+		if trimmed == name || trimmed == "" {
+			return name
+		}
+		if !strings.HasSuffix(trimmed, ".") && !strings.HasSuffix(trimmed, "-") {
+			return name
+		}
+		next := trimmed[:len(trimmed)-1]
+		if next == "" {
+			return name
+		}
+		name = next
+	}
+}
+
 func sameStringSlice(left []string, right []string) bool {
 	if len(left) != len(right) {
 		return false

internal/sandbox/command_prefix_test.go

@@ -0,0 +1,115 @@
+package sandbox
+
+import "testing"
+
+func TestUnsafeCommandPrefixLauncherIgnoresVersionAndExtension(t *testing.T) {
+	for _, program := range []string{
+		"python3", "python", "node", "git", "sh", "bash", "sudo",
+		"python3.11", "python3.12", "python2.7", "python3.13.1",
+		"python.exe", "python3.exe", "node.exe", "git.exe", "bash.exe",
+		"PYTHON3.11", "Node.EXE",
+		"cmd", "cmd.exe", "wsl", "wsl.exe",
+		// Windows resolves these to the same executable.
+		"python.", "python3.", "python3.11.", "python.exe.", "PYTHON.",
+		"python.cmd", "python.bat", "python.com", "python.ps1",
+		"PYTHON~1.EXE", "POWERS~1.EXE", "python.exe::$DATA",
+		// Versioned and ABI-suffixed interpreters.
+		"python3.7m", "python3.6dm", "pythonw3.11", "python-3.11",
+		// Free-threaded CPython (PEP 703) ships beside the GIL build.
+		"python3.13t", "python3.13t.exe", "python3.14t", "python3.13td",
+		// Distribution and build-channel spellings of the same interpreter.
+		"nodejs", "nodejs.exe", "node-nightly", "python3-dbg", "python3.11-dbg",
+		"python3.13t-dbg", "python3-debug", "python3.11-debug", "node-beta",
+		"pwsh-preview", "bash-static", "sudoedit",
+		// Twins of npm and npx, which the list already refuses.
+		"pnpm", "yarn", "bunx", "pnpm.cmd", "yarn.exe",
+		"perl5.36", "ruby3.2", "php8", "lua5.4", "node22.exe", "bash5",
+	} {
+		if !unsafeCommandPrefix([]string{program}) {
+			t.Errorf("unsafeCommandPrefix([%q]) = false, want true", program)
+		}
+	}
+}
+
+func TestUnsafeCommandPrefixKeepsOrdinaryCommandsGrantable(t *testing.T) {
+	for _, prefix := range [][]string{
+		{"cargo", "build"},
+		{"go", "test"},
+		{"ls", "-la"},
+		{"rg", "--json"},
+		{"7z", "l"},
+		{"base64", "-d"},
+		{"kubectl", "get", "pods"},
+		{"docker", "ps"},
+		{"sha256sum", "file"},
+		{"gcc-13", "-c"},
+		{"node-gyp", "build"},
+		{"python3-config", "--includes"},
+		{"s3cmd", "ls"},
+		{"cat", "file"},
+		{"zstd", "-d"},
+		{"sqlite3", "db"},
+		{"yt-dlp", "url"},
+		{"perl-doc", "-h"},
+		{"php-fpm", "-t"},
+		{"ruby-lsp", "--version"},
+		{"lua-language-server", "--check"},
+		{"nodemon", "app.js"},
+	} {
+		if unsafeCommandPrefix(prefix) {
+			t.Errorf("unsafeCommandPrefix(%q) = true, want false", prefix)
+		}
+	}
+}
+func TestNormalizeLauncherName(t *testing.T) {
+	for _, testCase := range []struct{ in, want string }{
+		{"python3.11", "python3"},
+		{"python2.7", "python2"},
+		{"pypy3", "pypy3"},
+		{"node.exe", "node"},
+		{"POWERSHELL.EXE", "powershell"},
+		{"7z", "7z"},
+		{"base64", "base64"},
+		{"sha256sum", "sha256sum"},
+		{"python.", "python"},
+		{"python3.11.exe.", "python3"},
+		{"python3.7m", "python3"},
+		{"python3.6dm", "python3"},
+		{"python3.13t", "python3"},
+		{"python3.13td", "python3"},
+		{"python3.11-dbg", "python3"},
+		{"node-nightly", "node"},
+		{"node-gyp", "node-gyp"},
+		{"python3-config", "python3-config"},
+		{"zstd", "zstd"},
+		{"cat", "cat"},
+		{"cargo", "cargo"},
+		{"", ""},
+	} {
+		if got := normalizeLauncherName(testCase.in); got != testCase.want {
+			t.Errorf("normalizeLauncherName(%q) = %q, want %q", testCase.in, got, testCase.want)
+		}
+	}
+}
+
+// Session-scoped grants are the agent's other entry point (loop.go grants them
+// for a single turn without touching the store), so they must refuse the same
+// launcher spellings the persisted path does.
+func TestGrantCommandPrefixForSessionRefusesLauncherVariants(t *testing.T) {
+	engine := &Engine{commandPrefixes: newCommandPrefixGrantSet()}
+	engine.GrantCommandPrefixForSession("bash", []string{"cargo", "build"})
+	if _, ok := engine.LookupCommandPrefixForSession("bash", []string{"cargo", "build", "--release"}); !ok {
+		t.Fatal("CONTROL BROKEN: an ordinary session prefix grant did not match")
+	}
+	for _, prefix := range [][]string{
+		{"python3.11", "-c"},
+		{"python.exe", "-c"},
+		{"cmd", "/c"},
+	} {
+		engine.GrantCommandPrefixForSession("bash", prefix)
+		command := append(append([]string(nil), prefix...), "whatever")
+		if _, ok := engine.LookupCommandPrefixForSession("bash", command); ok {
+			t.Errorf("session grant %q matched %q, want refused", prefix, command)
+		}
+	}
+}

internal/sandbox/grants_test.go

@@ -2,12 +2,15 @@ package sandbox
 
 import (
 	"errors"
+	"fmt"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strings"
 	"testing"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/execution"
 )
 
 func TestGrantStorePersistsListsRevokesAndClears(t *testing.T) {
@@ -343,7 +346,8 @@ func TestGrantStoreMigratesExactV1GrantAndReportsOnce(t *testing.T) {
 	if err != nil {
 		t.Fatalf("read rewritten grant file: %v", err)
 	}
-	if !strings.Contains(string(raw), `"schemaVersion": 3`) || !strings.Contains(string(raw), `"policyVersion": 1`) || !strings.Contains(string(raw), `"write_file": [`) {
+	policyLine := fmt.Sprintf(`"policyVersion": %d`, execution.PolicyVersion)
+	if !strings.Contains(string(raw), `"schemaVersion": 3`) || !strings.Contains(string(raw), policyLine) || !strings.Contains(string(raw), `"write_file": [`) {
 		t.Fatalf("grant file was not rewritten as a versioned grant store:\n%s", raw)
 	}
 	backup, err := os.ReadFile(path + ".v1.backup")
@@ -406,6 +410,62 @@ func TestGrantStorePolicyChangePreservesDeniesAndInvalidatesApprovals(t *testing
 	}
 }
 
+// A prefix grant recorded before launcher-name normalization existed (a
+// versioned or .exe-suffixed interpreter) no longer validates. It must reach
+// the user through the policy migration — backup, denies preserved, one notice
+// — and never make the whole grants file unreadable, which would silently drop
+// their persisted deny grants too.
+func TestGrantStoreMigratesPreNormalizationPrefixInsteadOfRejectingTheFile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "sandbox-grants.json")
+	original := `{"schemaVersion":3,"policyVersion":1,"grants":{"bash":[{"toolName":"bash","decision":"deny","approvedAt":"2026-06-05T14:30:00Z"}]},"commandPrefixes":{"bash":[{"toolName":"bash","prefix":["python3.11","-c"],"approvedAt":"2026-06-05T14:30:00Z"},{"toolName":"bash","prefix":["cargo","build"],"approvedAt":"2026-06-05T14:30:00Z"}]}}`
+	if err := writeText(path, original); err != nil {
+		t.Fatal(err)
+	}
+	store, err := NewGrantStore(StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatal(err)
+	}
+	grants, err := store.List()
+	if err != nil || len(grants) != 1 || grants[0].Decision != GrantDeny {
+		t.Fatalf("grants after migration = %#v err=%v, want the deny preserved", grants, err)
+	}
+	prefixes, err := store.ListCommandPrefixes()
+	if err != nil || len(prefixes) != 0 {
+		t.Fatalf("prefixes after migration = %#v err=%v, want all re-approved", prefixes, err)
+	}
+	if notice, err := store.ConsumeMigrationNotice(); err != nil || !strings.Contains(notice, "invalidated 2") {
+		t.Fatalf("notice = %q err=%v", notice, err)
+	}
+	if backup, err := os.ReadFile(path + ".policy-v1.backup"); err != nil || string(backup) != original {
+		t.Fatalf("backup = %q err=%v", backup, err)
+	}
+}
+
+// The legacy-schema path is a separate migration from the policy bump above and
+// prunes per grant rather than wholesale, so a pre-normalization prefix must be
+// dropped there while the valid prefix beside it survives.
+func TestGrantStoreLegacyMigrationDropsOnlyThePreNormalizationPrefix(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "sandbox-grants.json")
+	original := `{"schemaVersion":2,"grants":{},"commandPrefixes":{"bash":[{"toolName":"bash","prefix":["python3.11","-c"],"approvedAt":"2026-06-05T14:30:00Z"},{"toolName":"bash","prefix":["git","status"],"approvedAt":"2026-06-05T14:30:00Z"}]}}`
+	if err := writeText(path, original); err != nil {
+		t.Fatal(err)
+	}
+	store, err := NewGrantStore(StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prefixes, err := store.ListCommandPrefixes()
+	if err != nil {
+		t.Fatalf("ListCommandPrefixes after legacy migration returned error: %v", err)
+	}
+	if len(prefixes) != 1 || !sameStringSlice(prefixes[0].Prefix, []string{"git", "status"}) {
+		t.Fatalf("prefixes = %#v, want only [git status]", prefixes)
+	}
+	if notice, err := store.ConsumeMigrationNotice(); err != nil || !strings.Contains(notice, "migrated 1, invalidated 1") {
+		t.Fatalf("notice = %q err=%v", notice, err)
+	}
+}
+
 func TestGrantStorePersistsCommandPrefixes(t *testing.T) {
 	store, err := NewGrantStore(StoreOptions{
 		FilePath: filepath.Join(t.TempDir(), "sandbox-grants.json"),
@@ -481,6 +541,11 @@ func TestGrantStoreRejectsUnsafeCommandPrefixes(t *testing.T) {
 		{"python", "script.py"},
 		{"./script.sh"},
 		{"git"},
+		{"python3.11", "-c"},
+		{"python.exe", "-c"},
+		{"node.exe", "-e"},
+		{"git.exe"},
+		{"cmd", "/c"},
 	} {
 		if _, err := store.GrantCommandPrefix(CommandPrefixInput{ToolName: "bash", Prefix: prefix}); err == nil {
 			t.Fatalf("GrantCommandPrefix(%#v) succeeded, want validation error", prefix)

internal/sandbox/risk.go

@@ -281,11 +281,15 @@ func applyPatchPathBlock(request Request) *pathBlock {
 	if patch == "" {
 		return nil
 	}
+	// Only relative traversal is rejected up front. Absolute paths flow through
+	// the regular workspace-scope validation below (requestPaths), which accepts
+	// one inside the workspace and denies one outside — a model that echoes the
+	// absolute path read_file showed it must not be blocked for that alone.
 	for _, path := range applyPatchPaths(patch) {
 		if path == "" || path == "/dev/null" {
 			continue
 		}
-		if filepath.IsAbs(path) || path == ".." || strings.HasPrefix(path, "../") {
+		if path == ".." || strings.HasPrefix(path, "../") {
 			return &pathBlock{
 				Code:   BlockOutsideWorkspace,
 				Path:   path,
@@ -296,8 +300,35 @@ func applyPatchPathBlock(request Request) *pathBlock {
 	return nil
 }
 
+// structuredPatchMarkerPattern is the single classifier for structured-patch
+// markers, shared by the sandbox boundary and the apply_patch tool (which
+// imports this package). It accepts the canonical "*** Begin Patch" /
+// "*** End Patch" and the decorated spellings models emit ("*** Begin Patch ***",
+// "***Begin Patch", trailing whitespace). Both sides must agree: a spelling the
+// tool would apply but the sandbox did not recognise would make the sandbox
+// scan the patch as a unified diff, extract no targets, and validate nothing.
+var structuredPatchMarkerPattern = regexp.MustCompile(`^\*{3}\s*(Begin|End) Patch\s*\**\s*$`)
+
+// StructuredPatchMarker classifies a line as the "begin" or "end" marker of a
+// structured patch, or "" when it is neither.
+func StructuredPatchMarker(line string) string {
+	match := structuredPatchMarkerPattern.FindStringSubmatch(strings.TrimSpace(line))
+	if match == nil {
+		return ""
+	}
+	return strings.ToLower(match[1])
+}
+
+// IsStructuredPatch reports whether patch opens with a structured begin marker.
+// The tool applies exactly the patches this returns true for, so the sandbox
+// extracts structured header paths for exactly the same set.
+func IsStructuredPatch(patch string) bool {
+	first, _, _ := strings.Cut(strings.TrimSpace(strings.TrimPrefix(patch, "\ufeff")), "\n")
+	return StructuredPatchMarker(first) == "begin"
+}
+
 func applyPatchPaths(patch string) []string {
-	if strings.HasPrefix(strings.TrimSpace(strings.TrimPrefix(patch, "\ufeff")), "*** Begin Patch") {
+	if IsStructuredPatch(patch) {
 		return structuredPatchHeaderPaths(patch)
 	}
 	return patchHeaderPaths(patch)

internal/streamjson/streamjson.go

@@ -94,6 +94,8 @@ type Event struct {
 	Meta              map[string]string `json:"meta,omitempty"`
 	PromptTokens      *int              `json:"promptTokens,omitempty"`
 	CompletionTokens  *int              `json:"completionTokens,omitempty"`
+	CachedInputTokens *int              `json:"cachedInputTokens,omitempty"`
+	CacheWriteTokens  *int              `json:"cacheWriteTokens,omitempty"`
 	TotalTokens       *int              `json:"totalTokens,omitempty"`
 	CostUSD           *float64          `json:"costUsd,omitempty"`
 	Text              string            `json:"text,omitempty"`

internal/tools/apply_patch.go

@@ -3,11 +3,11 @@ package tools
 import (
 	"context"
 	"fmt"
-	"os"
-	"os/exec"
 	"path/filepath"
 	"strconv"
 	"strings"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
 )
 
 type applyPatchTool struct {
@@ -16,15 +16,110 @@ type applyPatchTool struct {
 	scope         PathScope
 }
 
