chore(main): release 0.2.0 (#370)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

.github/workflows/ci.yml

@@ -100,9 +100,10 @@ jobs:
           go-version-file: go.mod
           cache: true
 
-      # govulncheck and deadcode resolve a toolchain from their own modules, which
-      # can drop below the version go.mod requires and then fail to load our
-      # packages. Pin GOTOOLCHAIN to the go.mod toolchain so both run under it.
+      # govulncheck, deadcode, and golangci-lint each resolve a toolchain from
+      # their own modules, which can drop below the version go.mod requires and
+      # then fail to load our packages. Pin GOTOOLCHAIN to the go.mod toolchain
+      # so all three run under it.
       - name: Pin toolchain from go.mod
         run: |
           toolchain="$(awk '/^toolchain /{print $2}' go.mod)"
@@ -120,3 +121,13 @@ jobs:
       - name: deadcode (advisory)
         continue-on-error: true
         run: go run golang.org/x/tools/cmd/[email protected] -test=false ./...
+
+      # Advisory: catches what deadcode's whole-program reachability analysis
+      # doesn't, unused private functions/assignments reachable within a
+      # package but never actually called, plus staticcheck-style correctness
+      # and readability issues. Scoped to a few linters rather than the full
+      # default battery, and non-blocking, while the existing findings across
+      # the repo are cleaned up incrementally (see #527).
+      - name: golangci-lint (advisory)
+        continue-on-error: true
+        run: go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...

.gitignore

@@ -37,8 +37,13 @@ report.[0-9]_.[0-9]_.[0-9]_.[0-9]_.json
 
 /zero
 /zero.exe
+/zero-skills
 /snake-game.html
 
 # Superpowers skill artifacts (brainstorm mockups, specs, plans) — local only
 .superpowers/
 docs/superpowers/
+
+# Termux/Android build artifacts
+zero-termux-arm64
+zero-linux-sandbox

.release-please-manifest.json

@@ -1 +1 @@
-{".":"0.1.0"}
+{".":"0.2.0"}

AGENTS.md

@@ -47,6 +47,12 @@ Tips:
 - In a monorepo, drop a narrower `AGENTS.md` in each sub-tree (e.g. `services/api/AGENTS.md`). Zero picks those up automatically when you launch from inside the sub-tree.
 - A YAML frontmatter block (`---\n...\n---`) at the top is preserved verbatim in the injected prompt but is not parsed for `globs:` or `alwaysApply:` scoping today — keep the body self-contained.
 
+### Personal guidelines, across every project
+
+For preferences that follow *you*, not a specific repo (tone, tooling habits, workflow), drop a `ZERO.md` in your user config directory: `~/.config/zero/ZERO.md` on Linux/macOS, `%AppData%\Roaming\zero\ZERO.md` on Windows — the same directory as `config.json` and your personal specialists. Same format and 8 KiB cap as the project files above, and the same case-insensitive basename match.
+
+This file is injected as its own `## User guidelines` section, before the project's `AGENTS.md`/`ZERO.md`, and is labeled as personal preference in the prompt: project guidelines are the later, more specific instruction and take precedence over it when the two conflict.
+
 ## 2. Custom specialists
 
 Specialists are Zero's sub-agents. Three scopes, in priority order:
@@ -307,6 +313,7 @@ A team that wants every contributor's Zero to behave the same way commits:
 Each contributor adds only:
 
 - `~/.config/zero/config.json` — their personal API keys, theme, default mode.
+- `~/.config/zero/ZERO.md` — personal preferences that follow them across every project (see section 1).
 - `~/.local/share/zero/skills/` — personal skills they keep across projects.
 
 That's it. Run `zero` from the repo root and the agent has the team's full instruction set, every contributor's personal setup, and nothing else.

CHANGELOG.md

@@ -6,6 +6,71 @@ 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`.
 