+func (applyPatchTool) isBuiltInApplyPatch() {}
+
+// PrepareFreeformApplyPatchArguments converts native structured-patch input
+// into the ordinary apply_patch argument map used by permission checks and tool
+// execution. Absolute patch-header paths select a granted write root and are
+// rewritten relative to that root; relative-only patches keep workspace-root
+// behavior.
+func PrepareFreeformApplyPatchArguments(tool Tool, patch string) (map[string]any, error) {
+	preparer, ok := tool.(interface {
+		prepareFreeformArguments(string) (map[string]any, error)
+	})
+	if !ok {
+		return nil, fmt.Errorf("tool is not Zero's built-in apply_patch")
+	}
+	return preparer.prepareFreeformArguments(patch)
+}
+
+func (tool applyPatchTool) prepareFreeformArguments(patch string) (map[string]any, error) {
+	args := map[string]any{"patch": patch}
+	if !isStructuredPatch(patch) {
+		return args, nil
+	}
+
+	lines := strings.Split(strings.ReplaceAll(patch, "\r\n", "\n"), "\n")
+	selectedRoot := ""
+	hasRelativeHeader := false
+	for index, line := range lines {
+		prefix := ""
+		switch {
+		case strings.HasPrefix(line, structuredAddFile):
+			prefix = structuredAddFile
+		case strings.HasPrefix(line, structuredDeleteFile):
+			prefix = structuredDeleteFile
+		case strings.HasPrefix(line, structuredUpdateFile):
+			prefix = structuredUpdateFile
+		case strings.HasPrefix(line, structuredMoveTo):
+			prefix = structuredMoveTo
+		default:
+			continue
+		}
+		path := strings.TrimSpace(strings.TrimPrefix(line, prefix))
+		if !filepath.IsAbs(path) {
+			hasRelativeHeader = true
+			continue
+		}
+		root, relative, err := tool.freeformPatchRoot(path)
+		if err != nil {
+			return nil, err
+		}
+		if selectedRoot != "" && selectedRoot != root {
+			return nil, fmt.Errorf("freeform patch spans multiple write roots")
+		}
+		selectedRoot = root
+		lines[index] = prefix + filepath.ToSlash(relative)
+	}
+	if selectedRoot != "" {
+		roots, err := scopedRoots(tool.workspaceRoot, tool.scope)
+		if err != nil {
+			return nil, err
+		}
+		if hasRelativeHeader && selectedRoot != roots[0] {
+			return nil, fmt.Errorf("freeform patch mixes workspace-relative paths with an extra write root")
+		}
+		args["cwd"] = selectedRoot
+		args["patch"] = strings.Join(lines, "\n")
+	}
+	return args, nil
+}
+
+func (tool applyPatchTool) freeformPatchRoot(path string) (string, string, error) {
+	roots, err := scopedRoots(tool.workspaceRoot, tool.scope)
+	if err != nil {
+		return "", "", err
+	}
+	var firstErr error
+	for _, root := range roots {
+		candidate := sandbox.NormalizePrefixForRoot(path, root)
+		candidate, err = filepath.Abs(candidate)
+		if err != nil {
+			if firstErr == nil {
+				firstErr = err
+			}
+			continue
+		}
+		relative, err := filepath.Rel(root, candidate)
+		if err == nil && relative != ".." && !strings.HasPrefix(relative, ".."+string(filepath.Separator)) && !filepath.IsAbs(relative) {
+			return root, filepath.ToSlash(relative), nil
+		}
+		if firstErr == nil {
+			firstErr = outsideWorkspaceError(path)
+		}
+	}
+	return "", "", firstErr
+}
+
 func NewScopedApplyPatchTool(workspaceRoot string, scope PathScope) Tool {
 	return applyPatchTool{
 		baseTool: baseTool{
 			name:        "apply_patch",
-			description: "Apply a patch inside the workspace or an explicitly granted extra write root.",
+			description: "Apply a multi-hunk or multi-file patch inside the workspace (or a granted extra write root). Structured format:\n*** Begin Patch\n*** Update File: src/app.js\n@@\n unchanged context line\n-removed line\n+added line\n*** End Patch\nUse \"*** Add File: path\" with \"+\" lines to create a file and \"*** Delete File: path\" to remove one; several sections may follow each other. A unified diff (---/+++ headers with a/ b/ prefixes, @@ hunks) is also accepted and applied in-process. Paths are workspace-relative; absolute paths inside the workspace are fine. For a single targeted change, edit_file is simpler.",
 			parameters: Schema{
 				Type: "object",
 				Properties: map[string]PropertySchema{
-					"patch": {Type: "string", Description: "A unified diff or a structured *** Begin Patch patch to apply."},
+					"patch": {Type: "string", Description: "The structured (*** Begin Patch) or unified-diff patch text."},
 					"cwd":   {Type: "string", Description: "Directory where the patch should be applied. Relative paths stay in the workspace; use an absolute path to target a granted extra write root. Defaults to workspace root.", Default: "."},
 				},
 				Required:             []string{"patch"},
@@ -59,177 +154,14 @@ func (tool applyPatchTool) RunWithOptions(ctx context.Context, args map[string]a
 	if isStructuredPatch(patch) {
 		return tool.runStructuredPatch(applyRoot, relativeRoot, patch, options)
 	}
-	if err := validatePatchPaths(applyRoot, patch); err != nil {
-		return errorResult("Error applying patch: " + err.Error())
-	}
-
-	tempFile, err := os.CreateTemp("", "zero-patch-*.patch")
+	// Unified diffs are translated into the same operations and applied by
+	// the same os.Root engine, so neither format opens a target by pathname
+	// after validation (no check-to-use window) and git is not needed.
+	operations, err := parseUnifiedPatch(patch)
 	if err != nil {
 		return errorResult("Error applying patch: " + err.Error())
 	}
-	patchPath := tempFile.Name()
-	defer func() {
-		_ = os.Remove(patchPath)
-	}()
-	if _, err := tempFile.WriteString(patch); err != nil {
-		_ = tempFile.Close()
-		return errorResult("Error applying patch: " + err.Error())
-	}
-	if err := tempFile.Close(); err != nil {
-		return errorResult("Error applying patch: " + err.Error())
-	}
-
-	if err := recheckPatchWriteTargets(applyRoot, patch); err != nil {
-		return errorResult("Error applying patch: " + err.Error())
-	}
-	var createdTargets []string
-	var fullySuppliedTargets []string
-	wholeBefore := map[string]bool{}
-	if options.FileTracker != nil {
-		createdTargets = missingPatchTargets(applyRoot, patch)
-		fullySuppliedTargets = completeCreatedPatchTargets(applyRoot, patch)
-		for _, path := range patchHeaderPaths(patch) {
-			if path == "" || path == "/dev/null" {
-				continue
-			}
-			if absolute, _, rerr := resolveWorkspaceTargetPath(applyRoot, path); rerr == nil {
-				wholeBefore[absolute] = options.FileTracker.SeenWhole(absolute)
-			}
-		}
-	}
-
-	command := exec.CommandContext(ctx, "git", "apply", "--whitespace=nowarn", patchPath)
-	command.Dir = applyRoot
-	output, err := command.CombinedOutput()
-	if err != nil {
-		message := strings.TrimSpace(string(output))
-		if message == "" {
-			message = err.Error()
-		}
-		return errorResult("Error applying patch: " + message)
-	}
-
-	summary := "Patch applied successfully."
-	if relativeRoot != "." {
-		summary = "Patch applied successfully in " + relativeRoot + "."
-	}
-	result := okResult(summary)
-	result.ChangedFiles = changedFilesFromPatch(relativeRoot, patch)
-	result.Display = Display{Summary: summary, Kind: "diff", Preview: capPreviewDiff(patch)}
-	fullySupplied := make(map[string]bool, len(fullySuppliedTargets))
-	for _, absolute := range fullySuppliedTargets {
-		fullySupplied[absolute] = true
-	}
-	// Re-baseline files changed by this tool. When the model had already seen the
-	// whole input (or supplied a complete new file), the exact patch plus that
-	// input determines the whole output, so a follow-up edit needs no wasted read.
-	// Partial observations remain conservative and are cleared by Record.
-	for _, changed := range result.ChangedFiles {
-		if absolute, _, rerr := resolveScopedPath(tool.workspaceRoot, tool.scope, changed); rerr == nil {
-			content, readErr := os.ReadFile(absolute)
-			if readErr != nil {
-				options.FileTracker.Forget(absolute)
-				continue
-			}
-			info, _ := os.Stat(absolute)
-			wasWhole := wholeBefore[absolute] || fullySupplied[absolute]
-			options.FileTracker.Record(absolute, content, info)
-			if wasWhole {
-				lines := lineCount(string(content))
-				options.FileTracker.RecordSeenRange(absolute, 1, lines, lines)
-			}
-		}
-	}
-	recordCreatedPatchTargets(options.FileTracker, createdTargets)
-	return result
-}
-
-func missingPatchTargets(root string, patch string) []string {
-	seen := map[string]bool{}
-	var missing []string
-	for _, path := range patchHeaderPaths(patch) {
-		if path == "" || path == "/dev/null" {
-			continue
-		}
-		absolute, _, err := resolveWorkspaceTargetPath(root, path)
-		if err != nil || seen[absolute] {
-			continue
-		}
-		seen[absolute] = true
-		if _, err := os.Stat(absolute); os.IsNotExist(err) {
-			missing = append(missing, absolute)
-		}
-	}
-	return missing
-}
-
-// completeCreatedPatchTargets returns only files whose full contents are
-// supplied by a /dev/null creation patch. A missing rename/copy destination is
-// created by git too, but its bytes come from an unread source and must not gain
-// whole-file observation credit.
-func completeCreatedPatchTargets(root string, patch string) []string {
-	seen := map[string]bool{}
-	var created []string
-	oldRemaining, newRemaining := 0, 0
-	inHunk := false
-	fromDevNull := false
-	for _, line := range strings.Split(strings.ReplaceAll(patch, "\r\n", "\n"), "\n") {
-		if inHunk && (oldRemaining > 0 || newRemaining > 0) {
-			switch {
-			case strings.HasPrefix(line, "-"):
-				oldRemaining--
-			case strings.HasPrefix(line, "+"):
-				newRemaining--
-			case strings.HasPrefix(line, "\\"):
-			default:
-				oldRemaining--
-				newRemaining--
-			}
-			continue
-		}
-		inHunk = false
-		switch {
-		case strings.HasPrefix(line, "diff --git "):
-			fromDevNull = false
-		case strings.HasPrefix(line, "@@"):
-			oldRemaining, newRemaining = parseHunkCounts(line)
-			inHunk = oldRemaining > 0 || newRemaining > 0
-		case strings.HasPrefix(line, "--- "):
-			fromDevNull = patchFileHeaderPath(line) == "/dev/null"
-		case strings.HasPrefix(line, "+++ "):
-			path := patchFileHeaderPath(line)
-			if !fromDevNull || path == "" || path == "/dev/null" {
-				fromDevNull = false
-				continue
-			}
-			fromDevNull = false
-			absolute, _, err := resolveWorkspaceTargetPath(root, stripPatchPrefix(path))
-			if err != nil || seen[absolute] {
-				continue
-			}
-			if _, err := os.Stat(absolute); !os.IsNotExist(err) {
-				continue
-			}
-			seen[absolute] = true
-			created = append(created, absolute)
-		}
-	}
-	return created
-}
-
-func recordCreatedPatchTargets(tracker *FileTracker, missingBefore []string) {
-	if tracker == nil {
-		return
-	}
-	for _, absolute := range missingBefore {
-		if _, err := os.Stat(absolute); err != nil {
-			continue
-		}
-		if resolved, err := filepath.EvalSymlinks(absolute); err == nil {
-			absolute = resolved
-		}
-		tracker.RecordCreated(absolute)
-	}
+	return applyPatchOperations(applyRoot, relativeRoot, operations, options)
 }
 
 // changedFilesFromPatch extracts the unique, WORKSPACE-relative paths a patch
@@ -260,27 +192,35 @@ func changedFilesFromPatch(relativeRoot string, patch string) []string {
 	return paths
 }
 
+// normalizePatchPathForRoot resolves platform-level symlinks (macOS /var ->
+// /private/var) in the prefix of an absolute patch path that lies outside the
+// apply root, so a path the model copied from read_file compares equal to the
+// symlink-resolved root. Relative paths are returned unchanged.
+func normalizePatchPathForRoot(root string, path string) string {
+	if !filepath.IsAbs(path) {
+		return path
+	}
+	resolvedRoot, err := filepath.Abs(root)
+	if err != nil {
+		return path
+	}
+	if evaluated, err := filepath.EvalSymlinks(resolvedRoot); err == nil {
+		resolvedRoot = evaluated
+	}
+	return sandbox.NormalizePrefixForRoot(path, resolvedRoot)
+}
+
 func validatePatchPaths(root string, patch string) error {
 	for _, path := range patchHeaderPaths(patch) {
 		if path == "" || path == "/dev/null" {
 			continue
 		}
-		if filepath.IsAbs(path) || path == ".." || strings.HasPrefix(path, "../") {
+		// Relative traversal is rejected here; an absolute path is checked by
+		// resolveWorkspaceTargetPath, which only accepts one inside the root.
+		if path == ".." || strings.HasPrefix(path, "../") {
 			return fmt.Errorf("patch path %q must stay inside the workspace", path)
 		}
-		if _, _, err := resolveWorkspaceTargetPath(root, path); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-func recheckPatchWriteTargets(root string, patch string) error {
-	for _, path := range patchHeaderPaths(patch) {
-		if path == "" || path == "/dev/null" {
-			continue
-		}
-		if err := recheckWorkspaceWriteTarget(root, path); err != nil {
+		if _, _, err := resolveWorkspaceTargetPath(root, normalizePatchPathForRoot(root, path)); err != nil {
 			return err
 		}
 	}
@@ -351,7 +291,7 @@ func patchFileHeaderPath(line string) string {
 // tokens. Scanning the whole line would let a crafted heading like
 // "@@ -1,1 +1,1 @@ +1,999999" overwrite the real count, keep the parser stuck in
 // hunk mode, and swallow later "--- "/"+++ " file headers so they escape
-// validatePatchPaths / recheckPatchWriteTargets — a workspace-confinement bypass.
+// validatePatchPaths — a workspace-confinement bypass.
 func parseHunkCounts(line string) (int, int) {
 	_, rest, ok := strings.Cut(line, "@@")
 	if !ok {