+## [0.2.0](https://github.com/Gitlawb/zero/compare/v0.1.0...v0.2.0) (2026-07-06)
+
+
+### Features
+
+* add --auto flag for LLM-generated commit messages ([#423](https://github.com/Gitlawb/zero/issues/423)) ([b0abde7](https://github.com/Gitlawb/zero/commit/b0abde7d0697e808480cd59d69a6f4d0c6320475))
+* add zero changes push and pr subcommands, and extra repo-info metrics ([#391](https://github.com/Gitlawb/zero/issues/391)) ([2312abe](https://github.com/Gitlawb/zero/commit/2312abe5ddd95f4c6ef373cfb61cc03092f48cdd))
+* agent quality, caching, retry, and tooling upgrades ([#506](https://github.com/Gitlawb/zero/issues/506)) ([3c81fea](https://github.com/Gitlawb/zero/commit/3c81fea22873ee3df7fc97b10cb4f77792706c4b))
+* **agent:** curb over-engineering the solution in the editing discipline ([#517](https://github.com/Gitlawb/zero/issues/517)) ([f4c998a](https://github.com/Gitlawb/zero/commit/f4c998ac30c4f07ff313a2d706791e857293be49)), closes [#516](https://github.com/Gitlawb/zero/issues/516)
+* **agent:** inject per-user config.UserConfigDir()/zero/ZERO.md guidelines into system prompt ([#475](https://github.com/Gitlawb/zero/issues/475)) ([7b10aab](https://github.com/Gitlawb/zero/commit/7b10aab74bf14a01166b2cea22deab79bba9850b))
+* **openai:** forward prompt_cache_key for server-side prefix cache routing ([#515](https://github.com/Gitlawb/zero/issues/515)) ([87e7e69](https://github.com/Gitlawb/zero/commit/87e7e69afd18b5539579856f3a61c6a95bc445ae))
+* **providers:** add `zero providers models` to discover a provider's models ([#386](https://github.com/Gitlawb/zero/issues/386)) ([0bc8074](https://github.com/Gitlawb/zero/commit/0bc8074c97b0310e4a9d70c3f967003ee5e8a59f))
+* **providers:** add KiloCode and OpenCode provider support ([#388](https://github.com/Gitlawb/zero/issues/388)) ([b1ccb6d](https://github.com/Gitlawb/zero/commit/b1ccb6d9c1875377f5e5ea81a1304edd1e41ab4f))
+* **providers:** add Meituan LongCat catalog preset ([#424](https://github.com/Gitlawb/zero/issues/424)) ([b4275e3](https://github.com/Gitlawb/zero/commit/b4275e350472b2490212bf814709819d354c1216))
+* **providers:** split minimax zai into intl cn ([#398](https://github.com/Gitlawb/zero/issues/398)) ([aaad4d2](https://github.com/Gitlawb/zero/commit/aaad4d271270f41af837b6f3b60ae80beba0c645))
+* require manual approval before npm publish + drop release-as pin ([#369](https://github.com/Gitlawb/zero/issues/369)) ([bd89a1f](https://github.com/Gitlawb/zero/commit/bd89a1f451643c1b65ec803070abc7b116631ebe))
+* **sandbox:** unelevated Windows fallback tier instead of prompts-only degrade ([#427](https://github.com/Gitlawb/zero/issues/427)) ([b9ddd6f](https://github.com/Gitlawb/zero/commit/b9ddd6f42138312a1fee8d8bb67c46c8eb1dea2f))
+* support shift enter for composer newlines ([#462](https://github.com/Gitlawb/zero/issues/462)) ([daf65e0](https://github.com/Gitlawb/zero/commit/daf65e0af9a040314d4ab337b0ad59c55416b7bc))
+* **tui:** /loop — repeat a prompt or command on an interval or self-paced ([#502](https://github.com/Gitlawb/zero/issues/502)) ([387fe67](https://github.com/Gitlawb/zero/commit/387fe67ee7cd81317c9c969f5906a4437080fea3))
+* **tui:** add search/filter to provider picker in setup wizard ([#400](https://github.com/Gitlawb/zero/issues/400)) ([2fcea71](https://github.com/Gitlawb/zero/commit/2fcea71778d23e050c93409c471aef45b68c1621))
+* **update:** add zero upgrade command to apply self-updates ([#461](https://github.com/Gitlawb/zero/issues/461)) ([5f36349](https://github.com/Gitlawb/zero/commit/5f36349c1884e81fa9bc66bb5fe813b627e897b7))
+
+
+### Bug Fixes
+
+* **action:** keep provider key scoped to zero step ([#448](https://github.com/Gitlawb/zero/issues/448)) ([407a927](https://github.com/Gitlawb/zero/commit/407a92739ff508cba32d2c12b3f36f0efcdd54c3))
+* add android platform support for Termux npm install ([#455](https://github.com/Gitlawb/zero/issues/455)) ([9bd93c6](https://github.com/Gitlawb/zero/commit/9bd93c62f8d57fb74057284aa66a1b6e1429dcdd)), closes [#449](https://github.com/Gitlawb/zero/issues/449)
+* **agent:** reject a malformed additional_permissions payload before prompting ([#453](https://github.com/Gitlawb/zero/issues/453)) ([e4f760e](https://github.com/Gitlawb/zero/commit/e4f760ee8bd57299cd2fcb37e8e23130037c2607))
+* allow non-TLS connections to private-network provider endpoints ([#444](https://github.com/Gitlawb/zero/issues/444)) ([1d86384](https://github.com/Gitlawb/zero/commit/1d8638466ca31517eb9db2b9353d3dce1cbeeabc))
+* **auth:** route zero auth login chatgpt to the dedicated ChatGPT flow ([#443](https://github.com/Gitlawb/zero/issues/443)) ([305a62c](https://github.com/Gitlawb/zero/commit/305a62c954ca6cec00bc58d5398f933415156aff))
+* **config:** fall back to a usable saved provider instead of forcing full re-onboarding ([#410](https://github.com/Gitlawb/zero/issues/410)) ([c60ad87](https://github.com/Gitlawb/zero/commit/c60ad8729f79bb841114d352ee2d2fe29d5d0e41))
+* **config:** let a gateway ANTHROPIC_BASE_URL resolve as anthropic-compatible ([#497](https://github.com/Gitlawb/zero/issues/497)) ([30dd7c3](https://github.com/Gitlawb/zero/commit/30dd7c3112ad22d42fa12b5addd4e38f4beda42a)), closes [#479](https://github.com/Gitlawb/zero/issues/479)
+* **config:** unbrick first-run setup — default google/anthropic models, enter setup on fixable config errors ([#385](https://github.com/Gitlawb/zero/issues/385)) ([72eed06](https://github.com/Gitlawb/zero/commit/72eed06b4f94c43d75d31fe54a58d2f566de059e))
+* **config:** use ~/.config on macOS and enter setup when no provider ([#371](https://github.com/Gitlawb/zero/issues/371)) ([#372](https://github.com/Gitlawb/zero/issues/372)) ([027a8f2](https://github.com/Gitlawb/zero/commit/027a8f2768b17b89f5c8270887f156e2ccda69ea))
+* **docs:** rename AGENTS.MD > AGENTS.md ([#438](https://github.com/Gitlawb/zero/issues/438)) ([4266baf](https://github.com/Gitlawb/zero/commit/4266baf222df583ed2078b776687f12d496475b5))
+* **gemini:** strip unsupported JSON Schema fields from tool declarations ([#374](https://github.com/Gitlawb/zero/issues/374)) ([39e7100](https://github.com/Gitlawb/zero/commit/39e7100674150144a1152e3110c64c7cf0321d64)), closes [#373](https://github.com/Gitlawb/zero/issues/373)
+* **install:** persist install dir to user PATH on Windows ([#407](https://github.com/Gitlawb/zero/issues/407)) ([bdb1b0e](https://github.com/Gitlawb/zero/commit/bdb1b0ecd15859b1712a6037d296dace7f9c3c3f))
+* **mcp:** block cross-origin credential redirects ([#396](https://github.com/Gitlawb/zero/issues/396)) ([f915f70](https://github.com/Gitlawb/zero/commit/f915f70e5a3096e2419fa8d961a0f84a626fa4a9))
+* **oauth:** treat Windows ERROR_ACCESS_DENIED as lock contention in createSecretFile ([#445](https://github.com/Gitlawb/zero/issues/445)) ([d05e914](https://github.com/Gitlawb/zero/commit/d05e9148a7f79f67d1d3c31fca2775f21fbd331e))
+* **openai:** handle Ollama reasoning stream deltas ([#486](https://github.com/Gitlawb/zero/issues/486)) ([f6c0606](https://github.com/Gitlawb/zero/commit/f6c060631e18e082dda24cc4dc0903c31c2120d6))
+* preserve conversation context in exec prompts ([#460](https://github.com/Gitlawb/zero/issues/460)) ([949ee43](https://github.com/Gitlawb/zero/commit/949ee43f71e5cb7fab4695c5cb7b442fe4ecfbf7))
+* **provider-wizard:** allow multiple custom OpenAI-compatible providers ([#403](https://github.com/Gitlawb/zero/issues/403)) ([3fbbd28](https://github.com/Gitlawb/zero/commit/3fbbd28e4c586822cc4312c86232d94befe56e87))
+* **sandbox:** fix nested pipe creation under the Windows restricted token ([#456](https://github.com/Gitlawb/zero/issues/456)) ([563a6db](https://github.com/Gitlawb/zero/commit/563a6dbe91e65d5daeefd7626e8a77e30a6d8fb2))
+* **sandbox:** gate /tmp test assertions on GOOS, not path existence ([#426](https://github.com/Gitlawb/zero/issues/426)) ([f653dca](https://github.com/Gitlawb/zero/commit/f653dcac363fb69ad7be5b35e6e0fa6d2bce476d))
+* **sandbox:** self-heal a corrupt unelevated setup marker ([#437](https://github.com/Gitlawb/zero/issues/437)) ([8d0c5fe](https://github.com/Gitlawb/zero/commit/8d0c5feccb8bdbfb015df0508aa6e3bcbd1fd0e8))
+* **specialist:** cap max specialist nesting depth ([#491](https://github.com/Gitlawb/zero/issues/491)) ([177442c](https://github.com/Gitlawb/zero/commit/177442cfe4015bd8df04cc9894f98b468ee796d4))
+* Termux/Android support — PRoot scroll, SIGSYS sandbox, build docs ([#509](https://github.com/Gitlawb/zero/issues/509)) ([0f69d99](https://github.com/Gitlawb/zero/commit/0f69d995e9b586b774f66c066b21abab5e03024a))
+* **tools:** Block MSYS coreutils under Windows sandbox ([#476](https://github.com/Gitlawb/zero/issues/476)) ([81aad58](https://github.com/Gitlawb/zero/commit/81aad58d97839e51c068e2f08907618991fdc3fb))
+* **tools:** CRLF line ending mismatch in edit_file tool on Windows ([#378](https://github.com/Gitlawb/zero/issues/378)) ([33dc7ae](https://github.com/Gitlawb/zero/commit/33dc7ae2cc82c5389675531e1416856dae7151ce))
+* **tools:** fix cmd.exe /S/C corrupting commands with embedded quotes ([#465](https://github.com/Gitlawb/zero/issues/465)) ([190241b](https://github.com/Gitlawb/zero/commit/190241bd593f43211b766e0b13c8e89802d4bb37))
+* **tools:** flag piped POSIX utilities before running on Windows ([#412](https://github.com/Gitlawb/zero/issues/412)) ([5658a36](https://github.com/Gitlawb/zero/commit/5658a366274fc59a9d5336b06a21019c9c25cbf1))
+* **tools:** make grep and glob respect run cancellation ([#464](https://github.com/Gitlawb/zero/issues/464)) ([ba6c026](https://github.com/Gitlawb/zero/commit/ba6c0264697b7d7ed479f6e782fba9700a481e3d))
+* **tools:** require permission before web_search requests ([#382](https://github.com/Gitlawb/zero/issues/382)) ([960db96](https://github.com/Gitlawb/zero/commit/960db9660e4e31dc588fe8f7d6f116ff5e225566))
+* **tui:** compose help overlay through the viewport overlay pipeline ([#421](https://github.com/Gitlawb/zero/issues/421)) ([5b2b4de](https://github.com/Gitlawb/zero/commit/5b2b4dea1aaf9e0f68baa25e97e83296fb17b1a2))
+* **tui:** keep the profile name on /model switch so the stored key resolves ([#441](https://github.com/Gitlawb/zero/issues/441)) ([9134148](https://github.com/Gitlawb/zero/commit/9134148f4df3e4e556fba6c2f8babfdf6fcfeee1)), closes [#440](https://github.com/Gitlawb/zero/issues/440)
+* **tui:** resolve every permission request so the agent can't deadlock ([#397](https://github.com/Gitlawb/zero/issues/397)) ([952788f](https://github.com/Gitlawb/zero/commit/952788f72d32957659fe004521fcc8372b9ba9b4))
+* **tui:** show an M suffix for million-scale token counts ([#457](https://github.com/Gitlawb/zero/issues/457)) ([0562e3b](https://github.com/Gitlawb/zero/commit/0562e3bef7df2328610a48a1e81632a8da4aec64))
+* **tui:** title /model rows by model name, not the catalog description ([#395](https://github.com/Gitlawb/zero/issues/395)) ([cdf9d83](https://github.com/Gitlawb/zero/commit/cdf9d839ae57a729f292f36f7c5b0c67b41b288d))
+
+
+### Performance Improvements
+
+* cache TUI model registry ([#496](https://github.com/Gitlawb/zero/issues/496)) ([e7d88b4](https://github.com/Gitlawb/zero/commit/e7d88b4b518049733da25a8447c00144bd1da518))
+* universal tool-output ceiling with spill + async post-edit diagnostics ([#518](https://github.com/Gitlawb/zero/issues/518)) ([95ccd5b](https://github.com/Gitlawb/zero/commit/95ccd5bc327f6fb464ff0239f7229de789f578dc))
+
 ## 0.1.0 (2026-07-02)
 
 