internal/tools/apply_patch_tolerance_test.go

@@ -0,0 +1,712 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// Models routinely decorate the structured markers ("*** Begin Patch ***") and
+// write unified-diff ranges after "@@"; both used to fail on the first line and
+// pushed the model into whole-file rewrites.
+func TestApplyPatchToleratesDecoratedMarkersAndRangeHeaders(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "hello.txt"), "hello\nold\nbye\n")
+
+	patch := strings.Join([]string{
+		"*** Begin Patch ***",
+		"*** Update File: hello.txt",
+		"@@ -1,3 +1,3 @@",
+		" hello",
+		"-old",
+		"+new",
+		"*** End Patch ***",
+		"",
+	}, "\n")
+
+	if !isStructuredPatch(patch) {
+		t.Fatal("decorated begin marker must still be recognised as a structured patch")
+	}
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("decorated structured patch should apply, got %s: %s", result.Status, result.Output)
+	}
+	content, err := os.ReadFile(filepath.Join(root, "hello.txt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := strings.ReplaceAll(string(content), "\r\n", "\n"); got != "hello\nnew\nbye\n" {
+		t.Fatalf("patched content = %q", got)
+	}
+}
+
+func TestStructuredHunkAnchorKeepsHeadingAndPlainContext(t *testing.T) {
+	cases := map[string]string{
+		"-12,4 +12,6":                  "",
+		"-12 +12":                      "",
+		"-12,4 +12,6 @@":               "",
+		"-12,4 +12,6 @@ func main() {": "func main() {",
+		"func main() {":                "func main() {",
+		"":                             "",
+	}
+	for input, want := range cases {
+		if got := structuredHunkAnchor(input); got != want {
+			t.Fatalf("structuredHunkAnchor(%q) = %q, want %q", input, got, want)
+		}
+	}
+}
+
+func TestStructuredPatchMarkerSpellings(t *testing.T) {
+	begin := []string{"*** Begin Patch", "*** Begin Patch ***", "***Begin Patch", "  *** Begin Patch  ", "*** Begin Patch **"}
+	for _, line := range begin {
+		if structuredPatchMarker(line) != "begin" {
+			t.Fatalf("%q must read as the begin marker", line)
+		}
+	}
+	if structuredPatchMarker("*** End Patch ***") != "end" {
+		t.Fatal("decorated end marker must read as the end marker")
+	}
+	for _, line := range []string{"*** Update File: x", "Begin Patch", "*** Begin Patchwork"} {
+		if structuredPatchMarker(line) != "" {
+			t.Fatalf("%q must not read as a marker", line)
+		}
+	}
+}
+
+func TestStructuredPatchHeaderErrorCarriesFormatHint(t *testing.T) {
+	_, err := parseStructuredPatch("--- a/x\n+++ b/x\n")
+	if err == nil || !strings.Contains(err.Error(), "*** Update File: path") {
+		t.Fatalf("header error should teach the format, got %v", err)
+	}
+}
+
+func TestApplyPatchAcceptsAbsoluteInWorkspacePaths(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "hello.txt"), "hello\nold\n")
+	absolute := filepath.Join(root, "hello.txt")
+
+	structured := strings.Join([]string{
+		"*** Begin Patch",
+		"*** Update File: " + absolute,
+		"@@",
+		" hello",
+		"-old",
+		"+structured",
+		"*** End Patch",
+		"",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": structured})
+	if result.Status != StatusOK {
+		t.Fatalf("absolute in-workspace structured path should apply, got %s: %s", result.Status, result.Output)
+	}
+	content, _ := os.ReadFile(absolute)
+	if got := strings.ReplaceAll(string(content), "\r\n", "\n"); got != "hello\nstructured\n" {
+		t.Fatalf("structured content = %q", got)
+	}
+	if len(result.ChangedFiles) != 1 || result.ChangedFiles[0] != "hello.txt" {
+		t.Fatalf("changed files should be workspace-relative, got %v", result.ChangedFiles)
+	}
+
+	unified := strings.Join([]string{
+		"--- " + absolute,
+		"+++ " + absolute,
+		"@@ -1,2 +1,2 @@",
+		" hello",
+		"-structured",
+		"+unified",
+		"",
+	}, "\n")
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": unified})
+	if result.Status != StatusOK {
+		t.Fatalf("absolute in-workspace unified path should apply, got %s: %s", result.Status, result.Output)
+	}
+	content, _ = os.ReadFile(absolute)
+	if got := strings.ReplaceAll(string(content), "\r\n", "\n"); got != "hello\nunified\n" {
+		t.Fatalf("unified content = %q", got)
+	}
+}
+
+func TestApplyPatchStillRejectsAbsolutePathsOutsideWorkspace(t *testing.T) {
+	root := t.TempDir()
+	outside := filepath.Join(t.TempDir(), "escape.txt")
+	writeTestFile(t, outside, "old\n")
+
+	for name, patch := range map[string]string{
+		"structured": strings.Join([]string{"*** Begin Patch", "*** Update File: " + outside, "@@", "-old", "+new", "*** End Patch", ""}, "\n"),
+		"unified":    strings.Join([]string{"--- " + outside, "+++ " + outside, "@@ -1 +1 @@", "-old", "+new", ""}, "\n"),
+	} {
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+		if result.Status == StatusOK {
+			t.Fatalf("%s patch outside the workspace must be rejected", name)
+		}
+		if content, _ := os.ReadFile(outside); string(content) != "old\n" {
+			t.Fatalf("%s patch must not touch a file outside the workspace", name)
+		}
+	}
+}
+
+// Unified diffs are applied in-process through the same os.Root engine as
+// structured patches: git C-quoted paths with whitespace resolve to one file,
+// and the same header outside the workspace is refused before anything is
+// written.
+func TestUnifiedPatchQuotedWhitespacePaths(t *testing.T) {
+	root := t.TempDir()
+	dir := filepath.Join(root, "work tree")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	target := filepath.Join(dir, "file.txt")
+	writeTestFile(t, target, "hello\nold\n")
+
+	quoted := func(path string) string { return strconv.Quote(path) }
+	inside := strings.Join([]string{
+		"diff --git " + quoted("a/"+filepath.ToSlash(target)) + " " + quoted("b/"+filepath.ToSlash(target)),
+		"--- " + quoted(target),
+		"+++ " + quoted(target),
+		"@@ -1,2 +1,2 @@",
+		" hello",
+		"-old",
+		"+new",
+		"",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": inside})
+	if result.Status != StatusOK {
+		t.Fatalf("quoted whitespace path inside the workspace should apply, got %s: %s", result.Status, result.Output)
+	}
+	content, _ := os.ReadFile(target)
+	if got := strings.ReplaceAll(string(content), "\r\n", "\n"); got != "hello\nnew\n" {
+		t.Fatalf("content = %q", got)
+	}
+	if len(result.ChangedFiles) != 1 || result.ChangedFiles[0] != "work tree/file.txt" {
+		t.Fatalf("changed files = %v", result.ChangedFiles)
+	}
+
+	outsideDir := filepath.Join(t.TempDir(), "other tree")
+	if err := os.MkdirAll(outsideDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	outside := filepath.Join(outsideDir, "file.txt")
+	writeTestFile(t, outside, "hello\nold\n")
+	escape := strings.Join([]string{"--- " + quoted(outside), "+++ " + quoted(outside), "@@ -1,2 +1,2 @@", " hello", "-old", "+new", ""}, "\n")
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": escape})
+	if result.Status == StatusOK {
+		t.Fatal("quoted whitespace path outside the workspace must be rejected")
+	}
+	if content, _ := os.ReadFile(outside); string(content) != "hello\nold\n" {
+		t.Fatal("outside file must be untouched")
+	}
+}
+
+// A workspace path that is swapped for a symlink escaping the root — the
+// classic check-to-use attack — is refused by the os.Root engine for both
+// formats, and the outside file is never modified.
+func TestApplyPatchRefusesSymlinkEscapingWorkspace(t *testing.T) {
+	root := t.TempDir()
+	outside := filepath.Join(t.TempDir(), "secret.txt")
+	writeTestFile(t, outside, "hello\nold\n")
+	link := filepath.Join(root, "target.txt")
+	if err := os.Symlink(outside, link); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+	for name, patch := range map[string]string{
+		"unified":    strings.Join([]string{"--- a/target.txt", "+++ b/target.txt", "@@ -1,2 +1,2 @@", " hello", "-old", "+new", ""}, "\n"),
+		"structured": strings.Join([]string{"*** Begin Patch", "*** Update File: target.txt", "@@", " hello", "-old", "+new", "*** End Patch", ""}, "\n"),
+	} {
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+		if result.Status == StatusOK {
+			t.Fatalf("%s patch through an escaping symlink must be refused", name)
+		}
+		if content, _ := os.ReadFile(outside); string(content) != "hello\nold\n" {
+			t.Fatalf("%s patch must not modify the file outside the workspace", name)
+		}
+	}
+}
+
+func TestUnifiedPatchCreatesDeletesAndHonoursNoNewlineMarker(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "gone.txt"), "bye\n")
+	writeTestFile(t, filepath.Join(root, "keep.txt"), "a\nb\n")
+	patch := strings.Join([]string{
+		"diff --git a/new.txt b/new.txt",
+		"new file mode 100644",
+		"--- /dev/null",
+		"+++ b/new.txt",
+		"@@ -0,0 +1,2 @@",
+		"+first",
+		"+second",
+		"\\ No newline at end of file",
+		"diff --git a/gone.txt b/gone.txt",
+		"deleted file mode 100644",
+		"--- a/gone.txt",
+		"+++ /dev/null",
+		"@@ -1 +0,0 @@",
+		"-bye",
+		"--- a/keep.txt",
+		"+++ b/keep.txt",
+		"@@ -1,2 +1,2 @@",
+		" a",
+		"-b",
+		"+c",
+		"\\ No newline at end of file",
+		"",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("multi-file unified patch should apply, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "new.txt")); string(content) != "first\nsecond" {
+		t.Fatalf("created file = %q", string(content))
+	}
+	if _, err := os.Stat(filepath.Join(root, "gone.txt")); !os.IsNotExist(err) {
+		t.Fatal("deleted file must be gone")
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "keep.txt")); string(content) != "a\nc" {
+		t.Fatalf("updated file = %q", string(content))
+	}
+	if len(result.ChangedFiles) != 3 {
+		t.Fatalf("changed files = %v", result.ChangedFiles)
+	}
+}
+
+func TestUnifiedPatchUsesRangeHintThenFallsBackToContext(t *testing.T) {
+	root := t.TempDir()
+	// Two identical blocks: the hunk's range picks the second one, which
+	// content search alone would call ambiguous.
+	writeTestFile(t, filepath.Join(root, "dup.txt"), "x\ny\nx\ny\n")
+	patch := strings.Join([]string{"--- a/dup.txt", "+++ b/dup.txt", "@@ -3,2 +3,2 @@", " x", "-y", "+z", ""}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("range-hinted hunk should apply, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "dup.txt")); string(content) != "x\ny\nx\nz\n" {
+		t.Fatalf("content = %q", string(content))
+	}
+	// A stale range (file grew above the hunk) still applies by unique context.
+	writeTestFile(t, filepath.Join(root, "moved.txt"), "header\nhello\nold\n")
+	patch = strings.Join([]string{"--- a/moved.txt", "+++ b/moved.txt", "@@ -1,2 +1,2 @@", " hello", "-old", "+new", ""}, "\n")
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("offset hunk should apply by context, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "moved.txt")); string(content) != "header\nhello\nnew\n" {
+		t.Fatalf("content = %q", string(content))
+	}
+}
+
+func TestUnifiedPatchRenameWithHunkAndRejectsBinary(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "x.txt"), "hello\nold\n")
+	rename := strings.Join([]string{
+		"diff --git a/x.txt b/y.txt",
+		"similarity index 90%",
+		"rename from x.txt",
+		"rename to y.txt",
+		"--- a/x.txt",
+		"+++ b/y.txt",
+		"@@ -1,2 +1,2 @@",
+		" hello",
+		"-old",
+		"+new",
+		"",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": rename})
+	if result.Status != StatusOK {
+		t.Fatalf("rename with hunk should apply, got %s: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(filepath.Join(root, "x.txt")); !os.IsNotExist(err) {
+		t.Fatal("renamed source must be gone")
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "y.txt")); string(content) != "hello\nnew\n" {
+		t.Fatalf("renamed content = %q", string(content))
+	}
+	for name, patch := range map[string]string{
+		"binary": "diff --git a/x b/x\nBinary files a/x and b/x differ\n",
+		"empty":  "diff --git a/x b/x\n",
+	} {
+		if result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch}); result.Status == StatusOK {
+			t.Fatalf("%s patch must be rejected", name)
+		}
+	}
+}
+
+func TestUnifiedPatchCopyOperation(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "src.txt"), "hello\nold\n")
+	copyPatch := func(from, to string, hunk bool) string {
+		lines := []string{"diff --git a/" + from + " b/" + to, "similarity index 90%", "copy from " + from, "copy to " + to}
+		if hunk {
+			lines = append(lines, "--- a/"+from, "+++ b/"+to, "@@ -1,2 +1,2 @@", " hello", "-old", "+new")
+		}
+		return strings.Join(append(lines, ""), "\n")
+	}
+
+	// Success: source kept, destination created with the hunk applied, and the
+	// destination inherits the source's tracker state (read whole -> whole).
+	tracker := NewFileTracker()
+	registry := NewRegistry()
+	registry.Register(NewScopedReadFileTool(root, nil))
+	registry.Register(NewScopedApplyPatchTool(root, nil))
+	opts := RunOptions{PermissionGranted: true, FileTracker: tracker}
+	if r := registry.RunWithOptions(context.Background(), "read_file", map[string]any{"path": "src.txt"}, opts); r.Status != StatusOK {
+		t.Fatalf("read: %s", r.Output)
+	}
+	result := registry.RunWithOptions(context.Background(), "apply_patch", map[string]any{"patch": copyPatch("src.txt", "dst.txt", true)}, opts)
+	if result.Status != StatusOK {
+		t.Fatalf("copy with hunk should apply, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "src.txt")); string(content) != "hello\nold\n" {
+		t.Fatalf("copy must keep the source, got %q", string(content))
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "dst.txt")); string(content) != "hello\nnew\n" {
+		t.Fatalf("copy destination = %q", string(content))
+	}
+	if len(result.ChangedFiles) != 2 || result.ChangedFiles[0] != "src.txt" || result.ChangedFiles[1] != "dst.txt" {
+		t.Fatalf("changed files = %v", result.ChangedFiles)
+	}
+	destination, err := filepath.EvalSymlinks(filepath.Join(root, "dst.txt"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !tracker.SeenWhole(destination) {
+		t.Fatal("destination copied from a fully read source must be tracked as seen whole")
+	}
+	unreadSource, _ := filepath.EvalSymlinks(filepath.Join(root, "src.txt"))
+	if _, tracked := tracker.Version(unreadSource); !tracked {
+		t.Fatal("source must stay tracked after a copy")
+	}
+
+	// Failure: destination already exists — nothing is written.
+	writeTestFile(t, filepath.Join(root, "taken.txt"), "keep\n")
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": copyPatch("src.txt", "taken.txt", false)})
+	if result.Status == StatusOK || !strings.Contains(result.Output, "already exists") {
+		t.Fatalf("copy onto an existing destination must be refused, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "taken.txt")); string(content) != "keep\n" {
+		t.Fatal("existing destination must be untouched")
+	}
+
+	// Failure: copy onto itself.
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": copyPatch("src.txt", "src.txt", false)})
+	if result.Status == StatusOK || !strings.Contains(result.Output, "onto itself") {
+		t.Fatalf("copy onto itself must be refused, got %s: %s", result.Status, result.Output)
+	}
+}
+
+func TestUnifiedPatchRejectsOverDeclaredHunkCounts(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "a.txt"), "hello\nold\n")
+	writeTestFile(t, filepath.Join(root, "b.txt"), "x\n")
+	// "-1,3 +1,3" over-declares a two-line hunk, so without a guard the next
+	// file's ---/+++ headers would be swallowed as "-"/"+" content.
+	patch := strings.Join([]string{
+		"--- a/a.txt", "+++ b/a.txt", "@@ -1,3 +1,3 @@", " hello", "-old", "+new",
+		"--- a/b.txt", "+++ b/b.txt", "@@ -1 +1 @@", "-x", "+y", "",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status == StatusOK || !strings.Contains(result.Output, "declared line counts") {
+		t.Fatalf("over-declared hunk must be reported as malformed, got %s: %s", result.Status, result.Output)
+	}
+	for name, want := range map[string]string{"a.txt": "hello\nold\n", "b.txt": "x\n"} {
+		if content, _ := os.ReadFile(filepath.Join(root, name)); string(content) != want {
+			t.Fatalf("%s must be untouched after a malformed patch, got %q", name, string(content))
+		}
+	}
+	// Truncated at end of input is reported the same way.
+	truncated := "--- a/a.txt\n+++ b/a.txt\n@@ -1,2 +1,2 @@\n hello\n-old\n"
+	if result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": truncated}); result.Status == StatusOK || !strings.Contains(result.Output, "declared line counts") {
+		t.Fatalf("truncated hunk must be reported as malformed, got %s: %s", result.Status, result.Output)
+	}
+}
+
+func TestUnifiedPatchNoNewlineMarkerOnlyAfterFinalHunk(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "two.txt"), "a\nb\nc\nd\ne\nf\n")
+	// A marker after the first of two hunks describes nothing real and must
+	// not leak into the second hunk's result: the patch is rejected untouched.
+	early := strings.Join([]string{
+		"--- a/two.txt", "+++ b/two.txt",
+		"@@ -1,2 +1,2 @@", " a", "-b", "+B", "\\ No newline at end of file",
+		"@@ -5,2 +5,2 @@", " e", "-f", "+F", "",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": early})
+	if result.Status == StatusOK || !strings.Contains(result.Output, "must follow the last hunk") {
+		t.Fatalf("marker before a later hunk must be rejected, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "two.txt")); string(content) != "a\nb\nc\nd\ne\nf\n" {
+		t.Fatal("rejected patch must not modify the file")
+	}
+	// The same marker after the final hunk applies and strips the newline.
+	final := strings.Join([]string{
+		"--- a/two.txt", "+++ b/two.txt",
+		"@@ -1,2 +1,2 @@", " a", "-b", "+B",
+		"@@ -5,2 +5,2 @@", " e", "-f", "+F", "\\ No newline at end of file", "",
+	}, "\n")
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": final})
+	if result.Status != StatusOK {
+		t.Fatalf("two-hunk patch with a final marker should apply, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "two.txt")); string(content) != "a\nB\nc\nd\ne\nF" {
+		t.Fatalf("content = %q", string(content))
+	}
+}
+
+// A removed "-- …" line directly followed by an added "++ …" line looks like a
+// ---/+++ header pair; the boundary check must not mistake it for one.
+func TestUnifiedPatchKeepsAdjacentDashPlusContentLines(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "notes.sql"), "select 1;\n-- old comment\nselect 2;\n")
+	patch := strings.Join([]string{
+		"--- a/notes.sql", "+++ b/notes.sql",
+		"@@ -1,3 +1,3 @@", " select 1;", "--- old comment", "+++ new comment", " select 2;", "",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("adjacent -- / ++ content lines must apply, got %s: %s", result.Status, result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "notes.sql")); string(content) != "select 1;\n++ new comment\nselect 2;\n" {
+		t.Fatalf("content = %q", string(content))
+	}
+}
+
+// A unified deletion states the content it removes; it must be verified
+// against the current file before anything is deleted.
+func TestUnifiedPatchDeletionVerifiesExpectedContent(t *testing.T) {
+	deletion := func(path string, hunks ...string) string {
+		lines := []string{"diff --git a/" + path + " b/" + path, "deleted file mode 100644", "--- a/" + path, "+++ /dev/null"}
+		return strings.Join(append(append(lines, hunks...), ""), "\n")
+	}
+	t.Run("matching deletion succeeds", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "current\n")
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deletion("gone.txt", "@@ -1 +0,0 @@", "-current")})
+		if result.Status != StatusOK {
+			t.Fatalf("matching deletion should apply, got %s: %s", result.Status, result.Output)
+		}
+		if _, err := os.Stat(filepath.Join(root, "gone.txt")); !os.IsNotExist(err) {
+			t.Fatal("file must be deleted")
+		}
+	})
+	t.Run("stale deletion is refused untouched", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "current\n")
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deletion("gone.txt", "@@ -1 +0,0 @@", "-stale")})
+		if result.Status == StatusOK || !strings.Contains(result.Output, "does not match its current content") {
+			t.Fatalf("stale deletion must be refused, got %s: %s", result.Status, result.Output)
+		}
+		if content, _ := os.ReadFile(filepath.Join(root, "gone.txt")); string(content) != "current\n" {
+			t.Fatalf("file must be byte-for-byte unchanged, got %q", string(content))
+		}
+	})
+	t.Run("partial deletion is refused", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "one\ntwo\n")
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deletion("gone.txt", "@@ -1 +0,0 @@", "-one")})
+		if result.Status == StatusOK || !strings.Contains(result.Output, "does not match its current content") {
+			t.Fatalf("deletion that does not cover the file must be refused, got %s: %s", result.Status, result.Output)
+		}
+		if content, _ := os.ReadFile(filepath.Join(root, "gone.txt")); string(content) != "one\ntwo\n" {
+			t.Fatal("file must be unchanged")
+		}
+	})
+	t.Run("header-only deletion is refused", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "current\n")
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deletion("gone.txt")})
+		if result.Status == StatusOK || !strings.Contains(result.Output, "must include the hunk") {
+			t.Fatalf("header-only deletion must be refused, got %s: %s", result.Status, result.Output)
+		}
+		if _, err := os.Stat(filepath.Join(root, "gone.txt")); err != nil {
+			t.Fatal("file must still exist")
+		}
+	})
+	t.Run("multi-hunk deletion verifies everything before removing", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "a\nb\nc\nd\n")
+		good := deletion("gone.txt", "@@ -1,2 +0,0 @@", "-a", "-b", "@@ -3,2 +0,0 @@", "-c", "-d")
+		bad := deletion("gone.txt", "@@ -1,2 +0,0 @@", "-a", "-b", "@@ -3,2 +0,0 @@", "-c", "-stale")
+		result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": bad})
+		if result.Status == StatusOK {
+			t.Fatalf("deletion with one stale hunk must be refused: %s", result.Output)
+		}
+		if content, _ := os.ReadFile(filepath.Join(root, "gone.txt")); string(content) != "a\nb\nc\nd\n" {
+			t.Fatal("no hunk may be applied when any hunk is stale")
+		}
+		result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": good})
+		if result.Status != StatusOK {
+			t.Fatalf("multi-hunk deletion should apply, got %s: %s", result.Status, result.Output)
+		}
+		if _, err := os.Stat(filepath.Join(root, "gone.txt")); !os.IsNotExist(err) {
+			t.Fatal("file must be deleted")
+		}
+	})
+	t.Run("structured delete stays unconditional", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "anything\n")
+		patch := "*** Begin Patch\n*** Delete File: gone.txt\n*** End Patch\n"
+		if result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch}); result.Status != StatusOK {
+			t.Fatalf("structured delete should apply, got %s: %s", result.Status, result.Output)
+		}
+	})
+}
+
+// The source is re-read through the root immediately before commit, so a
+// change made after planning is refused instead of overwritten or removed.
+func TestApplyStructuredPatchChangeRefusesSourceChangedAfterPlanning(t *testing.T) {
+	root := t.TempDir()
+	path := filepath.Join(root, "file.txt")
+	writeTestFile(t, path, "planned\n")
+	workspace, err := os.OpenRoot(root)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer workspace.Close()
+	target := structuredPatchTarget{requested: "file.txt", absolute: path, relative: "file.txt"}
+	destination := structuredPatchTarget{requested: "copy.txt", absolute: filepath.Join(root, "copy.txt"), relative: "copy.txt"}
+	for name, change := range map[string]structuredPatchChange{
+		"update": {kind: structuredPatchUpdate, from: target, to: target, before: "planned\n", after: "new\n", mode: 0o644},
+		"delete": {kind: structuredPatchDelete, from: target, to: target, before: "planned\n", mode: 0o644},
+		"copy":   {kind: structuredPatchCopy, from: target, to: destination, before: "planned\n", after: "planned\n", mode: 0o644},
+	} {
+		writeTestFile(t, path, "changed after planning\n")
+		committed, err := applyStructuredPatchChange(workspace, change)
+		if err == nil || committed || !strings.Contains(err.Error(), "changed on disk between planning and commit") {
+			t.Fatalf("%s: expected a refusal, got committed=%v err=%v", name, committed, err)
+		}
+		if content, _ := os.ReadFile(path); string(content) != "changed after planning\n" {
+			t.Fatalf("%s: file must be untouched, got %q", name, string(content))
+		}
+		if _, statErr := os.Stat(filepath.Join(root, "copy.txt")); !os.IsNotExist(statErr) {
+			t.Fatalf("%s: refused change must not create copy.txt", name)
+		}
+	}
+}
+
+func TestUnifiedPatchDeletionIsByteExact(t *testing.T) {
+	deletion := func(hunks ...string) string {
+		return strings.Join(append(append([]string{"--- a/gone.txt", "+++ /dev/null"}, hunks...), ""), "\n")
+	}
+	for name, tc := range map[string]struct{ file, patch string }{
+		"trailing whitespace differs": {"current  \n", deletion("@@ -1 +0,0 @@", "-current")},
+		"indentation differs":         {"  current\n", deletion("@@ -1 +0,0 @@", "-current")},
+		"blank line not covered":      {"current\n\n", deletion("@@ -1 +0,0 @@", "-current")},
+		"space-only line not covered": {"current\n   \n", deletion("@@ -1 +0,0 @@", "-current")},
+		"missing final newline":       {"current", deletion("@@ -1 +0,0 @@", "-current")},
+	} {
+		t.Run(name, func(t *testing.T) {
+			root := t.TempDir()
+			writeTestFile(t, filepath.Join(root, "gone.txt"), tc.file)
+			result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": tc.patch})
+			if result.Status == StatusOK {
+				t.Fatalf("deletion must be byte-exact; %q was deleted by a non-matching hunk", tc.file)
+			}
+			if content, _ := os.ReadFile(filepath.Join(root, "gone.txt")); string(content) != tc.file {
+				t.Fatalf("file must be unchanged, got %q", string(content))
+			}
+		})
+	}
+	t.Run("no-newline marker on the removed side matches exactly", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "current")
+		patch := deletion("@@ -1 +0,0 @@", "-current", "\\ No newline at end of file")
+		if result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch}); result.Status != StatusOK {
+			t.Fatalf("deletion with a matching no-newline marker should apply, got %s: %s", result.Status, result.Output)
+		}
+		if _, err := os.Stat(filepath.Join(root, "gone.txt")); !os.IsNotExist(err) {
+			t.Fatal("file must be deleted")
+		}
+	})
+	t.Run("CRLF file matches an LF patch of the same content", func(t *testing.T) {
+		root := t.TempDir()
+		writeTestFile(t, filepath.Join(root, "gone.txt"), "a\r\nb\r\n")
+		if result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deletion("@@ -1,2 +0,0 @@", "-a", "-b")}); result.Status != StatusOK {
+			t.Fatalf("CRLF deletion should apply, got %s: %s", result.Status, result.Output)
+		}
+	})
+}
+
+// git describes an empty file's creation or deletion with header lines only.
+func TestUnifiedPatchHeaderOnlyEmptyFileForms(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "empty.txt"), "")
+	writeTestFile(t, filepath.Join(root, "full.txt"), "content\n")
+	writeTestFile(t, filepath.Join(root, "keep.txt"), "a\n")
+	deleteHeader := func(path string) string {
+		return "diff --git a/" + path + " b/" + path + "\ndeleted file mode 100644\nindex e69de29..0000000\n"
+	}
+	// A non-empty file is not deleted by the header-only form.
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": deleteHeader("full.txt")})
+	if result.Status == StatusOK || !strings.Contains(result.Output, "expects an empty file") {
+		t.Fatalf("header-only deletion of a non-empty file must be refused, got %s: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(filepath.Join(root, "full.txt")); err != nil {
+		t.Fatal("non-empty file must still exist")
+	}
+	// In a multi-file patch the empty-file deletion and creation are applied
+	// alongside a normal hunk.
+	patch := deleteHeader("empty.txt") +
+		"diff --git a/blank.txt b/blank.txt\nnew file mode 100644\nindex 0000000..e69de29\n" +
+		"diff --git a/keep.txt b/keep.txt\n--- a/keep.txt\n+++ b/keep.txt\n@@ -1 +1 @@\n-a\n+b\n"
+	result = NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status != StatusOK {
+		t.Fatalf("multi-file patch with header-only forms should apply, got %s: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(filepath.Join(root, "empty.txt")); !os.IsNotExist(err) {
+		t.Fatal("empty file must be deleted")
+	}
+	if content, err := os.ReadFile(filepath.Join(root, "blank.txt")); err != nil || len(content) != 0 {
+		t.Fatalf("empty file must be created, got err=%v content=%q", err, string(content))
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "keep.txt")); string(content) != "b\n" {
+		t.Fatalf("hunk file = %q", string(content))
+	}
+	if len(result.ChangedFiles) != 3 {
+		t.Fatalf("changed files = %v", result.ChangedFiles)
+	}
+}
+
+// When a later change fails, the error names exactly which files were already
+// committed so the caller knows what changed and what did not.
+func TestApplyPatchOperationsReportsCommittedPrefixOnFailure(t *testing.T) {
+	root := t.TempDir()
+	writeTestFile(t, filepath.Join(root, "first.txt"), "one\n")
+	writeTestFile(t, filepath.Join(root, "second.txt"), "two\n")
+	writeTestFile(t, filepath.Join(root, "third.txt"), "three\n")
+	// Make second.txt change after planning so its commit is refused.
+	structuredPatchBeforeCommit = func(change structuredPatchChange) {
+		if change.to.relative == "second.txt" {
+			writeTestFile(t, filepath.Join(root, "second.txt"), "changed\n")
+		}
+	}
+	defer func() { structuredPatchBeforeCommit = nil }()
+	patch := strings.Join([]string{
+		"*** Begin Patch",
+		"*** Update File: first.txt", "@@", "-one", "+ONE",
+		"*** Update File: second.txt", "@@", "-two", "+TWO",
+		"*** Update File: third.txt", "@@", "-three", "+THREE",
+		"*** End Patch", "",
+	}, "\n")
+	result := NewScopedApplyPatchTool(root, nil).Run(context.Background(), map[string]any{"patch": patch})
+	if result.Status == StatusOK {
+		t.Fatal("patch must fail when a later change is refused")
+	}
+	for _, want := range []string{"already committed: first.txt", "remaining files are unchanged"} {
+		if !strings.Contains(result.Output, want) {
+			t.Fatalf("error must report the committed prefix, got: %s", result.Output)
+		}
+	}
+	if strings.Contains(result.Output, "committed: first.txt, second.txt") || strings.Contains(result.Output, "third.txt") {
+		t.Fatalf("error must not list uncommitted files as committed: %s", result.Output)
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "first.txt")); string(content) != "ONE\n" {
+		t.Fatalf("first.txt must hold the committed change, got %q", string(content))
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "second.txt")); string(content) != "changed\n" {
+		t.Fatalf("second.txt must be untouched by the patch, got %q", string(content))
+	}
+	if content, _ := os.ReadFile(filepath.Join(root, "third.txt")); string(content) != "three\n" {
+		t.Fatalf("third.txt must be untouched, got %q", string(content))
+	}
+}