README.md

@@ -186,6 +186,7 @@ Common slash commands:
 | `/spec`, `/plan` | draft and review a plan before building |
 | `/image` | attach an image for vision-capable models |
 | `/resume`, `/rewind` | continue or roll back local sessions |
+| `/loop` | repeat a prompt or custom `/command` on an interval (`/loop 5m /babysit-prs`) or self-paced |
 | `/compact`, `/context` | manage context usage |
 | `/permissions`, `/tools` | inspect available tools and policy |
 | `/add-dir` | allow an extra write directory for this session |
@@ -283,6 +284,39 @@ zero cron             scheduled agent jobs
 zero update           check for newer releases
 ```
 
+## Extending Zero
+
+### Project and personal instructions
+
+Zero appends project-specific guidance to the system prompt from the first
+`AGENTS.md`, `ZERO.md`, or `.zero/AGENTS.md` file found in each directory from
+the git root down to your current working directory (checked in that order
+per directory). Files are injected general-to-specific, capped at 8 KiB per
+file and 32 KiB total.
+
+A personal `ZERO.md` under `config.UserConfigDir()/zero/ZERO.md`
+(`$XDG_CONFIG_HOME/zero/ZERO.md` or `~/.config/zero/ZERO.md` on Linux/macOS,
+`%AppData%\Roaming\zero\ZERO.md` on Windows) applies across every workspace, ahead of any project guidelines.
+
+### Plugins
+
+Plugins are discovered from `~/.config/zero/plugins/<name>/plugin.json` (user
+scope — `$XDG_CONFIG_HOME` or `~/.config` on every OS, independent of the
+`config.UserConfigDir()` path used above) and `<cwd>/.zero/plugins/<name>/plugin.json`
+(project scope — resolved from the current working directory, not the repo
+root), and managed with `zero plugins`. A manifest can declare:
+
+- `tools` — custom tools (`command`, `args`, `inputSchema`, and a
+  `permission` of `prompt` or `deny`; `allow` is honored only when manifest tool
+  auto-approval is enabled)
+- `hooks` — commands run on `beforeTool`, `afterTool`, `sessionStart`, or
+  `sessionEnd`
+- `prompts` and `skills` — additional prompt/skill files
+
+MCP servers (`zero mcp`) and standalone markdown skills (`zero skills`) use
+the same extension points and can also be wired up outside of a plugin
+manifest.
+
 ## Appearance And Accessibility
 
 | Control | Effect |

action.yml

@@ -180,6 +180,7 @@ runs:
         ZERO_INPUT_ADD_DIR: ${{ inputs.add-dir }}
         ZERO_INPUT_WORKTREE: ${{ inputs.worktree }}
         ZERO_INPUT_OUTPUT_FORMAT: ${{ inputs.output-format }}
+        ACTION_PATH: ${{ github.action_path }}
         RUNNER_TEMP: ${{ runner.temp }}
       run: |
         set -euo pipefail
@@ -247,10 +248,9 @@ runs:
         echo "exit-code=${code}" >> "${GITHUB_OUTPUT}"
         echo "output-file=${output_file}" >> "${GITHUB_OUTPUT}"
 
-        # Best-effort one-line summary: last non-empty line of the captured output.
-        summary="$(grep -v '^[[:space:]]*$' "${output_file}" | tail -n 1 || true)"
-        # Keep the summary single-line and bounded for the step output.
-        summary="$(printf '%s' "${summary}" | cut -c1-280)"
+        # Best-effort one-line summary. Structured output ends with lifecycle
+        # metadata, so parse final/error events instead of tailing stdout.
+        summary="$(node "${ACTION_PATH}/scripts/action-summary.mjs" "${ZERO_INPUT_OUTPUT_FORMAT}" "${output_file}" || true)"
         {
           echo "summary<<ZERO_EOF"
           printf '%s\n' "${summary}"

docs/INSTALL.md

@@ -173,6 +173,61 @@ Linux native sandboxing also requires Bubblewrap to be installed.
 
 macOS uses the system sandbox and does not need an extra helper binary.
 
+### Termux (Android)
+
+Zero can run natively on Android via [Termux](https://termux.dev/). Build with
+`GOOS=android` to avoid the `faccessat2` syscall that is blocked by Samsung's
+seccomp filter on Android:
+
+```bash
+# Install Go in Termux
+pkg install golang
+
+# Build Zero for Android
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+CGO_ENABLED=0 GOOS=android GOARCH=arm64 go build -ldflags="-s -w" -o zero ./cmd/zero
+
+# Move into PATH
+mv zero ~/.local/bin/
+```
+
+> **Why `GOOS=android`?** Go 1.26+ detects `runtime.GOOS == "android"` and skips
+> the `faccessat2` syscall inside `os/exec.findExecutable`, falling back to
+> permission-bit checks. Without this flag, Android's seccomp sends SIGSYS and
+> kills the process whenever Zero looks up a binary on `PATH` (git, sh, etc.).
+
+**DNS.** Android does not expose `/etc/resolv.conf`. Go's pure-Go DNS resolver
+needs one. Use `proot` to bind-mount Termux's resolver config:
+
+```bash
+pkg install proot
+proot -b "$PREFIX/etc/resolv.conf:/etc/resolv.conf" zero
+```
+
+Create a wrapper at `~/.local/bin/zero` to avoid typing proot every time:
+
+```bash
+#!/data/data/com.termux/files/usr/bin/bash
+exec proot -b "$PREFIX/etc/resolv.conf:/etc/resolv.conf" ~/.local/bin/zero.bin "$@"
+```
+
+**Scroll.** On native Termux (not under PRoot), mouse scrolling works out of the
+box. The TUI uses Bubble Tea's `AllMotion` mouse mode by default. If you run Zero
+inside PRoot (e.g. through proot-distro), the scroll fix activates `CellMotion`
+to avoid PRoot's ptrace interference with the 1003 escape sequence.
+
+**Providers.** Zero works with any OpenAI-compatible provider on Termux. For
+example, to use OpenCode Zen's free tier:
+
+```bash
+zero providers add opencode \
+  --name opencode \
+  --model deepseek-v4-flash-free \
+  --base-url https://opencode.ai/zen/v1 \
+  --set-active
+```
+
 Windows source builds can use the main `zero.exe` as the command runner and setup
 helper through Zero's built-in self-dispatch path. If you want a release-style
 layout anyway, build the standalone helper executables next to `zero.exe`:

docs/STREAM_JSON_PROTOCOL.md

@@ -41,6 +41,7 @@ events when present.
 
 ```json
 { "schemaVersion": 2, "type": "run_start", "runId": "run_20260603_abc123", "sessionId": "zero_20260603100000_abc123", "cwd": "/repo", "provider": "openai", "model": "gpt-4.1", "apiModel": "gpt-4.1" }
+{ "schemaVersion": 2, "type": "reasoning", "runId": "run_20260603_abc123", "delta": "Thinking..." }
 { "schemaVersion": 2, "type": "text", "runId": "run_20260603_abc123", "delta": "..." }
 { "schemaVersion": 2, "type": "tool_call", "runId": "run_20260603_abc123", "id": "call_1", "name": "read_file", "args": { "path": "README.md" }, "sideEffect": "read" }
 { "schemaVersion": 2, "type": "permission_request", "runId": "run_20260603_abc123", "id": "call_2", "name": "write_file", "action": "prompt", "permission": "prompt", "permissionMode": "ask", "sideEffect": "write", "reason": "Creates or overwrites files." }
@@ -51,6 +52,10 @@ events when present.
 { "schemaVersion": 2, "type": "run_end", "runId": "run_20260603_abc123", "status": "success", "exitCode": 0 }
 ```
 
+`reasoning` events carry live model reasoning/status deltas for providers that
+stream them separately from answer text. They are liveness/progress events only:
+they are not folded into `text` or the final answer.
+
 Permission events may include structured sandbox metadata:
 
 ```json

internal/agent/async_diagnostics.go

@@ -0,0 +1,185 @@
+package agent
+
+import (
+	"context"
+	"path/filepath"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Post-edit LSP diagnostics used to run synchronously inside edit_file /
+// write_file: every edit blocked its tool result on a fresh-publish wait with
+// a quiet-period debounce (≥300ms floor that reset on each publish, 10s cap) —
+// the largest recurring latency tax in an edit-heavy session. Collection is
+// now asynchronous: a file is enqueued the moment a mutating tool finishes, a
+// single worker checks it while the rest of the tool batch (and any
+// self-correct verification) runs, and the loop drains completed results just
+// before building the NEXT provider request, appending errors as a user-role
+// nudge. The model always issues another request after an edit turn — it has
+// to read its tool results — so it still sees introduced errors at the same
+// decision point as before, without any tool call stalling on the language
+// server.
+//
+// A file re-edited before its check runs is checked once in its final state
+// (the checker re-reads the file); a file re-edited after its check completed
+// is re-enqueued and the newer result wins. When the worker is still busy at
+// drain time, the drain gives up after a short wait and the results are
+// delivered on the following turn instead — they stay accurate because any
+// further edit re-enqueues the file.
+
+// asyncDiagnosticsDrainTimeout bounds how long a turn waits on the in-flight
+// check before deferring delivery to the next turn. Most checks finish well
+// under it (debounce floor + analysis); a cold language server on a large
+// repo does not get to stall the loop the way the old inline 10s cap could.
+// A var so tests can shorten the wait.
+var asyncDiagnosticsDrainTimeout = 3 * time.Second
+
+// asyncDiagnosticsFinalDrainTimeout is the wait used when the run is about to
+// FINALIZE (natural completion or the max-turns summary): there is no later
+// turn to defer to, so the gate waits out the old inline-era budget rather
+// than silently dropping an error the last edit introduced. A var so tests
+// can shorten the wait.
+var asyncDiagnosticsFinalDrainTimeout = 10 * time.Second
+
+// asyncDiagnosticsNudge prefixes the drained diagnostics blocks; phrased like
+// the old inline block so the model reacts the same way.
+const asyncDiagnosticsNudge = "Diagnostics after your recent edits (fix any errors you introduced):\n"
+
+type asyncDiagnostics struct {
+	check         func(context.Context, string) string
+	workspaceRoot string
+
+	mu      sync.Mutex
+	queue   []string
+	queued  map[string]bool
+	results map[string]string
+	// working is non-nil while the worker goroutine runs and is closed when it
+	// exits, so drain can wait for quiescence without polling.
+	working chan struct{}
+}
+
+// newAsyncDiagnostics returns nil when check is nil, and every method on a nil
+// receiver is a no-op — callers wire it unconditionally.
+func newAsyncDiagnostics(check func(context.Context, string) string, workspaceRoot string) *asyncDiagnostics {
+	if check == nil {
+		return nil
+	}
+	return &asyncDiagnostics{
+		check:         check,
+		workspaceRoot: workspaceRoot,
+		queued:        map[string]bool{},
+		results:       map[string]string{},
+	}
+}
+
+// enqueue schedules changed files for a background check. Paths are
+// workspace-relative from Result.ChangedFiles (absolute when under a granted
+// extra write root), mirroring SelfCorrector's resolution.
+func (diagnostics *asyncDiagnostics) enqueue(ctx context.Context, files []string) {
+	if diagnostics == nil || len(files) == 0 {
+		return
+	}
+	diagnostics.mu.Lock()
+	defer diagnostics.mu.Unlock()
+	for _, file := range files {
+		path := diagnostics.absPath(file)
+		if diagnostics.queued[path] {
+			continue
+		}
+		diagnostics.queued[path] = true
+		// A completed result for this file describes a pre-edit state now.
+		delete(diagnostics.results, path)
+		diagnostics.queue = append(diagnostics.queue, path)
+	}
+	if diagnostics.working == nil && len(diagnostics.queue) > 0 {
+		done := make(chan struct{})
+		diagnostics.working = done
+		go diagnostics.work(ctx, done)
+	}
+}
+
+func (diagnostics *asyncDiagnostics) work(ctx context.Context, done chan struct{}) {
+	defer close(done)
+	for {
+		diagnostics.mu.Lock()
+		if len(diagnostics.queue) == 0 || ctx.Err() != nil {
+			diagnostics.working = nil
+			diagnostics.mu.Unlock()
+			return
+		}
+		path := diagnostics.queue[0]
+		diagnostics.queue = diagnostics.queue[1:]
+		// Un-mark before checking so an edit that lands DURING the check
+		// re-enqueues the file and it is checked again in its final state.
+		delete(diagnostics.queued, path)
+		diagnostics.mu.Unlock()
+
+		block := diagnostics.check(ctx, path)
+
+		diagnostics.mu.Lock()
+		if block != "" {
+			diagnostics.results[path] = block
+		} else {
+			delete(diagnostics.results, path)
+		}
+		diagnostics.mu.Unlock()
+	}
+}
+
+// drain waits briefly for the worker to go quiet, then formats and consumes
+// all completed error blocks as one nudge. Returns "" when there is nothing
+// to report or the worker is still busy (results then surface next turn).
+func (diagnostics *asyncDiagnostics) drain(ctx context.Context) string {
+	return diagnostics.drainWithin(ctx, asyncDiagnosticsDrainTimeout)
+}
+
+// drainFinal is drain with the finalization budget: called before the run's
+// last model request, where an undelivered result would otherwise be lost.
+func (diagnostics *asyncDiagnostics) drainFinal(ctx context.Context) string {
+	return diagnostics.drainWithin(ctx, asyncDiagnosticsFinalDrainTimeout)
+}
+
+func (diagnostics *asyncDiagnostics) drainWithin(ctx context.Context, timeout time.Duration) string {
+	if diagnostics == nil {
+		return ""
+	}
+	diagnostics.mu.Lock()
+	busy := diagnostics.working
+	diagnostics.mu.Unlock()
+	if busy != nil {
+		timer := time.NewTimer(timeout)
+		defer timer.Stop()
+		select {
+		case <-busy:
+		case <-timer.C:
+			return ""
+		case <-ctx.Done():
+			return ""
+		}
+	}
+	diagnostics.mu.Lock()
+	defer diagnostics.mu.Unlock()
+	if len(diagnostics.results) == 0 {
+		return ""
+	}
+	paths := make([]string, 0, len(diagnostics.results))
+	for path := range diagnostics.results {
+		paths = append(paths, path)
+	}
+	sort.Strings(paths)
+	blocks := make([]string, 0, len(paths))
+	for _, path := range paths {
+		blocks = append(blocks, diagnostics.results[path])
+	}
+	diagnostics.results = map[string]string{}
+	return asyncDiagnosticsNudge + strings.Join(blocks, "\n")
+}
+
+func (diagnostics *asyncDiagnostics) absPath(path string) string {
+	if filepath.IsAbs(path) {
+		return path
+	}
+	return filepath.Join(diagnostics.workspaceRoot, path)
+}

internal/agent/async_diagnostics_test.go

@@ -0,0 +1,338 @@
+package agent
+
+import (
+	"context"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestAsyncDiagnosticsNilCollectorNoOps(t *testing.T) {
+	var diagnostics *asyncDiagnostics
+	if diagnostics != newAsyncDiagnostics(nil, "/ws") {
+		t.Fatal("nil check must yield a nil collector")
+	}
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+	if nudge := diagnostics.drain(context.Background()); nudge != "" {
+		t.Fatalf("nil collector drain = %q, want empty", nudge)
+	}
+}
+
+func TestAsyncDiagnosticsCollectsAndDrainsOnce(t *testing.T) {
+	// A real absolute root: a literal "/ws" is not filepath.IsAbs on Windows
+	// (no drive letter), which is exactly what the assertions below check.
+	root := t.TempDir()
+	var mu sync.Mutex
+	var checked []string
+	check := func(_ context.Context, absPath string) string {
+		mu.Lock()
+		checked = append(checked, absPath)
+		mu.Unlock()
+		if filepath.Base(absPath) == "clean.go" {
+			return ""
+		}
+		return "ERR " + filepath.Base(absPath)
+	}
+	diagnostics := newAsyncDiagnostics(check, root)
+	diagnostics.enqueue(context.Background(), []string{"broken.go", "clean.go"})
+
+	nudge := diagnostics.drain(context.Background())
+	if !strings.HasPrefix(nudge, asyncDiagnosticsNudge) {
+		t.Fatalf("nudge missing prefix: %q", nudge)
+	}
+	if !strings.Contains(nudge, "ERR broken.go") || strings.Contains(nudge, "clean.go") {
+		t.Fatalf("nudge = %q, want broken.go errors only", nudge)
+	}
+	mu.Lock()
+	for _, path := range checked {
+		if !filepath.IsAbs(path) || !strings.HasPrefix(path, root+string(filepath.Separator)) {
+			t.Fatalf("check received %q, want path under workspace root %q", path, root)
+		}
+	}
+	mu.Unlock()
+	if second := diagnostics.drain(context.Background()); second != "" {
+		t.Fatalf("second drain = %q, want empty (results are consumed)", second)
+	}
+}
+
+// A check still running at drain time defers delivery to the next drain
+// instead of blocking the turn.
+func TestAsyncDiagnosticsSlowCheckDefersToNextDrain(t *testing.T) {
+	previous := asyncDiagnosticsDrainTimeout
+	asyncDiagnosticsDrainTimeout = 20 * time.Millisecond
+	defer func() { asyncDiagnosticsDrainTimeout = previous }()
+
+	release := make(chan struct{})
+	check := func(_ context.Context, absPath string) string {
+		<-release
+		return "ERR " + filepath.Base(absPath)
+	}
+	diagnostics := newAsyncDiagnostics(check, "/ws")
+	diagnostics.enqueue(context.Background(), []string{"slow.go"})
+
+	start := time.Now()
+	if nudge := diagnostics.drain(context.Background()); nudge != "" {
+		t.Fatalf("drain during in-flight check = %q, want empty", nudge)
+	}
+	if elapsed := time.Since(start); elapsed > time.Second {
+		t.Fatalf("drain blocked %v, want bounded by the (shortened) timeout", elapsed)
+	}
+
+	close(release)
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		if nudge := diagnostics.drain(context.Background()); nudge != "" {
+			if !strings.Contains(nudge, "ERR slow.go") {
+				t.Fatalf("deferred nudge = %q", nudge)
+			}
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("deferred diagnostics never delivered")
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+}
+
+// Re-enqueueing a file after its check completed replaces the stale result.
+func TestAsyncDiagnosticsReEditReplacesResult(t *testing.T) {
+	var mu sync.Mutex
+	response := "ERR old"
+	check := func(context.Context, string) string {
+		mu.Lock()
+		defer mu.Unlock()
+		return response
+	}
+	diagnostics := newAsyncDiagnostics(check, "/ws")
+
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+	waitForIdle(t, diagnostics)
+	mu.Lock()
+	response = "ERR new"
+	mu.Unlock()
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+
+	nudge := diagnostics.drain(context.Background())
+	if !strings.Contains(nudge, "ERR new") || strings.Contains(nudge, "ERR old") {
+		t.Fatalf("nudge = %q, want only the re-check result", nudge)
+	}
+}
+
+func waitForIdle(t *testing.T, diagnostics *asyncDiagnostics) {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		diagnostics.mu.Lock()
+		busy := diagnostics.working
+		diagnostics.mu.Unlock()
+		if busy == nil {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("worker never went idle")
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+}
+
+// changedFilesTool is a fake mutating tool reporting a changed file without
+// touching the filesystem.
+type changedFilesTool struct{}
+
+func (changedFilesTool) Name() string        { return "fake_edit" }
+func (changedFilesTool) Description() string { return "test mutating tool" }
+func (changedFilesTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", Properties: map[string]tools.PropertySchema{}}
+}
+func (changedFilesTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectWrite, Permission: tools.PermissionAllow, Reason: "test"}
+}
+func (changedFilesTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "edited", ChangedFiles: []string{"main.go"}}
+}
+
+// End-to-end through Run: an edit turn does NOT carry diagnostics in the tool
+// result (the old inline path), but the next request contains the background
+// diagnostics nudge, before the model finalizes.
+func TestRunDeliversAsyncDiagnosticsNudgeNextTurn(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+	root := t.TempDir()
+	var checkedPath string
+	var mu sync.Mutex
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      root,
+		FileDiagnostics: func(_ context.Context, absPath string) string {
+			mu.Lock()
+			checkedPath = absPath
+			mu.Unlock()
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 provider turns, got %d", len(provider.requests))
+	}
+
+	// The tool result itself must NOT block on / embed diagnostics.
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleTool && strings.Contains(message.Content, "boom") {
+			t.Fatalf("diagnostics leaked into the tool result: %q", message.Content)
+		}
+	}
+	// The nudge must be present in the second request as a user message.
+	found := false
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			if !strings.Contains(message.Content, "boom") {
+				t.Fatalf("nudge missing diagnostics: %q", message.Content)
+			}
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("no async diagnostics nudge in the follow-up request")
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if want := filepath.Join(root, "main.go"); checkedPath != want {
+		t.Fatalf("checked path = %q, want %q", checkedPath, want)
+	}
+}
+
+// When the run hits the maxTurns ceiling right after an edit turn, the
+// final-answer request must still carry the diagnostics nudge — otherwise an
+// error introduced by the last edit would go unreported in the summary.
+func TestRunDrainsAsyncDiagnosticsBeforeMaxTurnsFinalAnswer(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "summary"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      t.TempDir(),
+		MaxTurns: 1,
+		FileDiagnostics: func(context.Context, string) string {
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "summary" {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected the max-turns final-answer request, got %d requests", len(provider.requests))
+	}
+	found := false
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("max-turns final-answer request missing the diagnostics nudge")
+	}
+}
+
+// When the model finalizes while a slow check is still in flight (the
+// per-turn drain missed it), the finalization gate must wait it out and give
+// the model one more turn with the nudge instead of dropping the errors.
+func TestRunFinalizationGateDeliversLateDiagnostics(t *testing.T) {
+	previous := asyncDiagnosticsDrainTimeout
+	asyncDiagnosticsDrainTimeout = time.Millisecond
+	defer func() { asyncDiagnosticsDrainTimeout = previous }()
+
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done after fix"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      t.TempDir(),
+		FileDiagnostics: func(context.Context, string) string {
+			// Slower than the (shortened) per-turn drain, well under the
+			// finalization budget.
+			time.Sleep(100 * time.Millisecond)
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done after fix" {
+		t.Fatalf("final answer = %q, want the post-nudge answer", result.FinalAnswer)
+	}
+	if len(provider.requests) != 3 {
+		t.Fatalf("expected a third request carrying the late nudge, got %d requests", len(provider.requests))
+	}
+	found := false
+	for _, message := range provider.requests[2].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			if !strings.Contains(message.Content, "boom") {
+				t.Fatalf("nudge missing diagnostics: %q", message.Content)
+			}
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("finalization gate did not deliver the late diagnostics nudge")
+	}
+}