internal/tools/edit_file.go

@@ -168,7 +168,7 @@ func (tool editFileTool) RunWithOptions(ctx context.Context, args map[string]any
 	options.FileTracker.Record(absolutePath, []byte(updated), newInfo)
 	if updated == modelKnownContent {
 		if previouslySeenWhole {
-			options.FileTracker.RecordSeenRange(absolutePath, 1, lineCount(updated), lineCount(updated))
+			options.FileTracker.RecordSeenRange(absolutePath, 1, trackedLineTotal(updated), trackedLineTotal(updated))
 		} else {
 			for _, span := range editedSpans {
 				options.FileTracker.RecordSeenBytes(absolutePath, span.start, span.end, len(updated))

internal/tools/file_tools_test.go

@@ -26,8 +26,8 @@ func TestReadFileToolReadsLineRanges(t *testing.T) {
 	}
 	for _, want := range []string{
 		"File: notes.txt (lines 2-3 of 4)",
-		"2 | beta",
-		"3 | gamma",
+		"2→beta",
+		"3→gamma",
 	} {
 		if !strings.Contains(result.Output, want) {
 			t.Fatalf("expected output to contain %q, got %q", want, result.Output)
@@ -54,7 +54,7 @@ func TestReadFileToolReadsCanonicalLineRange(t *testing.T) {
 	if result.Truncated {
 		t.Fatalf("canonical limit defines the requested range and must not mark it truncated: %#v", result.Meta)
 	}
-	if !strings.Contains(result.Output, "2 | beta") || !strings.Contains(result.Output, "3 | gamma") {
+	if !strings.Contains(result.Output, "2→beta") || !strings.Contains(result.Output, "3→gamma") {
 		t.Fatalf("canonical range returned the wrong lines: %q", result.Output)
 	}
 	if strings.Contains(result.Output, "alpha") || strings.Contains(result.Output, "delta") {
@@ -94,7 +94,7 @@ func TestReadFileToolMixedLegacyRangesPreferLines(t *testing.T) {
 	if result.Status != StatusOK {
 		t.Fatalf("mixed legacy range should recover, got %s: %s", result.Status, result.Output)
 	}
-	if !strings.Contains(result.Output, "2 | beta") || !strings.Contains(result.Output, "3 | gamma") {
+	if !strings.Contains(result.Output, "2→beta") || !strings.Contains(result.Output, "3→gamma") {
 		t.Fatalf("mixed legacy range did not prefer lines: %q", result.Output)
 	}
 }
@@ -110,7 +110,7 @@ func TestReadFileToolCombinesLegacyStartWithCanonicalLimit(t *testing.T) {
 	if result.Status != StatusOK || result.Truncated {
 		t.Fatalf("mixed compatible range should be exact, got status=%s truncated=%v: %s", result.Status, result.Truncated, result.Output)
 	}
-	if !strings.Contains(result.Output, "2 | beta") || !strings.Contains(result.Output, "3 | gamma") {
+	if !strings.Contains(result.Output, "2→beta") || !strings.Contains(result.Output, "3→gamma") {
 		t.Fatalf("mixed compatible range returned the wrong lines: %q", result.Output)
 	}
 	if strings.Contains(result.Output, "alpha") || strings.Contains(result.Output, "delta") {
@@ -955,7 +955,7 @@ func TestReadFileToolRecoversBackwardsRange(t *testing.T) {
 	for _, want := range []string{
 		"end_line 2 was before start_line 3", // the note
 		"only line 3 was read",
-		"3 | gamma",
+		"3→gamma",
 	} {
 		if !strings.Contains(result.Output, want) {
 			t.Fatalf("expected %q in output, got %q", want, result.Output)
@@ -984,7 +984,7 @@ func TestReadFileToolValidRangeUnchangedByRecovery(t *testing.T) {
 	// Byte-for-byte: header, blank separator, the exact selected lines, and no
 	// recovery note or trailing content. An altered header/separator or a stray
 	// out-of-range line would fail this where a substring check would not.
-	const want = "File: notes.txt (lines 2-4 of 5)\n\n2 | beta\n3 | gamma\n4 | delta"
+	const want = "File: notes.txt (lines 2-4 of 5)\n\n2→beta\n3→gamma\n4→delta"
 	if result.Output != want {
 		t.Fatalf("valid-range output changed by the recovery path:\n got: %q\nwant: %q", result.Output, want)
 	}

internal/tools/model_visibility.go

@@ -1,9 +1,11 @@
 package tools
 
 // ModelVisible reports whether a registered tool is advertised to an agent.
-// edit_file remains implemented for compatibility with existing integrations,
-// but its text-only replacement contract cannot safely disambiguate repeated
-// source fragments. Agents use apply_patch for contextual existing-file edits.
+// Every registered tool is visible. edit_file (exact search/replace with a
+// uniqueness guard, CRLF and fuzzy fallbacks) was hidden for a while in favour
+// of apply_patch; head-to-head benchmarking showed that when a model misses the
+// patch grammar it degrades to whole-file rewrites, so both edit tools are
+// advertised and the prompt steers between them.
 func ModelVisible(tool Tool) bool {
-	return tool != nil && tool.Name() != "edit_file"
+	return tool != nil
 }

internal/tools/model_visibility_test.go

@@ -2,17 +2,28 @@ package tools
 
 import "testing"
 
-func TestToolSearchDoesNotSurfaceLegacyStringReplacement(t *testing.T) {
+func TestToolSearchSurfacesBothEditTools(t *testing.T) {
 	registry := NewRegistry()
 	registry.Register(NewScopedEditFileTool(t.TempDir(), nil))
 	registry.Register(NewScopedApplyPatchTool(t.TempDir(), nil))
 	search := NewToolSearchTool(registry).(toolSearchTool)
 
 	eager := search.visibleEagerToolNames(nil, nil, "ask")
-	if eager["edit_file"] {
-		t.Fatal("tool_search must not describe edit_file as available to the model")
+	if !eager["edit_file"] {
+		t.Fatal("tool_search must describe edit_file as available to the model")
 	}
 	if !eager["apply_patch"] {
-		t.Fatal("tool_search must retain apply_patch as the available edit tool")
+		t.Fatal("tool_search must retain apply_patch as an available edit tool")
+	}
+}
+
+func TestModelVisibleAdvertisesEveryRegisteredTool(t *testing.T) {
+	if ModelVisible(nil) {
+		t.Fatal("a nil tool must not be visible")
+	}
+	for _, tool := range []Tool{NewScopedEditFileTool(t.TempDir(), nil), NewScopedApplyPatchTool(t.TempDir(), nil), NewScopedWriteFileTool(t.TempDir(), nil)} {
+		if !ModelVisible(tool) {
+			t.Fatalf("%s must be visible to the model", tool.Name())
+		}
 	}
 }

internal/tools/plan_tool_test.go

@@ -9,6 +9,9 @@ import (
 
 func TestUpdatePlanToolStoresAndFormatsPlan(t *testing.T) {
 	tool := NewUpdatePlanTool()
+	if description := strings.ToLower(tool.Description()); !strings.Contains(description, "skip it for bounded changes in one component") {
+		t.Fatalf("update_plan description must preserve the bounded-task fast path: %q", description)
+	}
 
 	result := tool.Run(context.Background(), map[string]any{
 		"plan": []any{

internal/tools/read_file.go

@@ -9,12 +9,14 @@ import (
 	"io"
 	"os"
 	"strconv"
-	"strings"
 	"unicode/utf8"
 )
 
 const readFileByteChunkMax = 64 * 1024
 
+// readFileLinePrefixSeparator follows the line number on every read_file line.
+const readFileLinePrefixSeparator = "→"
+
 type readFileTool struct {
 	baseTool
 	workspaceRoot string
@@ -31,7 +33,7 @@ func NewScopedReadFileTool(workspaceRoot string, scope PathScope) Tool {
 	return readFileTool{
 		baseTool: baseTool{
 			name:        "read_file",
-			description: "Read exact file text with line numbers. Use for comments, formatting, or edits; prefer read_minified_file for initial code understanding. Use offset and limit for a line range.",
+			description: "Read exact file text, each line prefixed with its line number and →. Use directly for small files likely to be edited, or when comments, formatting, or line numbers matter. Prefer read_minified_file only for exploratory understanding of large or unfamiliar code. Use offset and limit for a line range.",
 			parameters: Schema{
 				Type: "object",
 				Properties: map[string]PropertySchema{
@@ -221,7 +223,6 @@ func renderReadFileRange(absolutePath string, relativePath string, total int, st
 	}
 
 	lastLine := startLine + selectedLines - 1
-	width := len(strconv.Itoa(lastLine))
 	header := fmt.Sprintf("File: %s (%d lines)", relativePath, total)
 	if startLine != 1 || endLine != total || maxLines > 0 {
 		header = fmt.Sprintf("File: %s (lines %d-%d of %d)", relativePath, startLine, lastLine, total)
@@ -244,7 +245,7 @@ func renderReadFileRange(absolutePath string, relativePath string, total int, st
 		budgetedOutput.WriteString("\n")
 	}
 	budgetedOutput.WriteString("\n")
-	if err := appendReadFileRange(budgetedOutput, absolutePath, startLine, selectedLines, width); err != nil {
+	if err := appendReadFileRange(budgetedOutput, absolutePath, startLine, selectedLines); err != nil {
 		return errorResult("Error reading file " + relativePath + ": " + err.Error())
 	}
 	if truncated {
@@ -357,7 +358,7 @@ func renderReadFileBytes(path, relativePath string, total, requestedStart, limit
 	return Result{Status: StatusOK, Output: output, Meta: map[string]string{"next_byte_offset": strconv.Itoa(end)}}, start, end
 }
 
-func appendReadFileRange(output *outputBudgetBuilder, path string, startLine int, selectedLines int, width int) error {
+func appendReadFileRange(output *outputBudgetBuilder, path string, startLine int, selectedLines int) error {
 	file, err := os.Open(path)
 	if err != nil {
 		return err
@@ -380,10 +381,11 @@ func appendReadFileRange(output *outputBudgetBuilder, path string, startLine int
 			if emitted > 0 {
 				output.WriteString("\n")
 			}
-			number := strconv.Itoa(lineNumber)
-			output.WriteString(strings.Repeat(" ", width-len(number)))
-			output.WriteString(number)
-			output.WriteString(" | ")
+			// Compact "N→" prefix: the padded "  N | " form cost ~19% of every
+			// read (about 1.5k tokens per 800-line file) and was re-sent on each
+			// later call; no consumer parses it, and models read "N→" natively.
+			output.WriteString(strconv.Itoa(lineNumber))
+			output.WriteString(readFileLinePrefixSeparator)
 			output.WriteString(string(trimLineBreak(raw, ended)))
 			emitted++
 		}

internal/tools/read_minified_file.go

@@ -28,7 +28,7 @@ func NewScopedReadMinifiedFileTool(workspaceRoot string, scope PathScope) Tool {
 	return readMinifiedFileTool{
 		baseTool: baseTool{
 			name:        "read_minified_file",
-			description: "Read source code in a token-efficient, language-aware form. Prefer this for initial understanding; use read_file for exact text or line numbers.",
+			description: "Read source code in a token-efficient, language-aware form. Use for exploratory understanding of large or unfamiliar source when no exact edit is planned; use read_file directly for likely edit targets. Do not immediately reread the same file exactly unless a new need for exact text or line numbers appears.",
 			parameters: Schema{
 				Type: "object",
 				Properties: map[string]PropertySchema{
@@ -171,14 +171,9 @@ func selectSourceLines(content []byte, offset, limit int) sourceSelection {
 }
 
 func sourceLineCount(content []byte) int {
-	if len(content) == 0 {
-		return 1
-	}
-	lines := strings.Count(string(content), "\n")
-	if content[len(content)-1] != '\n' {
-		lines++
-	}
-	return lines
+	// One implementation for reader and writer: the tracker contract lives on
+	// trackedLineTotal, and this must never drift from it.
+	return trackedLineTotal(string(content))
 }
 
 // lineCount reports the number of newline-separated lines in s (an empty string
@@ -189,3 +184,20 @@ func lineCount(s string) int {
 	}
 	return strings.Count(s, "\n") + 1
 }
+
+// trackedLineTotal reports a file's line count the way read_file reports it
+// to the FileTracker (newline-terminated lines, plus an unterminated last
+// line; an empty file is one line). Writers must record this same number:
+// RecordSeenRange resets every observation when the total changes, so a
+// writer that recorded lineCount (one higher for a trailing newline) made the
+// next partial read_file wipe whole-file knowledge and refuse the next edit.
+func trackedLineTotal(s string) int {
+	if s == "" {
+		return 1
+	}
+	total := strings.Count(s, "\n")
+	if !strings.HasSuffix(s, "\n") {
+		total++
+	}
+	return total
+}

internal/tools/read_minified_file_test.go

@@ -8,6 +8,29 @@ import (
 	"testing"
 )
 
+func TestReadToolsDescribeExploratoryAndExactUseWithoutRedundantRereads(t *testing.T) {
+	root := t.TempDir()
+	minified := strings.ToLower(NewScopedReadMinifiedFileTool(root, nil).Description())
+	for _, want := range []string{
+		"exploratory understanding",
+		"use read_file directly for likely edit targets",
+		"unless a new need for exact text or line numbers appears",
+	} {
+		if !strings.Contains(minified, want) {
+			t.Fatalf("read_minified_file description missing %q: %s", want, minified)
+		}
+	}
+	exact := strings.ToLower(NewScopedReadFileTool(root, nil).Description())
+	for _, want := range []string{
+		"use directly for small files likely to be edited",
+		"prefer read_minified_file only for exploratory understanding",
+	} {
+		if !strings.Contains(exact, want) {
+			t.Fatalf("read_file description missing %q: %s", want, exact)
+		}
+	}
+}
+
 func TestReadMinifiedFileStripsCommentsAndLineNumbers(t *testing.T) {
 	dir := t.TempDir()
 	src := "package demo\n\nimport \"fmt\"\n\n// secret doc comment\nfunc F() { fmt.Println(\"x\") }\n"

internal/tools/registry.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"sort"
 	"strings"
+	"sync"
 