internal/agent/command_prefix.go

@@ -2,6 +2,7 @@ package agent
 
 import (
 	"fmt"
+	"runtime"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
@@ -27,10 +28,20 @@ func proposedCommandPrefix(toolName string, args map[string]any) []string {
 		}
 		return nil
 	}
-	for _, tokens := range segments {
-		if !knownSafeCommandSegment(tokens) && sandbox.ValidCommandPrefix(tokens) {
-			return append([]string(nil), tokens...)
+	// Only propose approving a prefix of one segment when every other segment
+	// in the command is independently known-safe. Once a prefix is approved,
+	// shellExecutionArgsForApproval escalates the whole command (every
+	// segment, not just the approved one) to bypass the sandbox, so offering a
+	// prefix that leaves an MSYS-prone (or otherwise unsafe) segment uncovered
+	// would let that segment run unsandboxed without ever being reviewed.
+	for index, tokens := range segments {
+		if knownSafeCommandSegment(tokens) {
+			continue
 		}
+		if !sandbox.ValidCommandPrefix(tokens) || !otherSegmentsKnownSafe(segments, index) {
+			return nil
+		}
+		return append([]string(nil), tokens...)
 	}
 	if len(segments) == 0 || !sandbox.ValidCommandPrefix(segments[0]) {
 		return nil
@@ -38,6 +49,20 @@ func proposedCommandPrefix(toolName string, args map[string]any) []string {
 	return append([]string(nil), segments[0]...)
 }
 
+// otherSegmentsKnownSafe reports whether every segment other than the one at
+// skip is known-safe on its own.
+func otherSegmentsKnownSafe(segments [][]string, skip int) bool {
+	for index, tokens := range segments {
+		if index == skip {
+			continue
+		}
+		if !knownSafeCommandSegment(tokens) {
+			return false
+		}
+	}
+	return true
+}
+
 func matchCommandPrefix(toolName string, args map[string]any, options Options) (sandbox.CommandPrefixGrant, bool, bool) {
 	if !isShellCommandTool(toolName) || options.Sandbox == nil {
 		return sandbox.CommandPrefixGrant{}, false, false
@@ -168,10 +193,6 @@ func cleanPrefixRule(prefix []string) []string {
 	return cleaned
 }
 
-func safeRequestedPrefix(prefix []string, command []string) bool {
-	return safeRequestedPrefixForSegments(prefix, [][]string{command})
-}
-
 func safeRequestedPrefixForSegments(prefix []string, segments [][]string) bool {
 	if len(prefix) == 0 || !sandbox.ValidCommandPrefix(prefix) {
 		return false
@@ -196,14 +217,6 @@ func safeRequestedPrefixForSegments(prefix []string, segments [][]string) bool {
 	return matched
 }
 
-func safeShellCommandTokens(command string) ([]string, bool) {
-	segments, ok := safeShellCommandSegments(command)
-	if !ok || len(segments) != 1 {
-		return nil, false
-	}
-	return segments[0], true
-}
-
 func safeShellCommandSegments(command string) ([][]string, bool) {
 	command = strings.TrimSpace(command)
 	if command == "" {
@@ -275,7 +288,11 @@ func knownSafeCommandSegment(command []string) bool {
 	if len(command) == 0 {
 		return false
 	}
-	switch commandName(command[0]) {
+	name := commandName(command[0])
+	if runtime.GOOS == "windows" && tools.MsysProneCommandName(name) {
+		return false
+	}
+	switch name {
 	case "cat", "cd", "cut", "echo", "expr", "false", "grep", "head", "id",
 		"ls", "nl", "paste", "pwd", "rev", "seq", "stat", "tail", "tr",
 		"true", "uname", "uniq", "wc", "which", "whoami":

internal/agent/command_prefix_test.go

@@ -1,9 +1,11 @@
 package agent
 
 import (
+	"runtime"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
 )
 
 func TestProposedCommandPrefixUsesSafeSimpleCommands(t *testing.T) {
@@ -35,12 +37,37 @@ func TestProposedCommandPrefixHonorsValidatedRequestedPrefix(t *testing.T) {
 
 func TestProposedCommandPrefixSupportsSegmentedCommands(t *testing.T) {
 	got := proposedCommandPrefix("bash", map[string]any{"command": "ps aux | head -5"})
+	if runtime.GOOS == "windows" {
+		// head is MSYS-prone on Windows (#458), so proposedCommandPrefix must
+		// not offer "ps aux" as a reusable prefix here: approving it would
+		// escalate the whole command, including the uncovered head segment,
+		// to bypass the sandbox unreviewed. See
+		// TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered for
+		// the platform-independent regression coverage of this behavior.
+		if got != nil {
+			t.Fatalf("expected no prefix on Windows because head is MSYS-prone, got %#v", got)
+		}
+		return
+	}
 	want := []string{"ps", "aux"}
 	if !equalStringSlices(got, want) {
 		t.Fatalf("prefix = %#v, want %#v", got, want)
 	}
 }
 
+// TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered guards
+// against proposedCommandPrefix offering to approve one segment of a
+// multi-segment command (e.g. "ps aux") while a different segment (e.g. "npm
+// install") is not known-safe. shellExecutionArgsForApproval escalates the
+// entire command once any prefix is approved, so an uncovered unsafe segment
+// would bypass the sandbox unreviewed. Uses npm, which is never known-safe on
+// any platform, so the assertion does not depend on runtime.GOOS.
+func TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered(t *testing.T) {
+	if got := proposedCommandPrefix("bash", map[string]any{"command": "ps aux && npm install"}); got != nil {
+		t.Fatalf("expected no prefix because npm segment is not known-safe, got %#v", got)
+	}
+}
+
 func TestProposedCommandPrefixHonorsRequestedPrefixAcrossSegments(t *testing.T) {
 	got := proposedCommandPrefix("bash", map[string]any{
 		"command":     "git status --short && git status --branch",
@@ -120,13 +147,35 @@ func TestProposedCommandPrefixRejectsUnsafeLaunchers(t *testing.T) {
 func TestMatchCommandPrefixCoversSegmentedCommandWithSafeTail(t *testing.T) {
 	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
 	engine.GrantCommandPrefixForSession("bash", []string{"ps", "aux"})
+	// head is MSYS-prone on Windows (#458) and must not count as a known-safe tail.
+	command := "ps aux | head -5"
+	if runtime.GOOS == "windows" {
+		command = "ps aux | echo ok"
+	}
 