 	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
@@ -12,7 +13,17 @@ import (
 )
 
 type Registry struct {
-	tools map[string]Tool
+	mu         sync.RWMutex
+	tools      map[string]Tool
+	generation uint64
+}
+
+// RegistrySnapshot is one immutable registry generation. Tools is a detached,
+// name-sorted slice: later registrations cannot change the snapshot observed by
+// an in-flight agent turn.
+type RegistrySnapshot struct {
+	Tools      []Tool
+	Generation uint64
 }
 
 type RunOptions struct {
@@ -107,23 +118,69 @@ func IsDeferralEligible(t Tool) bool {
 }
 
 func (registry *Registry) Register(tool Tool) {
+	registry.mu.Lock()
+	defer registry.mu.Unlock()
 	registry.tools[tool.Name()] = tool
+	registry.generation++
+}
+
+// RegisterBatch publishes a group of tools as one registry generation. Readers
+// observe either the complete prior generation or the complete new one.
+func (registry *Registry) RegisterBatch(batch []Tool) {
+	if len(batch) == 0 {
+		return
+	}
+	registry.mu.Lock()
+	defer registry.mu.Unlock()
+	for _, tool := range batch {
+		registry.tools[tool.Name()] = tool
+	}
+	registry.generation++
 }
 
 func (registry *Registry) Get(name string) (Tool, bool) {
+	registry.mu.RLock()
+	defer registry.mu.RUnlock()
 	tool, ok := registry.tools[name]
 	return tool, ok
 }
 
 func (registry *Registry) All() []Tool {
+	return registry.Snapshot().Tools
+}
+
+// Snapshot returns a stable, sorted view of the registry at one generation.
+func (registry *Registry) Snapshot() RegistrySnapshot {
+	registry.mu.RLock()
 	tools := make([]Tool, 0, len(registry.tools))
 	for _, tool := range registry.tools {
 		tools = append(tools, tool)
 	}
+	generation := registry.generation
+	registry.mu.RUnlock()
 	sort.Slice(tools, func(left, right int) bool {
 		return tools[left].Name() < tools[right].Name()
 	})
-	return tools
+	return RegistrySnapshot{Tools: tools, Generation: generation}
+}
+
+// Clone returns an independent registry containing exactly one source snapshot.
+func (registry *Registry) Clone() *Registry {
+	clone := NewRegistry()
+	if registry == nil {
+		return clone
+	}
+	snapshot := registry.Snapshot()
+	clonedTools := make([]Tool, 0, len(snapshot.Tools))
+	for _, tool := range snapshot.Tools {
+		if cloner, ok := tool.(interface{ cloneForRegistry(*Registry) Tool }); ok {
+			tool = cloner.cloneForRegistry(clone)
+		}
+		clonedTools = append(clonedTools, tool)
+	}
+	clone.RegisterBatch(clonedTools)
+	clone.generation = snapshot.Generation
+	return clone
 }
 
 func (registry *Registry) Run(ctx context.Context, name string, args map[string]any) Result {

internal/tools/registry_test.go

@@ -2,10 +2,12 @@ package tools
 
 import (
 	"context"
+	"fmt"
 	"os"
 	"path/filepath"
 	"slices"
 	"strings"
+	"sync"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
@@ -95,6 +97,81 @@ func TestRegistryAllReturnsToolsSortedByName(t *testing.T) {
 	}
 }
 
+func TestRegistrySnapshotAndCloneStayOnOneGeneration(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(fakePlainTool{baseTool: baseTool{name: "alpha", description: "alpha"}})
+	first := registry.Snapshot()
+
+	registry.RegisterBatch([]Tool{
+		fakePlainTool{baseTool: baseTool{name: "charlie", description: "charlie"}},
+		fakePlainTool{baseTool: baseTool{name: "bravo", description: "bravo"}},
+	})
+	second := registry.Snapshot()
+	clone := registry.Clone()
+	registry.Register(fakePlainTool{baseTool: baseTool{name: "delta", description: "delta"}})
+
+	if got := toolNames(first.Tools); !slices.Equal(got, []string{"alpha"}) {
+		t.Fatalf("first snapshot changed after registration: %v", got)
+	}
+	if second.Generation != first.Generation+1 {
+		t.Fatalf("batch generation = %d, want %d", second.Generation, first.Generation+1)
+	}
+	if got := toolNames(second.Tools); !slices.Equal(got, []string{"alpha", "bravo", "charlie"}) {
+		t.Fatalf("second snapshot tools = %v", got)
+	}
+	if got := toolNames(clone.All()); !slices.Equal(got, toolNames(second.Tools)) {
+		t.Fatalf("clone changed with source registry: %v", got)
+	}
+	if clone.Snapshot().Generation != second.Generation {
+		t.Fatalf("clone generation = %d, want %d", clone.Snapshot().Generation, second.Generation)
+	}
+}
+
+func TestRegistryBatchIsAtomicForConcurrentReaders(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(fakePlainTool{baseTool: baseTool{name: "base", description: "base"}})
+	batch := make([]Tool, 128)
+	for index := range batch {
+		name := fmt.Sprintf("tool_%03d", index)
+		batch[index] = fakePlainTool{baseTool: baseTool{name: name, description: name}}
+	}
+
+	start := make(chan struct{})
+	done := make(chan struct{})
+	var readers sync.WaitGroup
+	for range 4 {
+		readers.Add(1)
+		go func() {
+			defer readers.Done()
+			<-start
+			for {
+				count := len(registry.Snapshot().Tools)
+				if count != 1 && count != len(batch)+1 {
+					t.Errorf("reader observed partial batch with %d tools", count)
+					return
+				}
+				select {
+				case <-done:
+					return
+				default:
+				}
+			}
+		}()
+	}
+	close(start)
+	registry.RegisterBatch(batch)
+	close(done)
+	readers.Wait()
+}
+
+func toolNames(toolset []Tool) []string {
+	names := make([]string, 0, len(toolset))
+	for _, tool := range toolset {
+		names = append(names, tool.Name())
+	}
+	return names
+}
+
 func TestCoreNetworkToolsExposeSafetyMetadata(t *testing.T) {
 	// web_search is only registered when a backend is configured.
 	t.Setenv("ZERO_WEBSEARCH_BASE_URL", "https://search.example/api")

internal/tools/structured_patch.go

@@ -6,9 +6,11 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"regexp"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/pathjail"
+	"github.com/Gitlawb/zero/internal/sandbox"
 )
 
 const (
@@ -28,6 +30,9 @@ const (
 	structuredPatchAdd structuredPatchKind = iota
 	structuredPatchDelete
 	structuredPatchUpdate
+	// structuredPatchCopy creates movePath from path's content (plus any
+	// hunks) and keeps path; produced by a unified diff's "copy from/to".
+	structuredPatchCopy
 )
 
 type structuredPatchOperation struct {
@@ -37,8 +42,29 @@ type structuredPatchOperation struct {
 	contents string
 	chunks   []structuredPatchChunk
 	line     int
+	// eofNewline lets a unified diff's "\ No newline at end of file" marker
+	// force the trailing-newline state of the result; structured patches keep
+	// the file's existing state.
+	eofNewline eofNewlineMode
+	// verifyDelete marks a deletion that carries the expected old content
+	// (a unified diff's "+++ /dev/null" hunks, or git's header-only form for
+	// an empty file): the removed lines must equal the current file byte for
+	// byte, otherwise the deletion is stale and refused. A structured
+	// "*** Delete File" means "delete this path" and has no chunks.
+	verifyDelete bool
+	// oldNoNewline records a unified diff's "\ No newline at end of file"
+	// on the removed side, i.e. the old content did not end with a newline.
+	oldNoNewline bool
 }
 
+type eofNewlineMode uint8
+
+const (
+	eofNewlineKeep eofNewlineMode = iota
+	eofNewlinePresent
+	eofNewlineAbsent
+)
+
 type structuredPatchChunk struct {
 	context          string
 	hasContext       bool
@@ -46,6 +72,11 @@ type structuredPatchChunk struct {
 	new              []string
 	newSourceOffsets []int
 	endOfFile        bool
+	// hint is the 0-based line the hunk is expected at (from a unified diff's
+	// "@@ -a,b" range). When the expected lines match there it is used as-is;
+	// otherwise the hunk is located by content like a structured hunk.
+	hint    int
+	hasHint bool
 }
 
 type structuredPatchTarget struct {
@@ -63,8 +94,36 @@ type structuredPatchChange struct {
 	mode   os.FileMode
 }
 
+// unifiedHunkRangePattern recognises a unified-diff range header ("-12,4 +12,6",
+// optionally followed by "@@ heading") written inside a structured hunk marker.
+// The line numbers carry no information for the context-anchored grammar, so
+// the range is dropped and only an explicit heading is kept as the anchor.
+var unifiedHunkRangePattern = regexp.MustCompile(`^-\d+(?:,\d+)?\s+\+\d+(?:,\d+)?\s*(?:@@\s*(.*))?$`)
+
+const structuredPatchFormatHint = ` (format: "*** Begin Patch", then "*** Update File: path" / "*** Add File: path" / "*** Delete File: path" sections whose hunks start with "@@" and use " " context, "-" removed and "+" added lines, then "*** End Patch")`
+
+// structuredPatchMarker classifies a line as the "begin" or "end" marker of a
+// structured patch, or "" when it is neither. The classifier is owned by the
+// sandbox package so the boundary check and the tool accept exactly the same
+// spellings; a strict byte-equal match here once made every patch from some
+// models fail on line 1 and pushed them into whole-file rewrites.
+func structuredPatchMarker(line string) string {
+	return sandbox.StructuredPatchMarker(line)
+}
+
+// structuredHunkAnchor normalises the text after a hunk's "@@ " marker. A
+// unified-diff range is dropped (its optional heading survives as the anchor);
+// anything else is used verbatim.
+func structuredHunkAnchor(context string) string {
+	context = strings.TrimSpace(context)
+	if match := unifiedHunkRangePattern.FindStringSubmatch(context); match != nil {
+		return strings.TrimSpace(match[1])
+	}
+	return context
+}
+
 func isStructuredPatch(patch string) bool {
-	return strings.HasPrefix(strings.TrimSpace(strings.TrimPrefix(patch, "\ufeff")), structuredPatchBegin)
+	return sandbox.IsStructuredPatch(patch)
 }
 
 func (tool applyPatchTool) runStructuredPatch(applyRoot, relativeRoot, patch string, options RunOptions) Result {
@@ -72,6 +131,14 @@ func (tool applyPatchTool) runStructuredPatch(applyRoot, relativeRoot, patch str
 	if err != nil {
 		return errorResult("Error applying patch: " + err.Error())
 	}
+	return applyPatchOperations(applyRoot, relativeRoot, operations, options)
+}
+
+// applyPatchOperations applies parsed operations (from either patch format)
+// through an opened workspace root: every stat, read, create and write is
+// descriptor-relative and refuses to follow a link out of the root, so there
+// is no pathname check-to-use window between validation and write.
+func applyPatchOperations(applyRoot, relativeRoot string, operations []structuredPatchOperation, options RunOptions) Result {
 	workspace, err := os.OpenRoot(applyRoot)
 	if err != nil {
 		return errorResult("Error applying patch: " + err.Error())
@@ -84,7 +151,7 @@ func (tool applyPatchTool) runStructuredPatch(applyRoot, relativeRoot, patch str
 	wholeBefore := make(map[string]bool, len(changes))
 	if options.FileTracker != nil {
 		for _, change := range changes {
-			if change.kind == structuredPatchUpdate {
+			if change.kind == structuredPatchUpdate || change.kind == structuredPatchCopy {
 				wholeBefore[change.from.absolute] = options.FileTracker.SeenWhole(change.from.absolute)
 			}
 		}
@@ -103,6 +170,10 @@ func (tool applyPatchTool) runStructuredPatch(applyRoot, relativeRoot, patch str
 			wasWhole := wholeBefore[change.from.absolute]
 			options.FileTracker.Forget(change.from.absolute)
 			recordStructuredPatchFile(options.FileTracker, change.to.absolute, false, wasWhole)
+		case structuredPatchCopy:
+			// The source is untouched; the destination inherits only what the
+			// model had actually seen of the source.
+			recordStructuredPatchFile(options.FileTracker, change.to.absolute, true, wholeBefore[change.from.absolute])
 		}
 	}
 
@@ -122,11 +193,11 @@ func parseStructuredPatch(patch string) ([]structuredPatchOperation, error) {
 		return nil, fmt.Errorf("structured patch is empty")
 	}
 	lines := strings.Split(normalized, "\n")
-	if strings.TrimSpace(lines[0]) != structuredPatchBegin {
-		return nil, fmt.Errorf("the first line of a structured patch must be %q", structuredPatchBegin)
+	if structuredPatchMarker(lines[0]) != "begin" {
+		return nil, fmt.Errorf("the first line of a structured patch must be %q%s", structuredPatchBegin, structuredPatchFormatHint)
 	}
-	if strings.TrimSpace(lines[len(lines)-1]) != structuredPatchEnd {
-		return nil, fmt.Errorf("the last line of a structured patch must be %q", structuredPatchEnd)
+	if structuredPatchMarker(lines[len(lines)-1]) != "end" {
+		return nil, fmt.Errorf("the last line of a structured patch must be %q%s", structuredPatchEnd, structuredPatchFormatHint)
 	}
 
 	var operations []structuredPatchOperation
@@ -191,8 +262,12 @@ func parseStructuredPatch(patch string) ([]structuredPatchOperation, error) {
 				}
 				if trimmed == "@@" || strings.HasPrefix(trimmed, structuredHunkContext) {
 					chunk := structuredPatchChunk{}
+					rest := ""
 					if trimmed != "@@" {
-						chunk.context = strings.TrimPrefix(trimmed, structuredHunkContext)
+						rest = strings.TrimPrefix(trimmed, structuredHunkContext)
+					}
+					if anchor := structuredHunkAnchor(rest); anchor != "" {
+						chunk.context = anchor
 						chunk.hasContext = true
 					}
 					op.chunks = append(op.chunks, chunk)
@@ -248,7 +323,7 @@ func structuredPatchPath(header, prefix string, line int) (string, error) {
 }
 
 func isStructuredPatchHeader(line string) bool {
-	return line == structuredPatchEnd || strings.HasPrefix(line, structuredAddFile) || strings.HasPrefix(line, structuredDeleteFile) || strings.HasPrefix(line, structuredUpdateFile)
+	return structuredPatchMarker(line) == "end" || strings.HasPrefix(line, structuredAddFile) || strings.HasPrefix(line, structuredDeleteFile) || strings.HasPrefix(line, structuredUpdateFile)
 }
 
 func structuredPatchLineError(line int, message string) error {
@@ -318,7 +393,7 @@ func planStructuredPatch(root *os.Root, operations []structuredPatchOperation, t
 				return nil, err
 			}
 			change.after = operation.contents
-		case structuredPatchDelete, structuredPatchUpdate:
+		case structuredPatchDelete, structuredPatchUpdate, structuredPatchCopy:
 			info, err := root.Stat(from.relative)
 			if err != nil {
 				return nil, fmt.Errorf("stating %s: %w", from.relative, err)
@@ -332,12 +407,20 @@ func planStructuredPatch(root *os.Root, operations []structuredPatchOperation, t
 				return nil, fmt.Errorf("%s", fileConflictMessage(from.relative))
 			}
 			change.before = string(content)
-			if operation.kind == structuredPatchUpdate {
+			if operation.kind == structuredPatchDelete && operation.verifyDelete {
+				if err := verifyUnifiedDeletion(change.before, from.relative, operation); err != nil {
+					return nil, err
+				}
+			}
+			if operation.kind == structuredPatchUpdate || operation.kind == structuredPatchCopy {
 				updated, err := applyStructuredPatchUpdate(change.before, from.relative, operation.chunks)
 				if err != nil {
 					return nil, err
 				}
-				change.after = updated
+				change.after = applyEOFNewline(updated, operation.eofNewline)
+				if operation.kind == structuredPatchCopy && from.absolute == to.absolute {
+					return nil, fmt.Errorf("cannot copy %s onto itself", from.relative)
+				}
 				if from.absolute != to.absolute {
 					if _, err := root.Lstat(to.relative); err == nil {
 						return nil, fmt.Errorf("cannot move %s to %s because the destination already exists", from.relative, to.relative)
@@ -353,10 +436,14 @@ func planStructuredPatch(root *os.Root, operations []structuredPatchOperation, t
 }
 
 func resolveStructuredPatchTarget(root, path string) (structuredPatchTarget, error) {
-	if filepath.IsAbs(path) || path == ".." || strings.HasPrefix(filepath.ToSlash(path), "../") {
+	// Relative traversal is rejected outright. An absolute path is allowed when
+	// it resolves inside the apply root (models routinely echo the absolute path
+	// they were shown by read_file); resolveWorkspaceTargetPath rejects anything
+	// that lands outside.
+	if path == ".." || strings.HasPrefix(filepath.ToSlash(path), "../") {
 		return structuredPatchTarget{}, fmt.Errorf("patch path %q must stay inside the workspace", path)
 	}
-	absolute, relative, err := resolveWorkspaceTargetPath(root, path)
+	absolute, relative, err := resolveWorkspaceTargetPath(root, normalizePatchPathForRoot(root, path))
 	if err != nil {
 		return structuredPatchTarget{}, err
 	}
@@ -389,9 +476,15 @@ func applyStructuredPatchUpdate(content, path string, chunks []structuredPatchCh
 		if len(chunk.old) == 0 {
 			// A context anchor identifies the line immediately before a pure
 			// insertion. Only an explicit end-of-file hunk (or an unanchored
-			// insertion) belongs at the file end.
+			// insertion) belongs at the file end. A unified diff names the
+			// insertion point in its range instead; a zero-context hunk (as
+			// from `git diff -U0`) has nothing to verify the position against,
+			// so it is position-trusting by design — the same semantics git
+			// apply uses without fuzz. Hunks that carry context are verified.
 			start := lineIndex
-			if chunk.endOfFile || !chunk.hasContext {
+			if chunk.hasHint {
+				start = max(chunk.hint, lineIndex)
+			} else if chunk.endOfFile || !chunk.hasContext {
 				start = len(lines)
 			}
 			if start > len(lines) {
@@ -405,7 +498,13 @@ func applyStructuredPatchUpdate(content, path string, chunks []structuredPatchCh
 			lineIndex = start
 			continue
 		}
-		index, ambiguous := findStructuredPatchSequence(lines, chunk.old, lineIndex, chunk.endOfFile)
+		var index int
+		ambiguous := false
+		if chunk.hasHint && chunk.hint >= lineIndex && sequenceMatchesAt(lines, chunk.old, chunk.hint) {
+			index = chunk.hint
+		} else {
+			index, ambiguous = findStructuredPatchSequence(lines, chunk.old, lineIndex, chunk.endOfFile)
+		}
 		if ambiguous {
 			return "", fmt.Errorf("expected lines are ambiguous in %s; provide more surrounding context:\n%s", path, strings.Join(chunk.old, "\n"))
 		}
@@ -440,6 +539,64 @@ func applyStructuredPatchUpdate(content, path string, chunks []structuredPatchCh
 	return updated, nil
 }
 