-	grant, ok, session := matchCommandPrefix("bash", map[string]any{"command": "ps aux | head -5"}, Options{Sandbox: engine})
+	grant, ok, session := matchCommandPrefix("bash", map[string]any{"command": command}, Options{Sandbox: engine})
 	if !ok || !session || !equalStringSlices(grant.Prefix, []string{"ps", "aux"}) {
 		t.Fatalf("match = %#v ok=%v session=%v, want session ps aux prefix", grant, ok, session)
 	}
 }
 
+func TestKnownSafeCommandSegmentRejectsMsysProneOnWindows(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows-only known-safe MSYS guard")
+	}
+	for _, command := range [][]string{{"head", "-5"}, {"cat", "file.txt"}, {"grep", "pat"}} {
+		if knownSafeCommandSegment(command) {
+			t.Fatalf("expected %q to be unsafe on Windows, got known-safe", command)
+		}
+	}
+	if !knownSafeCommandSegment([]string{"echo", "ok"}) {
+		t.Fatal("expected echo to remain known-safe on Windows")
+	}
+	if !tools.MsysProneCommandName("head") {
+		t.Fatal("expected head to be MSYS-prone")
+	}
+}
+
 func TestMatchCommandPrefixRejectsUncoveredSegment(t *testing.T) {
 	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
 	engine.GrantCommandPrefixForSession("bash", []string{"ps", "aux"})

internal/agent/compaction.go

@@ -325,6 +325,43 @@ type compactionState struct {
 	// OnText is deliberately NOT forwarded (compaction stays invisible to the user),
 	// but its token COST must still be counted so usage reports and budgets include it.
 	onUsage func(Usage)
+
+	// calibrationRatio scales the raw byte/4 token estimate toward the provider's
+	// real prompt-token count. ApproxTextTokens over-counts code-heavy content by
+	// ~15-20%, which would trip compaction early (at ~60% of true capacity). It
+	// starts at 1.0 and converges via an EMA as each turn reports actual usage, so
+	// later turns compact nearer to real capacity. Zero is treated as 1.0.
+	calibrationRatio float64
+}
+
+// calibrate folds one turn's (rawEstimate, actualPromptTokens) sample into the
+// running calibration ratio. A single sample is clamped to a sane band so an
+// outlier (a huge cache-read turn, a provider-overhead spike) can't skew the
+// estimate enough to disable or thrash compaction.
+func (state *compactionState) calibrate(rawEstimate int, actualPromptTokens int) {
+	if !state.enabled || rawEstimate <= 0 || actualPromptTokens <= 0 {
+		return
+	}
+	sample := float64(actualPromptTokens) / float64(rawEstimate)
+	if sample < 0.5 {
+		sample = 0.5
+	} else if sample > 2.0 {
+		sample = 2.0
+	}
+	if state.calibrationRatio <= 0 {
+		state.calibrationRatio = 1.0
+	}
+	const alpha = 0.3 // weight on the newest sample; smooths jitter across turns
+	state.calibrationRatio = state.calibrationRatio*(1-alpha) + sample*alpha
+}
+
+// calibratedTokens applies the learned ratio to a raw estimate. Before any sample
+// arrives (ratio unset) it returns the raw estimate unchanged.
+func (state *compactionState) calibratedTokens(raw int) int {
+	if state.calibrationRatio <= 0 {
+		return raw
+	}
+	return int(float64(raw) * state.calibrationRatio)
 }
 
 func newCompactionState(options Options) *compactionState {
@@ -353,7 +390,7 @@ func (state *compactionState) maybeCompact(
 	// the messages; both the threshold check and the shrink check below use the
 	// same term so they stay consistent.
 	toolTokens := estimateToolDefTokens(tools)
-	size := estimateTokens(messages) + toolTokens
+	size := state.calibratedTokens(estimateTokens(messages) + toolTokens)
 	if size <= state.threshold {
 		return messages
 	}
@@ -370,7 +407,7 @@ func (state *compactionState) maybeCompact(
 	// entirely and preserve recent turns verbatim.
 	if pruned, reclaimed := pruneStaleToolOutput(messages, state.preserveLast); reclaimed > 0 {
 		messages = pruned
-		size = estimateTokens(messages) + toolTokens
+		size = state.calibratedTokens(estimateTokens(messages) + toolTokens)
 		if size <= state.threshold {
 			state.lowWaterMark = size
 			return messages
@@ -386,7 +423,7 @@ func (state *compactionState) maybeCompact(
 		// later turn) can try again; we never drop messages on failure here.
 		return messages
 	}
-	newSize := estimateTokens(compacted) + toolTokens
+	newSize := state.calibratedTokens(estimateTokens(compacted) + toolTokens)
 	if newSize >= size {
 		// Compaction did not actually shrink anything (e.g. nothing to
 		// summarize). Leave the history untouched and don't churn next turn.
@@ -444,7 +481,7 @@ func (state *compactionState) recover(
 	// the SAME combined (messages + tool-defs) domain maybeCompact uses, so the
 	// proactive shrink-guard compares like with like.
 	state.reactiveAttempted = true
-	state.lowWaterMark = estimateTokens(result) + estimateToolDefTokens(tools)
+	state.lowWaterMark = state.calibratedTokens(estimateTokens(result) + estimateToolDefTokens(tools))
 	return result, true, nil
 }
 

internal/agent/compaction_calibrate_test.go

@@ -0,0 +1,50 @@
+package agent
+
+import "testing"
+
+func TestCalibrationConvergesTowardActual(t *testing.T) {
+	state := &compactionState{enabled: true}
+
+	// Before any sample, the estimate is unchanged.
+	if got := state.calibratedTokens(1000); got != 1000 {
+		t.Fatalf("uncalibrated tokens = %d, want 1000", got)
+	}
+
+	// The estimator over-counts: our raw estimate is 1000 but the provider reports
+	// 850 (a 0.85 ratio). Feeding that sample repeatedly should pull the ratio
+	// toward 0.85, so a 1000-token raw estimate calibrates downward.
+	for range 20 {
+		state.calibrate(1000, 850)
+	}
+	got := state.calibratedTokens(1000)
+	if got >= 1000 || got < 820 || got > 880 {
+		t.Fatalf("calibrated tokens = %d, want ~850 after convergence", got)
+	}
+}
+
+func TestCalibrateIgnoresDegenerateSamples(t *testing.T) {
+	state := &compactionState{enabled: true}
+	state.calibrate(0, 500) // zero estimate
+	state.calibrate(500, 0) // zero actual
+	if state.calibrationRatio != 0 {
+		t.Fatalf("degenerate samples must not move the ratio, got %v", state.calibrationRatio)
+	}
+	// Disabled state never calibrates.
+	disabled := &compactionState{enabled: false}
+	disabled.calibrate(1000, 850)
+	if disabled.calibrationRatio != 0 {
+		t.Fatalf("disabled compaction must not calibrate, got %v", disabled.calibrationRatio)
+	}
+}
+
+func TestCalibrateClampsOutliers(t *testing.T) {
+	state := &compactionState{enabled: true}
+	// A wild outlier (10x) is clamped to 2.0, so one bad sample can't blow up the
+	// ratio and disable compaction.
+	for range 50 {
+		state.calibrate(1000, 10000)
+	}
+	if state.calibrationRatio > 2.0 {
+		t.Fatalf("ratio should be clamped at 2.0, got %v", state.calibrationRatio)
+	}
+}

internal/agent/compaction_preserve.go

@@ -20,6 +20,16 @@ const (
 	toolNameUpdatePlan = "update_plan"
 	toolNameToolSearch = "tool_search"
 	toolNameSkill      = "skill"
+	toolNameWriteFile  = "write_file"
+	toolNameEditFile   = "edit_file"
+)
+
+// maxRecentEdits caps how many edited files are carried across a compaction, and
+// maxEditNoteBytes caps each edit's one-line note, so a long editing run can't
+// bloat the preserved-state block.
+const (
+	maxRecentEdits   = 20
+	maxEditNoteBytes = 160
 )
 
 const (
@@ -100,12 +110,123 @@ func formatPlanArguments(arguments string) string {
 }
 
 // skillEntry is a named preserved body. It began as loaded-skill state and is
-// reused for loaded tools and project instruction blocks.
+// reused for loaded tools, project instruction blocks, and recent file edits.
 type skillEntry struct {
 	name string
 	body string
 }
 