+// verifyUnifiedDeletion checks, byte for byte, that the lines a unified
+// deletion removes are exactly the file's current content. A deletion is not
+// recoverable, so none of the whitespace tolerance used to locate update
+// hunks applies here; only the file's own line-ending style is normalised.
+// With no hunks (git's header-only form) the file must be empty.
+func verifyUnifiedDeletion(current, path string, operation structuredPatchOperation) error {
+	var removed []string
+	for _, chunk := range operation.chunks {
+		if len(chunk.new) > 0 {
+			return fmt.Errorf("deletion of %s must not add lines", path)
+		}
+		removed = append(removed, chunk.old...)
+	}
+	expected := strings.Join(removed, "\n")
+	if len(removed) > 0 && !operation.oldNoNewline {
+		expected += "\n"
+	}
+	actual := current
+	if structuredPatchLineEnding(current) == "\r\n" {
+		actual = strings.ReplaceAll(current, "\r\n", "\n")
+	}
+	if actual != expected {
+		if len(removed) == 0 {
+			return fmt.Errorf("deletion of %s expects an empty file but it has content; include the removed lines in the patch", path)
+		}
+		return fmt.Errorf("deletion of %s does not match its current content; the removed lines must equal the whole file", path)
+	}
+	return nil
+}
+
+// sequenceMatchesAt reports whether wanted appears verbatim at lines[start:].
+func sequenceMatchesAt(lines, wanted []string, start int) bool {
+	if start < 0 || start+len(wanted) > len(lines) {
+		return false
+	}
+	for offset, line := range wanted {
+		if lines[start+offset] != line {
+			return false
+		}
+	}
+	return true
+}
+
+// applyEOFNewline forces or strips a single trailing line ending when a
+// unified diff's end-of-file marker asked for it.
+func applyEOFNewline(content string, mode eofNewlineMode) string {
+	switch mode {
+	case eofNewlinePresent:
+		if content != "" && !strings.HasSuffix(content, "\n") {
+			return content + structuredPatchLineEnding(content)
+		}
+	case eofNewlineAbsent:
+		content = strings.TrimSuffix(content, "\n")
+		content = strings.TrimSuffix(content, "\r")
+	}
+	return content
+}
+
 func structuredPatchReplacement(chunk structuredPatchChunk, source []string, start int) ([]string, error) {
 	if len(chunk.newSourceOffsets) != len(chunk.new) {
 		return nil, fmt.Errorf("context mapping has %d entries for %d output lines", len(chunk.newSourceOffsets), len(chunk.new))
@@ -516,14 +673,19 @@ func findStructuredPatchSequence(lines, wanted []string, start int, endOfFile bo
 }
 
 func applyStructuredPatchChanges(root *os.Root, changes []structuredPatchChange, tracker *FileTracker) error {
-	mutated := false
+	// committed lists, in order, the paths whose change reached disk before a
+	// later change failed, so the caller (and the model) knows exactly which
+	// files now hold the patched content and which were never touched.
+	var committed []string
 	for _, change := range changes {
-		committed, err := applyStructuredPatchChange(root, change)
-		mutated = mutated || committed
+		done, err := applyStructuredPatchChange(root, change)
+		if done {
+			committed = append(committed, structuredPatchChangePaths(change)...)
+		}
 		if err != nil {
 			forgetStructuredPatchFiles(tracker, changes)
-			if mutated {
-				return fmt.Errorf("%w; structured patch was partially applied; re-read affected files before retrying", err)
+			if len(committed) > 0 {
+				return fmt.Errorf("%w; patch was partially applied — already committed: %s; the remaining files are unchanged; re-read the committed files before retrying", err, strings.Join(committed, ", "))
 			}
 			return err
 		}
@@ -531,6 +693,19 @@ func applyStructuredPatchChanges(root *os.Root, changes []structuredPatchChange,
 	return nil
 }
 
+// structuredPatchChangePaths names the workspace-relative paths a committed
+// change touched: the destination, plus the source of a move or copy.
+func structuredPatchChangePaths(change structuredPatchChange) []string {
+	if change.kind == structuredPatchDelete {
+		return []string{change.from.relative}
+	}
+	paths := []string{change.to.relative}
+	if change.from.absolute != change.to.absolute && change.from.relative != "" {
+		paths = append([]string{change.from.relative}, paths...)
+	}
+	return paths
+}
+
 func forgetStructuredPatchFiles(tracker *FileTracker, changes []structuredPatchChange) {
 	if tracker == nil {
 		return
@@ -543,14 +718,35 @@ func forgetStructuredPatchFiles(tracker *FileTracker, changes []structuredPatchC
 	}
 }
 
+// structuredPatchBeforeCommit, when set, runs before each change's pre-commit
+// recheck. Tests use it to alter the filesystem between planning and commit
+// deterministically; it is nil in production.
+var structuredPatchBeforeCommit func(change structuredPatchChange)
+
 func applyStructuredPatchChange(root *os.Root, change structuredPatchChange) (bool, error) {
+	if structuredPatchBeforeCommit != nil {
+		structuredPatchBeforeCommit(change)
+	}
+	// The planner validated the source against the bytes it read; re-read
+	// through the same root immediately before committing so a change made
+	// by another process in between is refused rather than overwritten or
+	// removed.
+	if change.kind != structuredPatchAdd {
+		current, err := root.ReadFile(change.from.relative)
+		if err != nil {
+			return false, fmt.Errorf("re-reading %s before commit: %w", change.from.relative, err)
+		}
+		if string(current) != change.before {
+			return false, fmt.Errorf("%s changed on disk between planning and commit; re-read it and retry", change.from.relative)
+		}
+	}
 	switch change.kind {
 	case structuredPatchDelete:
 		if err := root.Remove(change.from.relative); err != nil {
 			return false, fmt.Errorf("deleting %s: %w", change.from.relative, err)
 		}
 		return true, nil
-	case structuredPatchAdd:
+	case structuredPatchAdd, structuredPatchCopy:
 		return writeStructuredPatchFile(root, change.to, change.after, change.mode, true)
 	case structuredPatchUpdate:
 		moving := change.from.absolute != change.to.absolute
@@ -706,7 +902,7 @@ func recordStructuredPatchFile(tracker *FileTracker, absolute string, created, s
 	info, _ := os.Stat(absolute)
 	tracker.Record(absolute, content, info)
 	if seenWhole {
-		lines := lineCount(string(content))
+		lines := trackedLineTotal(string(content))
 		tracker.RecordSeenRange(absolute, 1, lines, lines)
 	}
 	if created {

internal/tools/tool_search.go

@@ -58,6 +58,11 @@ func NewToolSearchTool(registry *Registry) Tool {
 	}
 }
 
+func (tool toolSearchTool) cloneForRegistry(registry *Registry) Tool {
+	tool.registry = registry
+	return tool
+}
+
 // Run satisfies the Tool interface; actual dispatch goes through RunWithOptions.
 func (tool toolSearchTool) Run(ctx context.Context, args map[string]any) Result {
 	return tool.RunWithOptions(ctx, args, RunOptions{})

internal/tools/tool_search_test.go

@@ -73,6 +73,37 @@ func newDeferredFixtureRegistry() *Registry {
 	return reg
 }
 
+func TestRegistryCloneKeepsToolSearchOnFrozenSnapshot(t *testing.T) {
+	live := newDeferredFixtureRegistry()
+	live.Register(NewToolSearchTool(live))
+	runRegistry := live.Clone()
+	live.Register(searchFakeTool{
+		name:        "late_lookup",
+		description: "Registered after the run snapshot.",
+		parameters:  Schema{Type: "object", AdditionalProperties: false},
+	})
+
+	result := runRegistry.Run(context.Background(), ToolSearchToolName, map[string]any{
+		"query": "select:late_lookup",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("tool_search status = %s: %s", result.Status, result.Output)
+	}
+	if loaded := result.Meta["load_tools"]; loaded != "" {
+		t.Fatalf("snapshot tool_search loaded late tool %q", loaded)
+	}
+	if direct := runRegistry.Run(context.Background(), "late_lookup", map[string]any{}); direct.Status != StatusError {
+		t.Fatalf("late tool unexpectedly executable in run snapshot: %#v", direct)
+	}
+
+	known := runRegistry.Run(context.Background(), ToolSearchToolName, map[string]any{
+		"query": "select:weather_lookup",
+	})
+	if loaded := known.Meta["load_tools"]; loaded != "weather_lookup" {
+		t.Fatalf("snapshot tool_search lost existing tool: load_tools=%q output=%q", loaded, known.Output)
+	}
+}
+
 func TestToolSearchExposesExpectedSafetyAndSchema(t *testing.T) {
 	tool := NewToolSearchTool(NewRegistry())
 

internal/tools/tracked_line_total_test.go

@@ -0,0 +1,59 @@
+package tools
+
+import (
+	"context"
+	"fmt"
+	"path/filepath"
+	"testing"
+)
+
+func TestTrackedLineTotalMatchesReadFileStats(t *testing.T) {
+	root := t.TempDir()
+	for name, content := range map[string]string{"trailing": "a\nb\nc\n", "unterminated": "a\nb\nc", "single": "x\n", "empty": "", "crlf": "a\r\nb\r\n"} {
+		path := filepath.Join(root, name+".txt")
+		writeTestFile(t, path, content)
+		stats, err := scanReadFileStats(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if got := trackedLineTotal(content); got != stats.lines {
+			t.Fatalf("%s: trackedLineTotal = %d, read_file reports %d", name, got, stats.lines)
+		}
+	}
+}
+
+// A whole-file read, an edit, then a partial re-read must keep the file "seen
+// whole": the writer and reader now agree on the line total, so the partial
+// read accumulates instead of resetting and a later edit far from the re-read
+// window is not refused.
+func TestEditThenPartialReadKeepsWholeFileObservation(t *testing.T) {
+	root := t.TempDir()
+	path := filepath.Join(root, "notes.txt")
+	content := ""
+	for i := 1; i <= 40; i++ {
+		content += fmt.Sprintf("line %02d\n", i)
+	}
+	writeTestFile(t, path, content)
+	tracker := NewFileTracker()
+	registry := NewRegistry()
+	registry.Register(NewScopedReadFileTool(root, nil))
+	registry.Register(NewScopedEditFileTool(root, nil))
+	opts := RunOptions{PermissionGranted: true, FileTracker: tracker}
+
+	if r := registry.RunWithOptions(context.Background(), "read_file", map[string]any{"path": "notes.txt"}, opts); r.Status != StatusOK {
+		t.Fatalf("read: %s", r.Output)
+	}
+	if r := registry.RunWithOptions(context.Background(), "edit_file", map[string]any{"path": "notes.txt", "old_string": "line 01\nline 02\n", "new_string": "line 01\nline two\n"}, opts); r.Status != StatusOK {
+		t.Fatalf("first edit: %s", r.Output)
+	}
+	if r := registry.RunWithOptions(context.Background(), "read_file", map[string]any{"path": "notes.txt", "limit": 5}, opts); r.Status != StatusOK {
+		t.Fatalf("partial read: %s", r.Output)
+	}
+	absolute, _ := filepath.EvalSymlinks(path)
+	if !tracker.SeenWhole(absolute) {
+		t.Fatal("partial re-read of an unchanged file must not discard whole-file knowledge")
+	}
+	if r := registry.RunWithOptions(context.Background(), "edit_file", map[string]any{"path": "notes.txt", "old_string": "line 39\nline 40\n", "new_string": "line 39\nline forty\n"}, opts); r.Status != StatusOK {
+		t.Fatalf("edit far from the re-read window must succeed: %s", r.Output)
+	}
+}

internal/tools/types.go

@@ -211,6 +211,14 @@ type Tool interface {
 	Run(ctx context.Context, args map[string]any) Result
 }
 
+// IsBuiltInApplyPatch reports whether tool is Zero's scoped patch tool. The
+// marker method is package-private so external and MCP tools cannot claim this
+// identity merely by sharing the public name "apply_patch".
+func IsBuiltInApplyPatch(tool Tool) bool {
+	_, ok := tool.(interface{ isBuiltInApplyPatch() })
+	return ok
+}
+
 // ArgsPermissioner is an optional interface a Tool can implement to refine its
 // permission for a SPECIFIC call based on its arguments. When a tool implements
 // it, the agent loop consults PermissionForArgs(args) instead of the static

internal/tools/unified_patch.go

@@ -0,0 +1,393 @@
+package tools
+
+import (
+	"fmt"
+	"strings"
+)
+
+// parseUnifiedPatch converts a unified diff into the same operations a
+// structured patch produces, so both formats are applied by Zero itself
+// through an opened workspace root (descriptor-relative, no-follow writes)
+// instead of handing pathnames to git apply after validation.
+//
+// Supported: file modifications, creations (--- /dev/null) and deletions
+// (+++ /dev/null, verified against the content the hunks expect to remove),
+// multiple hunks per file, "\ No newline at end of file"
+// markers, a/ and b/ prefixes, git C-quoted paths, CRLF patches, and git's
+// "rename from/to" and "copy from/to" headers (with or without hunks). Headers
+// git emits around a hunk (diff --git, index, mode lines) are skipped; binary
+// patches are rejected.
+func parseUnifiedPatch(patch string) ([]structuredPatchOperation, error) {
+	normalized := strings.TrimPrefix(strings.ReplaceAll(patch, "\r\n", "\n"), "\ufeff")
+	lines := strings.Split(normalized, "\n")
+	// The element after a final "\n" is the line terminator, not an empty
+	// context line; counting it would let a truncated hunk look complete.
+	if len(lines) > 0 && lines[len(lines)-1] == "" {
+		lines = lines[:len(lines)-1]
+	}
+
+	var operations []structuredPatchOperation
+	var current *structuredPatchOperation
+	var chunk *structuredPatchChunk
+	var added []string // collected "+" lines for a file creation
+	oldPath, newPath := "", ""
+	pendingFrom, pendingKind := "", structuredPatchUpdate
+	// git's header-only forms: "deleted file mode" / "new file mode" after a
+	// "diff --git" line with no ---/+++ pair describe an empty file.
+	diffPath, headerOnly := "", byte(0) // headerOnly is 'd' (deleted) or 'n' (new)
+	oldRemaining, newRemaining := 0, 0
+	inHunk := false
+	lastSide := byte(0) // '+' or '-' for the most recent content line
+
+	finish := func() error {
+		if current == nil {
+			return nil
+		}
+		switch current.kind {
+		case structuredPatchAdd:
+			if len(added) > 0 {
+				current.contents = strings.Join(added, "\n")
+				if current.eofNewline != eofNewlineAbsent {
+					current.contents += "\n"
+				}
+			}
+		case structuredPatchUpdate, structuredPatchCopy:
+			if !allStructuredPatchChunksHaveContent(current.chunks) {
+				return fmt.Errorf("unified diff for %s has an empty hunk", current.path)
+			}
+			if len(current.chunks) == 0 && current.movePath == "" {
+				return fmt.Errorf("unified diff for %s has no hunk lines", current.path)
+			}
+		case structuredPatchDelete:
+			if len(current.chunks) == 0 || !allStructuredPatchChunksHaveContent(current.chunks) {
+				return fmt.Errorf("unified deletion of %s must include the hunk with the content being removed", current.path)
+			}
+		}
+		operations = append(operations, *current)
+		current, chunk, added = nil, nil, nil
+		return nil
+	}
+	flushHeaderOnly := func(line int) error {
+		if headerOnly == 0 {
+			return nil
+		}
+		if diffPath == "" {
+			return fmt.Errorf("invalid unified diff at line %d: file mode header without a diff --git path", line)
+		}
+		if err := finish(); err != nil {
+			return err
+		}
+		switch headerOnly {
+		case 'd':
+			operations = append(operations, structuredPatchOperation{kind: structuredPatchDelete, path: diffPath, line: line, verifyDelete: true})
+		case 'n':
+			operations = append(operations, structuredPatchOperation{kind: structuredPatchAdd, path: diffPath, line: line})
+		}
+		headerOnly, diffPath = 0, ""
+		return nil
+	}
+	startFile := func(line int) error {
+		// A ---/+++ pair means the file has hunks; it is not header-only.
+		headerOnly, diffPath = 0, ""
+		if oldPath == "" || newPath == "" {
+			return fmt.Errorf("invalid unified diff at line %d: hunk before a ---/+++ header pair", line)
+		}
+		// A ---/+++ pair after a rename/copy header names the same files; keep
+		// accumulating that operation's hunks instead of starting a new one.
+		if current != nil && current.movePath != "" && current.path == oldPath && (current.movePath == newPath || oldPath == newPath) {
+			return nil
+		}
+		if err := finish(); err != nil {
+			return err
+		}
+		op := structuredPatchOperation{line: line}
+		switch {
+		case oldPath == "/dev/null" && newPath == "/dev/null":
+			return fmt.Errorf("invalid unified diff at line %d: both sides are /dev/null", line)
+		case oldPath == "/dev/null":
+			op.kind, op.path = structuredPatchAdd, newPath
+		case newPath == "/dev/null":
+			// A unified deletion states the content it expects to remove; keep
+			// its hunks so the planner verifies them before deleting.
+			op.kind, op.path, op.verifyDelete = structuredPatchDelete, oldPath, true
+		default:
+			op.kind, op.path = structuredPatchUpdate, oldPath
+			if newPath != oldPath {
+				op.movePath = newPath
+			}
+		}
+		current = &op
+		return nil
+	}
+
+	for index, raw := range lines {
+		lineNumber := index + 1
+		if inHunk && (oldRemaining > 0 || newRemaining > 0) {
+			// A "--- " line followed by "+++ " and then a "@@" hunk header is
+			// the next file's header pair, not a removed "-- x" line followed by
+			// an added "++ y" line: an over-declared count would otherwise
+			// swallow both and the hunk would later fail against the wrong
+			// file. The "@@" requirement keeps a genuine adjacent pair intact.
+			if strings.HasPrefix(raw, "--- ") && index+2 < len(lines) && strings.HasPrefix(lines[index+1], "+++ ") && strings.HasPrefix(lines[index+2], "@@") {
+				return nil, fmt.Errorf("invalid unified diff at line %d: hunk ended before its declared line counts", lineNumber)
+			}
+			switch {
+			case strings.HasPrefix(raw, "\\"):
+				// "\ No newline at end of file" qualifies the previous line's side.
+				if current != nil && lastSide == '+' {
+					current.eofNewline = eofNewlineAbsent
+				} else if current != nil && lastSide == '-' {
+					current.oldNoNewline = true
+					if current.eofNewline == eofNewlineKeep {
+						current.eofNewline = eofNewlinePresent
+					}
+				}
+				continue
+			case strings.HasPrefix(raw, "-"):
+				oldRemaining--
+				lastSide = '-'
+				if chunk != nil {
+					chunk.old = append(chunk.old, raw[1:])
+				}
+			case strings.HasPrefix(raw, "+"):
+				newRemaining--
+				lastSide = '+'
+				if current != nil && current.kind == structuredPatchAdd {
+					added = append(added, raw[1:])
+				} else if chunk != nil {
+					chunk.new = append(chunk.new, raw[1:])
+					chunk.newSourceOffsets = append(chunk.newSourceOffsets, -1)
+				}
+			default:
+				content := raw
+				if strings.HasPrefix(raw, " ") {
+					content = raw[1:]
+				} else if raw != "" {
+					return nil, fmt.Errorf("invalid unified diff at line %d: hunk lines must start with ' ', '+', '-' or '\\'", lineNumber)
+				}
+				oldRemaining--
+				newRemaining--
+				lastSide = ' '
+				if chunk != nil {
+					offset := len(chunk.old)
+					chunk.old = append(chunk.old, content)
+					chunk.new = append(chunk.new, content)
+					chunk.newSourceOffsets = append(chunk.newSourceOffsets, offset)
+				}
+			}
+			continue
+		}
+		inHunk = false
+		trimmed := strings.TrimSpace(raw)
+		if oldRemaining > 0 || newRemaining > 0 {
+			// The hunk ended before its declared counts were consumed. Report
+			// it here rather than absorbing the next file's ---/+++ headers as
+			// "-"/"+" content and failing later against the wrong file.
+			return nil, fmt.Errorf("invalid unified diff at line %d: hunk ended before its declared line counts", lineNumber)
+		}
+		switch {
+		case trimmed == "":
+			continue
+		case strings.HasPrefix(raw, "\\"):
+			// "\ No newline at end of file" directly after a hunk's last line.
+			if current != nil && lastSide == '+' {
+				current.eofNewline = eofNewlineAbsent
+			} else if current != nil && lastSide == '-' {
+				current.oldNoNewline = true
+				if current.eofNewline == eofNewlineKeep {
+					current.eofNewline = eofNewlinePresent
+				}
+			}
+			continue
+		case strings.HasPrefix(raw, "diff --git "):
+			if err := flushHeaderOnly(lineNumber); err != nil {
+				return nil, err
+			}
+			diffPath = diffGitNewPath(raw)
+			continue
+		case strings.HasPrefix(raw, "deleted file mode "):
+			headerOnly = 'd'
+			continue
+		case strings.HasPrefix(raw, "new file mode "):
+			headerOnly = 'n'
+			continue
+		case strings.HasPrefix(raw, "index "), strings.HasPrefix(raw, "old mode "), strings.HasPrefix(raw, "new mode "):
+			continue
+		case strings.HasPrefix(raw, "similarity index "), strings.HasPrefix(raw, "dissimilarity index "):
+			continue
+		case strings.HasPrefix(raw, "rename from "), strings.HasPrefix(raw, "copy from "):
+			pendingFrom = strings.TrimSpace(unquoteGitPath(strings.TrimPrefix(strings.TrimPrefix(raw, "rename from "), "copy from ")))
+			pendingKind = structuredPatchUpdate
+			if strings.HasPrefix(raw, "copy from ") {
+				pendingKind = structuredPatchCopy
+			}
+			if pendingFrom == "" {
+				return nil, fmt.Errorf("invalid unified diff at line %d: missing source path", lineNumber)
+			}
+		case strings.HasPrefix(raw, "rename to "), strings.HasPrefix(raw, "copy to "):
+			to := strings.TrimSpace(unquoteGitPath(strings.TrimPrefix(strings.TrimPrefix(raw, "rename to "), "copy to ")))
+			if pendingFrom == "" || to == "" {
+				return nil, fmt.Errorf("invalid unified diff at line %d: rename/copy destination without a source", lineNumber)
+			}
+			if err := finish(); err != nil {
+				return nil, err
+			}
+			current = &structuredPatchOperation{kind: pendingKind, path: pendingFrom, movePath: to, line: lineNumber}
+			oldPath, newPath, pendingFrom = pendingFrom, to, ""
+		case strings.HasPrefix(raw, "Binary files "), strings.HasPrefix(raw, "GIT binary patch"):
+			return nil, fmt.Errorf("invalid unified diff at line %d: binary patches are not supported", lineNumber)
+		case strings.HasPrefix(raw, "--- "):
+			oldPath = stripPatchPrefix(patchFileHeaderPath(raw))
+			newPath = ""
+			if oldPath == "" {
+				return nil, fmt.Errorf("invalid unified diff at line %d: missing path in --- header", lineNumber)
+			}
+		case strings.HasPrefix(raw, "+++ "):
+			newPath = stripPatchPrefix(patchFileHeaderPath(raw))
+			if newPath == "" {
+				return nil, fmt.Errorf("invalid unified diff at line %d: missing path in +++ header", lineNumber)
+			}
+			if err := startFile(lineNumber); err != nil {
+				return nil, err
+			}
+		case strings.HasPrefix(raw, "@@"):
+			if current == nil {
+				return nil, fmt.Errorf("invalid unified diff at line %d: hunk before a ---/+++ header pair", lineNumber)
+			}
+			oldStart, oldCount, newCount, ok := parseHunkRange(raw)
+			if !ok {
+				return nil, fmt.Errorf("invalid unified diff at line %d: malformed hunk header", lineNumber)
+			}
+			// "\ No newline at end of file" describes the end of the file, so it
+			// may only follow a file's final hunk; a marker before another hunk
+			// would otherwise leak into that later hunk's result.
+			if current.eofNewline != eofNewlineKeep {
+				return nil, fmt.Errorf("invalid unified diff at line %d: \"\\ No newline at end of file\" must follow the last hunk of a file", lineNumber)
+			}
+			oldRemaining, newRemaining = oldCount, newCount
+			inHunk = oldRemaining > 0 || newRemaining > 0
+			lastSide = 0
+			if current.kind != structuredPatchAdd {
+				next := structuredPatchChunk{hasHint: true, hint: oldStart - 1}
+				if oldCount == 0 {
+					// A pure insertion's range names the line after which the
+					// new lines go, so the insertion index is oldStart itself.
+					next.hint = oldStart
+				}
+				if next.hint < 0 {
+					next.hint = 0
+				}
+				current.chunks = append(current.chunks, next)
+				chunk = &current.chunks[len(current.chunks)-1]
+			}
+		default:
+			return nil, fmt.Errorf("invalid unified diff at line %d: unexpected %q", lineNumber, trimmed)
+		}
+	}
+	// Strict at end of input on purpose: a hunk cut short (typically a model
+	// response that stopped mid-patch) would otherwise apply as a partial
+	// change — removals without their replacements — which is worse than a
+	// clear error and a retry. Miscounted ranges with complete content are
+	// still tolerated at every other boundary.
+	if inHunk && (oldRemaining > 0 || newRemaining > 0) {
+		return nil, fmt.Errorf("invalid unified diff at line %d: hunk ended before its declared line counts", len(lines))
+	}
+	if err := flushHeaderOnly(len(lines)); err != nil {
+		return nil, err
+	}
+	if err := finish(); err != nil {
+		return nil, err
+	}
+	if len(operations) == 0 {
+		return nil, fmt.Errorf("unified diff contains no file changes")
+	}
+	return operations, nil
+}
+
+// diffGitNewPath returns the post-image path named by a "diff --git a/X b/Y"
+// line, handling git's C-quoted form; "" when it cannot be determined.
+func diffGitNewPath(line string) string {
+	rest := strings.TrimSpace(strings.TrimPrefix(line, "diff --git "))
+	if strings.HasPrefix(rest, "\"") {
+		// Two quoted tokens: skip the first, unquote the second.
+		end := strings.Index(rest[1:], "\"")
+		for end > 0 && rest[end] == '\\' {
+			next := strings.Index(rest[end+2:], "\"")
+			if next < 0 {
+				return ""
+			}
+			end += next + 2
+		}
+		if end < 0 || end+2 > len(rest) {
+			return ""
+		}
+		rest = strings.TrimSpace(rest[end+2:])
+		return stripPatchPrefix(unquoteGitPath(rest))
+	}
+	fields := strings.Fields(rest)
+	if len(fields) < 2 {
+		return ""
+	}
+	last := fields[len(fields)-1]
+	if strings.HasPrefix(last, "\"") {
+		return stripPatchPrefix(unquoteGitPath(last))
+	}
+	return stripPatchPrefix(last)
+}
+
+// parseHunkRange reads "@@ -a[,b] +c[,d] @@" and returns a, b and d; a missing
+// count means 1 per unified-diff convention.
+func parseHunkRange(line string) (oldStart, oldCount, newCount int, ok bool) {
+	_, rest, found := strings.Cut(line, "@@")
+	if !found {
+		return 0, 0, 0, false
+	}
+	rangeSection := rest
+	if before, _, found := strings.Cut(rest, "@@"); found {
+		rangeSection = before
+	}
+	fields := strings.Fields(rangeSection)
+	if len(fields) != 2 || !strings.HasPrefix(fields[0], "-") || !strings.HasPrefix(fields[1], "+") {
+		return 0, 0, 0, false
+	}
+	parse := func(spec string) (int, int, bool) {
+		startText, countText, hasCount := strings.Cut(spec, ",")
+		start, err := parseNonNegativeInt(startText)
+		if err != nil {
+			return 0, 0, false
+		}
+		count := 1
+		if hasCount {
+			if count, err = parseNonNegativeInt(countText); err != nil {
+				return 0, 0, false
+			}
+		}
+		return start, count, true
+	}
+	oldStart, oldCount, ok = parse(fields[0][1:])
+	if !ok {
+		return 0, 0, 0, false
+	}
+	_, newCount, ok = parse(fields[1][1:])
+	if !ok {
+		return 0, 0, 0, false
+	}
+	return oldStart, oldCount, newCount, true
+}
+
+func parseNonNegativeInt(text string) (int, error) {
+	if text == "" {
+		return 0, fmt.Errorf("empty number")
+	}
+	value := 0
+	for _, r := range text {
+		if r < '0' || r > '9' {
+			return 0, fmt.Errorf("not a number: %q", text)
+		}
+		value = value*10 + int(r-'0')
+		if value > 1<<30 {
+			return 0, fmt.Errorf("number too large: %q", text)
+		}
+	}
+	return value, nil
+}

internal/tools/update_plan.go

@@ -26,7 +26,7 @@ func NewUpdatePlanTool() *updatePlanTool {
 	return &updatePlanTool{
 		baseTool: baseTool{
 			name: "update_plan",
-			description: "Create or update the in-memory plan for implementation or investigation with at least three meaningful dependent steps. Do not use it for simple lookups, explanations, or code navigation. " +
+			description: "Create or update the in-memory plan for work spanning multiple components or workstreams, sequencing uncertainty, or sustained tracking across many tool calls. Skip it for bounded changes in one component, simple lookups, explanations, and code navigation. " +
 				"Pass the full ordered list of steps each call; it replaces the previous plan. " +
 				"Each item needs a `content` string; `status` defaults to \"pending\" and `id` is " +
 				"auto-numbered, so you only need to supply `content` (and `status` as the task progresses). " +

internal/tools/write_file.go

@@ -120,7 +120,7 @@ func (tool writeFileTool) RunWithOptions(ctx context.Context, args map[string]an
 	newInfo, _ := os.Stat(absolutePath)
 	options.FileTracker.Record(absolutePath, []byte(content), newInfo)
 	if content == modelKnownContent {
-		options.FileTracker.RecordSeenRange(absolutePath, 1, lineCount(content), lineCount(content))
+		options.FileTracker.RecordSeenRange(absolutePath, 1, trackedLineTotal(content), trackedLineTotal(content))
 	}
 	if !existed {
 		options.FileTracker.RecordCreated(absolutePath)

internal/tools/write_tools_test.go

@@ -866,7 +866,9 @@ func TestStructuredPatchFailedDeleteDoesNotRecreateMissingFile(t *testing.T) {
 	}
 }
 
-func TestStructuredPatchMoveFailureLeavesPublishedDestination(t *testing.T) {
+// A move whose source vanishes after planning is refused by the pre-commit
+// recheck before the destination is published, so nothing is left half-done.
+func TestStructuredPatchMoveWithMissingSourceIsRefusedBeforePublishing(t *testing.T) {
 	root := t.TempDir()
 	workspace, err := os.OpenRoot(root)
 	if err != nil {
@@ -875,22 +877,35 @@ func TestStructuredPatchMoveFailureLeavesPublishedDestination(t *testing.T) {
 	defer workspace.Close()
 	sourcePath := filepath.Join(root, "source.txt")
 	destinationPath := filepath.Join(root, "destination.txt")
+	writeTestFile(t, sourcePath, "removed source\n")
 	change := structuredPatchChange{
 		kind:   structuredPatchUpdate,
 		from:   structuredPatchTarget{absolute: sourcePath, relative: "source.txt"},
 		to:     structuredPatchTarget{absolute: destinationPath, relative: "destination.txt"},
 		before: "removed source\n", after: "moved content\n", mode: 0o644,
 	}
+	// The source exists at planning time and disappears just before commit.
+	removed := false
+	structuredPatchBeforeCommit = func(structuredPatchChange) {
+		if err := os.Remove(sourcePath); err != nil {
+			t.Fatal(err)
+		}
+		removed = true
+	}
+	defer func() { structuredPatchBeforeCommit = nil }()
 
 	err = applyStructuredPatchChanges(workspace, []structuredPatchChange{change}, nil)
-	if err == nil || !strings.Contains(err.Error(), "partially applied") {
-		t.Fatalf("move with a missing source = %v, want partial-application error", err)
+	if !removed {
+		t.Fatal("pre-commit hook did not run")
+	}
+	if err == nil || !strings.Contains(err.Error(), "before commit") || strings.Contains(err.Error(), "partially applied") {
+		t.Fatalf("move with a missing source = %v, want a pre-commit refusal with nothing applied", err)
 	}
 	if _, statErr := os.Stat(sourcePath); !os.IsNotExist(statErr) {
 		t.Fatalf("failed move recreated a missing source: %v", statErr)
 	}
-	if got := mustReadTestFile(t, destinationPath); got != "moved content\n" {
-		t.Fatalf("published move destination = %q", got)
+	if _, statErr := os.Stat(destinationPath); !os.IsNotExist(statErr) {
+		t.Fatalf("refused move must not publish the destination: %v", statErr)
 	}
 }
 

internal/trace/trace.go

@@ -53,6 +53,7 @@ const (
 	CounterModelSwitches     = "model_switches"
 	CounterInputTokens       = "input_tokens"
 	CounterCachedInputTokens = "cached_input_tokens"
+	CounterCacheWriteTokens  = "cache_write_tokens"
 	CounterOutputTokens      = "output_tokens"
 	CounterPrewarmAttempts   = "prewarm_attempts"
 	CounterPrefixStable      = "prefix_stable"
@@ -64,6 +65,12 @@ const (
 	CounterPostureEscalations = "posture_escalations"
 )
 
+const (
+	CounterResponseChainReused   = "response_chain_reused"
+	CounterResponseChainReset    = "response_chain_reset"
+	CounterResponsesHTTPFallback = "responses_http_fallback"
+)
+
 // Span is one named wall interval attributed to part of a run. Each stamp is
 // its own entry (spans are not merged by name): a run that streams two model
 // completions has two generation entries, which lets the recorder derive
@@ -290,6 +297,7 @@ func OptionalEventKeys() []string {
 		"span:" + SpanVerification,
 		"counter:" + CounterToolCalls,
 		"counter:" + CounterCachedInputTokens,
+		"counter:" + CounterCacheWriteTokens,
 		"counter:" + CounterRetryCount,
 		"counter:" + CounterReconnectCount,
 		"counter:" + CounterCompactionCount,
@@ -301,6 +309,9 @@ func OptionalEventKeys() []string {
 		"counter:" + CounterPrewarmAttempts,
 		"counter:" + CounterPrefixStable,
 		"counter:" + CounterPrefixDrift,
+		"counter:" + CounterResponseChainReused,
+		"counter:" + CounterResponseChainReset,
+		"counter:" + CounterResponsesHTTPFallback,
 		"counter:" + CounterPostureEscalations,
 		"event:prefix_hash",
 		"task_state",

internal/trace/trace_test.go

@@ -614,11 +614,21 @@ func TestCounterPrecisionRoundTrip(t *testing.T) {
 	}
 }
 
-func TestOptionalEventKeysIncludePostureEscalations(t *testing.T) {
+func TestOptionalEventKeysIncludeRuntimeCounters(t *testing.T) {
+	keys := make(map[string]bool)
 	for _, key := range OptionalEventKeys() {
-		if key == "counter:"+CounterPostureEscalations {
-			return
+		keys[key] = true
+	}
+	for _, counter := range []string{
+		CounterCacheWriteTokens,
+		CounterResponseChainReused,
+		CounterResponseChainReset,
+		CounterResponsesHTTPFallback,
+		CounterPostureEscalations,
+	} {
+		key := "counter:" + counter
+		if !keys[key] {
+			t.Errorf("%s missing from OptionalEventKeys: %v", key, OptionalEventKeys())
 		}
 	}
-	t.Fatalf("posture_escalations missing from OptionalEventKeys: %v", OptionalEventKeys())
 }

internal/tui/model.go

@@ -85,10 +85,12 @@ type model struct {
 	savedProviders              []config.ProviderProfile
 	provider                    zeroruntime.Provider
 	newProvider                 func(config.ProviderProfile) (zeroruntime.Provider, error)
+	newTurnSessionProvider      func(config.ProviderProfile, zeroruntime.Provider) zeroruntime.TurnSessionProvider
 	probeProviderHealth         func(context.Context, providerhealth.Options) providerhealth.Result
 	discoverProviderModels      func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
 	discoverOllamaContextWindow func(ctx context.Context, baseURL string, model string) (int, error)
 	registry                    *tools.Registry
+	awaitToolReadiness          func(context.Context)
 	// lspManager is created once per session and reused across prompts so gopls (and
 	// other language servers) stay warm — a fresh manager per run would cold-start
 	// the server on the first edit of every turn. Nil when cwd is unknown; runs then
@@ -233,17 +235,20 @@ type model struct {
 	leaderHelpOverlay bool
 	// leaderPending is true after Ctrl+X until a second key, Esc, or timeout
 	// resolves the chord (see leader.go). leaderSeq invalidates a stale tick.
-	leaderPending         bool
-	leaderSeq             int
-	transcriptBodyHeights *transcriptBodyHeightCache
-	input                 textinput.Model
-	composer              composerState
-	composerActive        bool
-	composerCursorVisible bool
-	composerPastePreviews []composerPastePreview
-	composerSelection     composerSelectionState
-	dictation             dictationController
-	sttKeyPrompt          *sttKeyPromptState
+	leaderPending          bool
+	leaderSeq              int
+	transcriptBodyHeights  *transcriptBodyHeightCache
+	input                  textinput.Model
+	composer               composerState
+	composerActive         bool
+	composerCursorVisible  bool
+	composerBlinkSeq       int
+	composerBlinkIdleTicks int
+	composerBlinkTicking   bool
+	composerPastePreviews  []composerPastePreview
+	composerSelection      composerSelectionState
+	dictation              dictationController
+	sttKeyPrompt           *sttKeyPromptState
 	// plan holds the sticky plan panel state (steps, expansion, timings)
 	// synced from the update_plan tool. See plan_panel.go.
 	plan            planPanelState
@@ -491,10 +496,11 @@ type model struct {
 	// transientNotice is a single, replaceable confirmation shown above the
 	// composer. Unlike copyStatus it is shared by lightweight slash commands
 	// and is never persisted into the transcript.
-	transientNotice    transientNotice
-	transientNoticeSeq int
-	exitConfirmActive  bool
-	exitConfirmSeq     int
+	transientNotice         transientNotice
+	transientNoticeSeq      int
+	transientNoticeTimerSeq int
+	exitConfirmActive       bool
+	exitConfirmSeq          int
 	// cancelConfirmActive/cancelConfirmSeq mirror exitConfirmActive/exitConfirmSeq
 	// (same seq-gated tea.Tick pattern) but guard a DIFFERENT action: Esc
 	// cancelling a running turn. The two are deliberately separate state (not a
@@ -927,6 +933,13 @@ func newModel(ctx context.Context, options Options) model {
 	// terminal bracketed paste (Paste: true) is the single paste path.
 	input.KeyMap.Paste.SetEnabled(false)
 	input.Focus()
+	// The main composer paints its own cursor through composerCursorVisible.
+	// Keep the embedded textinput cursor static so the shared textinput.Blink
+	// message cannot start a second, invisible 530ms tick chain. Setup inputs
+	// still use the normal Bubbles cursor and continue blinking independently.
+	inputStyles := input.Styles()
+	inputStyles.Cursor.Blink = false
+	input.SetStyles(inputStyles)
 
 	runSpinner := spinner.New(spinner.WithSpinner(spinner.MiniDot))
 	runSpinner.Spinner.FPS = activeAnimationFrameInterval
@@ -951,6 +964,8 @@ func newModel(ctx context.Context, options Options) model {
 		userCommands:                loadedUserCommands,
 		loadSkills:                  options.LoadSkills,
 		composerCursorVisible:       true,
+		composerBlinkSeq:            1,
+		composerBlinkTicking:        true,
 		terminalFocused:             true,
 		userConfigPath:              options.UserConfigPath,
 		doctorUserConfigPath:        doctorUserConfigPath,
@@ -966,10 +981,12 @@ func newModel(ctx context.Context, options Options) model {
 		recapsEnabled:               options.RecapsEnabled,
 		provider:                    options.Provider,
 		newProvider:                 options.NewProvider,
+		newTurnSessionProvider:      options.NewTurnSessionProvider,
 		probeProviderHealth:         options.ProbeProviderHealth,
 		discoverProviderModels:      options.DiscoverProviderModels,
 		discoverOllamaContextWindow: options.DiscoverOllamaContextWindow,
 		registry:                    registry,
+		awaitToolReadiness:          options.AwaitToolReadiness,
 		sessionStore:                sessionStore,
 		peerService:                 options.PeerService,
 		sandboxStore:                sandboxStore,
@@ -1090,6 +1107,11 @@ const (
 // composerCursorBlinkInterval is the on/off period of the composer text cursor.
 const composerCursorBlinkInterval = 530 * time.Millisecond
 