+// recentEdits returns the files mutated by write_file/edit_file calls in messages
+// — latest note per path, in last-seen order — as skillEntry{name: path, body:
+// note}. After compaction elides the editing turns, this tells the model WHAT it
+// changed in each file (from the tool's result) so it need not re-read to
+// rediscover its own footprint. Capped at maxRecentEdits paths.
+//
+// Ordering is by LAST edit, not first: re-editing a file moves it to the newest
+// position so the tail cap below keeps the file the model most recently touched
+// rather than an earlier, now-stale entry.
+func recentEdits(messages []zeroruntime.Message) []skillEntry {
+	pathByID := map[string]string{}
+	sequence := make([]string, 0)
+	for _, message := range messages {
+		for _, call := range message.ToolCalls {
+			if call.Name != toolNameWriteFile && call.Name != toolNameEditFile {
+				continue
+			}
+			path := editPathFromArguments(call.Arguments)
+			if path == "" {
+				continue
+			}
+			if call.ID != "" {
+				pathByID[call.ID] = path
+			}
+			sequence = append(sequence, path)
+		}
+	}
+	order := lastSeenOrder(sequence)
+	if len(order) == 0 {
+		return nil
+	}
+
+	noteByPath := map[string]string{}
+	for _, message := range messages {
+		if message.Role != zeroruntime.MessageRoleTool || message.ToolCallID == "" {
+			continue
+		}
+		if path, ok := pathByID[message.ToolCallID]; ok {
+			noteByPath[path] = editNote(message.Content)
+		}
+	}
+
+	// Keep the most recent maxRecentEdits paths (the tail of last-seen order).
+	if len(order) > maxRecentEdits {
+		order = order[len(order)-maxRecentEdits:]
+	}
+	entries := make([]skillEntry, 0, len(order))
+	for _, path := range order {
+		entries = append(entries, skillEntry{name: path, body: noteByPath[path]})
+	}
+	return entries
+}
+
+// lastSeenOrder dedupes paths keeping each at its LAST occurrence, preserving
+// chronological order otherwise. A re-edited path therefore lands at the newest
+// position rather than staying pinned to where it first appeared.
+func lastSeenOrder(paths []string) []string {
+	lastIdx := make(map[string]int, len(paths))
+	for i, p := range paths {
+		lastIdx[p] = i
+	}
+	order := make([]string, 0, len(lastIdx))
+	for i, p := range paths {
+		if lastIdx[p] == i {
+			order = append(order, p)
+		}
+	}
+	return order
+}
+
+// editPathFromArguments pulls the target file path from a write_file/edit_file
+// call's JSON arguments (path/file/file_path/filename aliases). Returns "" on
+// malformed arguments or no path.
+func editPathFromArguments(arguments string) string {
+	var parsed struct {
+		Path     string `json:"path"`
+		File     string `json:"file"`
+		FilePath string `json:"file_path"`
+		Filename string `json:"filename"`
+	}
+	if err := json.Unmarshal([]byte(strings.TrimSpace(arguments)), &parsed); err != nil {
+		return ""
+	}
+	for _, candidate := range []string{parsed.Path, parsed.File, parsed.FilePath, parsed.Filename} {
+		if s := strings.TrimSpace(candidate); s != "" {
+			return s
+		}
+	}
+	return ""
+}
+
+// editNote reduces an edit tool's result to a single short line (its first
+// non-empty line, byte-capped on a rune boundary) for the preserved summary.
+func editNote(content string) string {
+	for _, line := range strings.Split(content, "\n") {
+		s := strings.TrimSpace(line)
+		if s == "" {
+			continue
+		}
+		if len(s) > maxEditNoteBytes {
+			limit := maxEditNoteBytes
+			for limit > 0 && !utf8.RuneStart(s[limit]) {
+				limit--
+			}
+			s = s[:limit] + "…"
+		}
+		return s
+	}
+	return ""
+}
+
 // loadedSkills returns the skills loaded via the skill tool in messages — the
 // latest body per name, in first-seen order — matching each skill tool call to
 // its tool result by id.
@@ -310,11 +431,17 @@ func capBody(body string) string {
 // preservedState is the JSON shape of the carried-across-compaction block.
 type preservedState struct {
 	Plan                string                 `json:"plan,omitempty"`
+	RecentEdits         []preservedEdit        `json:"recent_edits,omitempty"`
 	Tools               []preservedTool        `json:"tools,omitempty"`
 	Skills              []preservedSkill       `json:"skills,omitempty"`
 	ProjectInstructions []preservedInstruction `json:"project_instructions,omitempty"`
 }
 
+type preservedEdit struct {
+	Path string `json:"path"`
+	Note string `json:"note,omitempty"`
+}
+
 type preservedTool struct {
 	Name string `json:"name"`
 	Body string `json:"body"`
@@ -347,6 +474,12 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		plan = priorState.Plan
 	}
 
+	// Recent edits: merge edits preserved earlier with fresh write_file/edit_file
+	// results in middle (newer note per path wins AND moves to the newest
+	// position), so the tail cap keeps the files the model most recently touched
+	// across repeated compactions rather than dropping a just-re-edited file.
+	edits := capRecentEdits(mergeRecentEdits(preservedEditsToEntries(priorState.RecentEdits), recentEdits(middle)))
+
 	// Tools: preserve deferred tool_search schemas from the transcript. Fresh
 	// loads override older carried copies by name.
 	tools := mergeSkillEntries(preservedToolsToEntries(priorState.Tools), loadedToolSchemas(middle))
@@ -360,19 +493,76 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		projectInstructionEntries(middle),
 	)
 
-	if block := formatPreservedState(plan, tools, skills, instructions); block != "" {
+	if block := formatPreservedState(plan, edits, tools, skills, instructions); block != "" {
 		summary += "\n\n" + block
 	}
 	return summary
 }
 
+// mergeRecentEdits overlays fresh edits onto edits preserved by an earlier
+// compaction. Unlike mergeSkillEntries (which keeps refreshed entries in their
+// original slot), a path touched again by a fresh edit MOVES to the newest
+// position and takes the fresh note, so capRecentEdits — which keeps only the
+// most-recent tail — never drops a file the model just re-edited. Paths present
+// only in the older set keep their relative order, ahead of the fresh ones.
+func mergeRecentEdits(older, newer []skillEntry) []skillEntry {
+	if len(newer) == 0 {
+		return older
+	}
+	freshBody := make(map[string]string, len(newer))
+	freshOrder := make([]string, 0, len(newer))
+	for _, e := range newer {
+		if _, ok := freshBody[e.name]; !ok {
+			freshOrder = append(freshOrder, e.name)
+		}
+		freshBody[e.name] = e.body
+	}
+	merged := make([]skillEntry, 0, len(older)+len(freshOrder))
+	// Older entries that were NOT re-edited keep their position, oldest first.
+	for _, e := range older {
+		if _, refreshed := freshBody[e.name]; refreshed {
+			continue // re-appended at its newest position below
+		}
+		merged = append(merged, e)
+	}
+	// Fresh edits follow, in last-seen order, each carrying the fresh note — so
+	// the most recently touched files sit at the tail the cap preserves.
+	for _, name := range freshOrder {
+		merged = append(merged, skillEntry{name: name, body: freshBody[name]})
+	}
+	return merged
+}
+
+// capRecentEdits bounds the preserved edit list to the most recent maxRecentEdits
+// entries after a merge, so repeated compactions can't grow it without limit.
+func capRecentEdits(entries []skillEntry) []skillEntry {
+	if len(entries) > maxRecentEdits {
+		return entries[len(entries)-maxRecentEdits:]
+	}
+	return entries
+}
+
+func preservedEditsToEntries(edits []preservedEdit) []skillEntry {
+	entries := make([]skillEntry, 0, len(edits))
+	for _, e := range edits {
+		if e.Path == "" {
+			continue
+		}
+		entries = append(entries, skillEntry{name: e.Path, body: e.Note})
+	}
+	return entries
+}
+
 // formatPreservedState renders state as the labelled, single-line
 // JSON block. Returns "" when there is nothing to preserve.
-func formatPreservedState(plan string, tools, skills, instructions []skillEntry) string {
-	if plan == "" && len(tools) == 0 && len(skills) == 0 && len(instructions) == 0 {
+func formatPreservedState(plan string, edits, tools, skills, instructions []skillEntry) string {
+	if plan == "" && len(edits) == 0 && len(tools) == 0 && len(skills) == 0 && len(instructions) == 0 {
 		return ""
 	}
 	state := preservedState{Plan: plan}
+	for _, e := range edits {
+		state.RecentEdits = append(state.RecentEdits, preservedEdit{Path: e.name, Note: e.body})
+	}
 	for _, t := range tools {
 		state.Tools = append(state.Tools, preservedTool{Name: t.name, Body: t.body})
 	}