+// composerBlinkIdleTicksBeforeSettle bounds the software cursor animation.
+// Once the terminal is focused but otherwise idle, the cursor finishes a short
+// blink sequence and settles visible instead of waking the TUI forever.
+const composerBlinkIdleTicksBeforeSettle = 4
+
 // composerTypingIdleThreshold is how long a typing pause must last before the
 // cursor resumes blinking; comfortably above normal inter-keystroke gaps
 // (~150-300ms) so it won't flicker mid-sentence.
@@ -1098,16 +1120,27 @@ const composerTypingIdleThreshold = 500 * time.Millisecond
 // composerBlinkMsg toggles the composer cursor's visibility each tick. The custom
 // composer render draws its own cursor (not textinput's), so it drives its own
 // blink rather than relying on textinput.Blink.
-type composerBlinkMsg struct{}
+type composerBlinkMsg struct{ seq int }
 
-func composerBlinkCmd() tea.Cmd {
+func composerBlinkCmd(seq int) tea.Cmd {
 	return tea.Tick(composerCursorBlinkInterval, func(time.Time) tea.Msg {
-		return composerBlinkMsg{}
+		return composerBlinkMsg{seq: seq}
 	})
 }
 
+func (m model) armComposerBlink() (model, tea.Cmd) {
+	m.composerCursorVisible = true
+	m.composerBlinkIdleTicks = 0
+	if m.composerBlinkTicking {
+		return m, nil
+	}
+	m.composerBlinkTicking = true
+	m.composerBlinkSeq++
+	return m, composerBlinkCmd(m.composerBlinkSeq)
+}
+
 func (m model) Init() tea.Cmd {
-	cmds := []tea.Cmd{textinput.Blink, composerBlinkCmd()}
+	cmds := []tea.Cmd{textinput.Blink, composerBlinkCmd(m.composerBlinkSeq)}
 	if m.petAnimation != nil && !m.reducedMotion {
 		cmds = append(cmds, petTickCmd(m.petTickSeq, m.petFrameDelay()))
 	}
@@ -1290,7 +1323,15 @@ func (m model) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 	nm = nm.syncChatScroll()
 	nm, mouseCmd := nm.syncMouseCapture()
 	nm, flushCmd := nm.settleTranscript()
-	return nm, batchCommands(cmd, mouseCmd, flushCmd, recapActivityCmd)
+	var composerCmd tea.Cmd
+	switch msg.(type) {
+	case tea.KeyPressMsg, tea.PasteMsg, tea.FocusMsg:
+		if nm.terminalFocused {
+			nm, composerCmd = nm.armComposerBlink()
+		}
+	}
+	nm, noticeCmd := nm.ensureTransientNoticeTimer()
+	return nm, batchCommands(cmd, mouseCmd, flushCmd, recapActivityCmd, composerCmd, noticeCmd)
 }
 
 func batchCommands(cmds ...tea.Cmd) tea.Cmd {
@@ -1353,16 +1394,27 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		}
 		return m, nil
 	case composerBlinkMsg:
-		m = m.expireTransientNotice()
+		if !m.composerBlinkTicking || msg.seq != m.composerBlinkSeq {
+			return m, nil
+		}
 		switch {
 		case !m.terminalFocused:
-			m.composerCursorVisible = false // hidden while unfocused
+			m.composerCursorVisible = false
+			m.composerBlinkTicking = false
+			return m, nil
 		case m.now().Sub(m.lastCharTime) < composerTypingIdleThreshold:
-			m.composerCursorVisible = true // solid while actively typing
+			m.composerCursorVisible = true
+			m.composerBlinkIdleTicks = 0
 		default:
-			m.composerCursorVisible = !m.composerCursorVisible // idle + focused: blink as before
+			m.composerBlinkIdleTicks++
+			if m.composerBlinkIdleTicks >= composerBlinkIdleTicksBeforeSettle {
+				m.composerCursorVisible = true
+				m.composerBlinkTicking = false
+				return m, nil
+			}
+			m.composerCursorVisible = !m.composerCursorVisible
 		}
-		return m, composerBlinkCmd()
+		return m, composerBlinkCmd(m.composerBlinkSeq)
 	case tea.BackgroundColorMsg:
 		// A background reading changes only the contrast direction used for named
 		// palettes. It never repaints the terminal canvas.
@@ -2189,6 +2241,8 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		}
 		m.terminalFocused = false
 		m.composerCursorVisible = false
+		m.composerBlinkTicking = false
+		m.composerBlinkSeq++
 		if m.notifier != nil {
 			m.notifier.SetFocused(false)
 		}
@@ -5372,6 +5426,9 @@ func (m model) runAgentWithOptions(runID int, runCtx context.Context, prompt str
 		sessionEvents := []pendingSessionEvent{}
 		usageModelID := m.modelName
 		var specReview *pendingSpecReviewPrompt
+		if m.awaitToolReadiness != nil {
+			m.awaitToolReadiness(runCtx)
+		}
 		options := m.agentOptions
 		options.Registry = cloneToolRegistry(m.registry)
 		goalAwareRun := !runOptions.specDraft && m.activeLoopID == "" &&
@@ -5414,6 +5471,9 @@ func (m model) runAgentWithOptions(runID int, runCtx context.Context, prompt str
 		options.ResponseStyle = m.responseStyle
 		options.Cwd = m.cwd
 		options.Images = images
+		if m.newTurnSessionProvider != nil && m.provider != nil {
+			options.TurnSessionProvider = m.newTurnSessionProvider(m.providerProfile, m.provider)
+		}
 		if m.captureRunImages != nil {
 			m.captureRunImages(images)
 		}

internal/tui/model_test.go

@@ -109,6 +109,81 @@ func TestPromptSubmitInjectsLiveSessionModelContext(t *testing.T) {
 	}
 }
 
+func TestPromptWaitsForToolReadinessBeforeSnapshot(t *testing.T) {
+	t.Setenv("XDG_DATA_HOME", t.TempDir())
+	registry := tools.NewRegistry()
+	provider := &fakeProvider{events: []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}
+	awaited := false
+	m := newModel(context.Background(), Options{
+		Cwd:          t.TempDir(),
+		ProviderName: "test",
+		ModelName:    "test-model",
+		Provider:     provider,
+		Registry:     registry,
+		AwaitToolReadiness: func(context.Context) {
+			awaited = true
+			registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
+		},
+	})
+	m.input.SetValue("inspect the workspace")
+
+	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	next := updated.(model)
+	if cmd == nil {
+		t.Fatal("expected prompt submit to start an agent run")
+	}
+	next.Update(execCmd(cmd))
+
+	if !awaited {
+		t.Fatal("agent run did not await tool readiness")
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("provider requests = %d, want 1", len(provider.requests))
+	}
+	for _, definition := range provider.requests[0].Tools {
+		if definition.Name == "read_file" {
+			return
+		}
+	}
+	t.Fatal("provider request omitted tool published by readiness callback")
+}
+
+func TestPromptBuildsTurnSessionForCurrentProvider(t *testing.T) {
+	t.Setenv("XDG_DATA_HOME", t.TempDir())
+	provider := &fakeProvider{events: []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}
+	profile := config.ProviderProfile{Name: "chatgpt", Model: "gpt-test"}
+	called := false
+	m := newModel(context.Background(), Options{
+		Cwd:             t.TempDir(),
+		ProviderName:    profile.Name,
+		ModelName:       profile.Model,
+		ProviderProfile: profile,
+		Provider:        provider,
+		Registry:        tools.NewRegistry(),
+		NewTurnSessionProvider: func(gotProfile config.ProviderProfile, gotProvider zeroruntime.Provider) zeroruntime.TurnSessionProvider {
+			called = true
+			if gotProfile.Name != profile.Name || gotProfile.Model != profile.Model || gotProvider != provider {
+				t.Fatalf("turn session factory inputs = %#v/%#v", gotProfile, gotProvider)
+			}
+			return zeroruntime.NewProviderTurnSessionProvider(gotProvider, zeroruntime.ProviderCapabilities{})
+		},
+	})
+	m.input.SetValue("inspect the workspace")
+
+	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	next := updated.(model)
+	next.Update(execCmd(cmd))
+	if !called {
+		t.Fatal("agent run did not build a turn session for the current provider")
+	}
+}
+
 func TestPromptSubmitStoresReasoningSeparatelyFromAnswer(t *testing.T) {
 	t.Setenv("XDG_DATA_HOME", t.TempDir())
 	provider := &fakeProvider{events: []zeroruntime.StreamEvent{
@@ -548,10 +623,10 @@ func TestModelCommandSwitchesSessionModel(t *testing.T) {
 	})
 	m.input.SetValue("/model gpt-4.1-mini")
 
-	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	updated, _ := m.Update(testKey(tea.KeyEnter))
 	next := updated.(model)
 
-	if cmd != nil {
+	if next.pending {
 		t.Fatal("expected /model to be handled without starting an agent run")
 	}
 	if next.modelName != "gpt-4.1-mini" || next.provider != nextProvider {
@@ -591,10 +666,10 @@ func TestModelCommandAcceptsChatGPTCatalogModelID(t *testing.T) {
 	}
 	m.input.SetValue("/model openai/gpt-5.6-sol")
 
-	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	updated, _ := m.Update(testKey(tea.KeyEnter))
 	next := updated.(model)
 
-	if cmd != nil {
+	if next.pending {
 		t.Fatal("expected /model to be handled without starting an agent run")
 	}
 	if next.modelName != "gpt-5.6-sol" || next.provider != nextProvider {
@@ -663,10 +738,10 @@ func TestModelCommandPersistsSelectedModelToUserConfig(t *testing.T) {
 	})
 	m.input.SetValue("/model gpt-4.1-mini")
 
-	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	updated, _ := m.Update(testKey(tea.KeyEnter))
 	next := updated.(model)
 
-	if cmd != nil {
+	if next.pending {
 		t.Fatal("expected /model to be handled without starting an agent run")
 	}
 	if next.modelName != "gpt-4.1-mini" {
@@ -2439,7 +2514,7 @@ func TestAssistantNarrationBeforeToolCardGetsBlankSeparator(t *testing.T) {
 	m.transcript = append(m.transcript,
 		transcriptRow{kind: rowUser, text: "run it"},
 		transcriptRow{kind: rowAssistant, text: "I'll inspect the existing file, then run it."},
-		transcriptRow{kind: rowToolResult, id: "t1", tool: "read_file", status: tools.StatusOK, detail: "File: time_test.py\n\n  1 | print('x')"},
+		transcriptRow{kind: rowToolResult, id: "t1", tool: "read_file", status: tools.StatusOK, detail: "File: time_test.py\n\n1→print('x')"},
 	)
 	items := m.transcriptBodyItems(m.chatColumnWidth(), "", false)
 	toolIdx := -1
@@ -2868,6 +2943,8 @@ func TestComposerBlinkStaysSolidWhileTyping(t *testing.T) {
 		terminalFocused:       true,
 		lastCharTime:          base,
 		composerCursorVisible: true,
+		composerBlinkSeq:      1,
+		composerBlinkTicking:  true,
 	}
 
 	// Each iteration simulates a keystroke (refreshing lastCharTime) followed by
@@ -2876,7 +2953,7 @@ func TestComposerBlinkStaysSolidWhileTyping(t *testing.T) {
 	for i := 0; i < 3; i++ {
 		now = now.Add(200 * time.Millisecond)
 		m.lastCharTime = now
-		updated, _ := m.Update(composerBlinkMsg{})
+		updated, _ := m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
 		m = updated.(model)
 		if !m.composerCursorVisible {
 			t.Fatalf("tick %d: expected cursor to stay visible while typing, got hidden", i)
@@ -2884,6 +2961,13 @@ func TestComposerBlinkStaysSolidWhileTyping(t *testing.T) {
 	}
 }
 
+func TestMainComposerDisablesEmbeddedCursorBlink(t *testing.T) {
+	m := newModel(t.Context(), Options{})
+	if m.input.Styles().Cursor.Blink {
+		t.Fatal("embedded composer cursor blink must stay disabled")
+	}
+}
+
 func TestComposerBlinkHiddenWhileUnfocused(t *testing.T) {
 	base := time.Date(2026, 7, 7, 12, 0, 0, 0, time.UTC)
 	m := model{
@@ -2920,10 +3004,10 @@ func TestComposerBlinkResumesAfterRefocusAndIdle(t *testing.T) {
 		t.Fatal("expected terminalFocused to be true after FocusMsg")
 	}
 
-	updated, _ = m.Update(composerBlinkMsg{})
+	updated, _ = m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
 	m = updated.(model)
 	first := m.composerCursorVisible
-	updated, _ = m.Update(composerBlinkMsg{})
+	updated, _ = m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
 	m = updated.(model)
 	second := m.composerCursorVisible
 	if first == second {
@@ -2931,6 +3015,34 @@ func TestComposerBlinkResumesAfterRefocusAndIdle(t *testing.T) {
 	}
 }
 
+func TestComposerBlinkRejectsTickFromBeforeRefocus(t *testing.T) {
+	base := time.Date(2026, 7, 7, 12, 0, 0, 0, time.UTC)
+	m := model{
+		now:                   func() time.Time { return base },
+		terminalFocused:       true,
+		lastCharTime:          base.Add(-time.Hour),
+		composerCursorVisible: true,
+		composerBlinkSeq:      5,
+		composerBlinkTicking:  true,
+	}
+	staleSeq := m.composerBlinkSeq
+
+	updated, _ := m.Update(tea.BlurMsg{})
+	m = updated.(model)
+	updated, refocusCmd := m.Update(tea.FocusMsg{})
+	m = updated.(model)
+	if refocusCmd == nil || m.composerBlinkSeq == staleSeq {
+		t.Fatal("refocus did not re-arm composer blinking with a new sequence")
+	}
+	visible := m.composerCursorVisible
+
+	updated, staleCmd := m.Update(composerBlinkMsg{seq: staleSeq})
+	m = updated.(model)
+	if staleCmd != nil || m.composerCursorVisible != visible {
+		t.Fatalf("stale tick changed composer state: cmd=%v visible=%v want=%v", staleCmd != nil, m.composerCursorVisible, visible)
+	}
+}
+
 func TestComposerBlinkTogglesWhenIdleAndFocused(t *testing.T) {
 	base := time.Date(2026, 7, 7, 12, 0, 0, 0, time.UTC)
 	m := model{
@@ -2938,20 +3050,53 @@ func TestComposerBlinkTogglesWhenIdleAndFocused(t *testing.T) {
 		terminalFocused:       true,
 		lastCharTime:          base.Add(-time.Hour),
 		composerCursorVisible: true,
+		composerBlinkSeq:      1,
+		composerBlinkTicking:  true,
 	}
 
-	updated, _ := m.Update(composerBlinkMsg{})
+	updated, _ := m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
 	m = updated.(model)
 	if m.composerCursorVisible {
 		t.Fatal("expected cursor to toggle off on first idle+focused tick")
 	}
-	updated, _ = m.Update(composerBlinkMsg{})
+	updated, _ = m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
 	m = updated.(model)
 	if !m.composerCursorVisible {
 		t.Fatal("expected cursor to toggle back on on second idle+focused tick")
 	}
 }
 
+func TestComposerBlinkSettlesVisibleWithoutRescheduling(t *testing.T) {
+	m := model{
+		now:                   func() time.Time { return time.Unix(100, 0) },
+		terminalFocused:       true,
+		lastCharTime:          time.Unix(0, 0),
+		composerCursorVisible: true,
+		composerBlinkSeq:      1,
+		composerBlinkTicking:  true,
+	}
+	var cmd tea.Cmd
+	for range composerBlinkIdleTicksBeforeSettle {
+		var updated tea.Model
+		updated, cmd = m.Update(composerBlinkMsg{seq: m.composerBlinkSeq})
+		m = updated.(model)
+	}
+	if m.composerBlinkTicking || cmd != nil || !m.composerCursorVisible {
+		t.Fatalf("settled cursor = ticking:%v cmd:%v visible:%v", m.composerBlinkTicking, cmd != nil, m.composerCursorVisible)
+	}
+}
+
+func TestComposerInputRestartsSettledBlink(t *testing.T) {
+	m := newModel(t.Context(), Options{})
+	m.composerBlinkTicking = false
+	m.composerCursorVisible = true
+	updated, cmd := m.Update(tea.KeyPressMsg(tea.Key{Code: 'a', Text: "a"}))
+	m = updated.(model)
+	if !m.composerBlinkTicking || cmd == nil || !m.composerCursorVisible {
+		t.Fatalf("restarted cursor = ticking:%v cmd:%v visible:%v", m.composerBlinkTicking, cmd != nil, m.composerCursorVisible)
+	}
+}
+
 func TestComposerCursorShowsImmediatelyOnKeypress(t *testing.T) {
 	// The blink phase may have just hidden the caret when the user starts
 	// typing; the typed character must render with a caret immediately, not