internal/agent/compaction_recent_edits_test.go

@@ -0,0 +1,121 @@
+package agent
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// recentEdits extracts each mutated file's path and a one-line note from the
+// matching tool result, latest note per path in last-seen order.
+func TestRecentEditsExtractsPathsAndNotes(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e1", Name: "write_file", Arguments: `{"path":"internal/foo.go","content":"package foo"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e1", Content: "Wrote internal/foo.go (12 lines)"},
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e2", Name: "edit_file", Arguments: `{"path":"internal/bar.go","old_string":"a","new_string":"b"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e2", Content: "Applied edit to internal/bar.go"},
+	}
+
+	edits := recentEdits(messages)
+	if len(edits) != 2 {
+		t.Fatalf("expected 2 edited files, got %d: %#v", len(edits), edits)
+	}
+	if edits[0].name != "internal/foo.go" || !strings.Contains(edits[0].body, "12 lines") {
+		t.Fatalf("first edit = %#v, want foo.go with its note", edits[0])
+	}
+	if edits[1].name != "internal/bar.go" || !strings.Contains(edits[1].body, "Applied edit") {
+		t.Fatalf("second edit = %#v, want bar.go with its note", edits[1])
+	}
+}
+
+// After compaction elides the editing turns, the preserved-state block still
+// names the edited files and what changed, so the model needn't re-read them.
+func TestCompactionPreservesRecentEdits(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "system"},
+		{Role: zeroruntime.MessageRoleUser, Content: "add a flag"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "editing", ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e1", Name: "write_file", Arguments: `{"path":"cmd/main.go","content":"..."}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e1", Content: "Wrote cmd/main.go (adds --version flag)"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "done"},
+		{Role: zeroruntime.MessageRoleUser, Content: "continue"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "continuing"},
+	}
+	summary := compactStateConversation(t, messages)
+
+	state := parsePreservedStateBlock(summary)
+	if len(state.RecentEdits) != 1 {
+		t.Fatalf("expected 1 preserved edit, got %#v", state.RecentEdits)
+	}
+	if state.RecentEdits[0].Path != "cmd/main.go" || !strings.Contains(state.RecentEdits[0].Note, "--version") {
+		t.Fatalf("preserved edit = %#v, want cmd/main.go + its note", state.RecentEdits[0])
+	}
+}
+
+// A fresh edit note for a path overrides the one carried from an earlier
+// compaction (newer wins) and moves the path to the newest position rather than
+// duplicating it or leaving it pinned to its old slot.
+func TestRecentEditsMergeNewerWins(t *testing.T) {
+	prior := preservedState{RecentEdits: []preservedEdit{{Path: "a.go", Note: "old note"}, {Path: "z.go", Note: "z"}}}
+	older := preservedEditsToEntries(prior.RecentEdits)
+	newer := []skillEntry{{name: "a.go", body: "new note"}, {name: "b.go", body: "added"}}
+
+	merged := mergeRecentEdits(older, newer)
+	// z.go (untouched) keeps its lead; a.go is re-edited so it takes the new note
+	// and moves behind z.go, ahead of the brand-new b.go: [z.go, a.go, b.go].
+	if len(merged) != 3 {
+		t.Fatalf("expected z.go + refreshed a.go + b.go, got %#v", merged)
+	}
+	if merged[0].name != "z.go" {
+		t.Fatalf("untouched z.go should stay first, got %#v", merged)
+	}
+	if merged[1].name != "a.go" || merged[1].body != "new note" {
+		t.Fatalf("re-edited a.go should move to the newest edits with the new note, got %#v", merged[1])
+	}
+	if merged[2].name != "b.go" {
+		t.Fatalf("brand-new b.go should be last, got %#v", merged[2])
+	}
+}
+
+// Regression: once more than maxRecentEdits distinct files are tracked, re-editing
+// an early file must move it into the capped tail rather than leaving it to be
+// dropped — otherwise the next compaction omits exactly the file the model most
+// recently touched.
+func TestRecentEditsCapKeepsReeditedFile(t *testing.T) {
+	// Earlier compaction preserved exactly maxRecentEdits files, f0 the oldest.
+	older := make([]skillEntry, 0, maxRecentEdits)
+	for i := 0; i < maxRecentEdits; i++ {
+		older = append(older, skillEntry{name: fmt.Sprintf("f%d.go", i), body: "old"})
+	}
+	// Fresh window re-edits the OLDEST file (f0) and adds a brand-new file, so the
+	// merged list exceeds the cap and something must be dropped.
+	newer := []skillEntry{{name: "f0.go", body: "re-edited"}, {name: "fresh.go", body: "new"}}
+
+	merged := capRecentEdits(mergeRecentEdits(older, newer))
+	if len(merged) != maxRecentEdits {
+		t.Fatalf("cap should keep %d entries, got %d", maxRecentEdits, len(merged))
+	}
+
+	var f0 *skillEntry
+	for i := range merged {
+		if merged[i].name == "f0.go" {
+			f0 = &merged[i]
+		}
+		if merged[i].name == "f1.go" {
+			t.Fatalf("f1.go (now the least-recently touched) should have been dropped, got %#v", merged)
+		}
+	}
+	if f0 == nil {
+		t.Fatalf("re-edited f0.go must survive the cap, got %#v", merged)
+	}
+	if f0.body != "re-edited" {
+		t.Fatalf("surviving f0.go should carry the fresh note, got %q", f0.body)
+	}
+}

internal/agent/file_diagnostics.go

@@ -0,0 +1,62 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/lsp"
+)
+
+// fileDiagnosticsTimeout bounds one inline post-edit diagnostics check so a
+// slow or wedged language server can never hang a tool call; on timeout the
+// edit simply reports without a diagnostics block.
+const fileDiagnosticsTimeout = 10 * time.Second
+
+// NewFileDiagnostics adapts an *lsp.Manager to the per-edit inline diagnostics
+// callback (tools.RunOptions.Diagnostics): it reads the just-written file,
+// checks it against the file's language server, and formats error-severity
+// diagnostics for the model. Warnings and hints are excluded — nagging about
+// style on every edit is noise, while a type error the edit just introduced is
+// exactly what the model should see before its next step. Diagnostics are
+// rendered with workspace-relative paths: the absolute path would put the
+// local username/home directory into the model prompt and session transcript
+// on every edit. Returns nil when manager is nil, disabling inline diagnostics
+// entirely.
+func NewFileDiagnostics(manager *lsp.Manager, workspaceRoot string) func(context.Context, string) string {
+	if manager == nil {
+		return nil
+	}
+	return func(ctx context.Context, absPath string) string {
+		text, err := os.ReadFile(absPath)
+		if err != nil {
+			return ""
+		}
+		checkCtx, cancel := context.WithTimeout(ctx, fileDiagnosticsTimeout)
+		defer cancel()
+		diagnostics, err := manager.Check(checkCtx, absPath, string(text))
+		if err != nil {
+			return ""
+		}
+		errors := lsp.FilterBySeverity(diagnostics, lsp.SeverityError)
+		if len(errors) == 0 {
+			return ""
+		}
+		return lsp.FormatDiagnostics(diagnosticsDisplayPath(workspaceRoot, absPath), errors)
+	}
+}
+
+// diagnosticsDisplayPath renders absPath relative to the workspace root for
+// model-facing output, falling back to the file's base name when the path is
+// outside the workspace (a bare name still identifies the file without
+// exposing the directory layout).
+func diagnosticsDisplayPath(workspaceRoot, absPath string) string {
+	if workspaceRoot != "" {
+		if rel, err := filepath.Rel(workspaceRoot, absPath); err == nil && !strings.HasPrefix(rel, "..") {
+			return rel
+		}
+	}
+	return filepath.Base(absPath)
+}

internal/agent/file_diagnostics_test.go

@@ -0,0 +1,24 @@
+package agent
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+// Diagnostics are model-facing: absolute paths would leak the local username
+// and directory layout into the prompt and session transcript on every edit.
+func TestDiagnosticsDisplayPath(t *testing.T) {
+	root := filepath.Join("/Users", "someone", "project")
+	cases := []struct {
+		root, abs, want string
+	}{
+		{root, filepath.Join(root, "internal", "a.go"), filepath.Join("internal", "a.go")},
+		{root, filepath.Join("/etc", "other.go"), "other.go"}, // outside root -> base name only
+		{"", filepath.Join("/home", "user", "x.go"), "x.go"},  // no root -> base name only
+	}
+	for _, c := range cases {
+		if got := diagnosticsDisplayPath(c.root, c.abs); got != c.want {
+			t.Errorf("diagnosticsDisplayPath(%q, %q) = %q, want %q", c.root, c.abs, got, c.want)
+		}
+	}
+}

internal/agent/guardrails.go

@@ -22,6 +22,12 @@ const (
 	// calls have executed since the last update_plan call.
 	staleToolCallThreshold = 10
 
+	// stalePlanTurnThreshold injects the same one-shot reminder once this many
+	// turns have passed since the last update_plan while plan items are still
+	// pending — the turn-based complement to staleToolCallThreshold, catching a
+	// plan that drifts stale across many turns that each make few tool calls.
+	stalePlanTurnThreshold = 8
+
 	// toolOnlyProgressReminderAt injects a one-shot progress nudge after this
 	// many consecutive turns contain tool calls but no visible assistant text.
 	// It does not stop the run; it tells the model to synthesize what it already
@@ -43,8 +49,12 @@ const (
 	toolFailureHintAt = 2
 	// toolFailureStopAt halts the run after a tool fails this many times in a row
 	// with the same error, so NO model (weak or strong) burns turns looping on a
-	// bad call.
-	toolFailureStopAt = 4
+	// bad call. Set to 6 (not 4): a corrective hint fires at toolFailureHintAt (2),
+	// and a model iterating on a genuinely tricky edit can legitimately fail a few
+	// times after the hint while converging — stopping at 4 cut those runs short.
+	// The streak still resets the moment the tool succeeds or hits a different
+	// error, so this only affects true same-error loops.
+	toolFailureStopAt = 6
 
 	// maxContinueNudges bounds how many times the headless completion gate
 	// (Options.RequireCompletionSignal) re-prompts a model that stopped without a
@@ -354,8 +364,13 @@ type guardState struct {
 	emptyTurns               int
 	totalToolCalls           int
 	toolCallsSincePlanUpdate int
-	planEverCalled           bool
-	notCalledReminderSent    bool
+	// turnsSincePlanUpdate counts turns (not individual tool calls) since the last
+	// update_plan, so a plan that goes stale across many low-tool-call turns is
+	// still caught — the tool-call counter alone can take many turns to trip when
+	// the model makes only one call per turn.
+	turnsSincePlanUpdate  int
+	planEverCalled        bool
+	notCalledReminderSent bool
 	// staleReminderSent records whether the stale reminder has already fired for
 	// the current stale interval. It is cleared when a plan update opens a new
 	// interval, making the reminder one-shot per interval rather than per turn.
@@ -414,8 +429,9 @@ func (state *guardState) observeToolResult(name string, failed bool, output stri
 func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (stop bool) {
 	hasToolCalls := len(collected.ToolCalls) > 0
 	hasVisibleText := strings.TrimSpace(collected.Text) != ""
+	hasReasoning := collected.HasReasoning || len(collected.ReasoningBlocks) > 0
 
-	if hasToolCalls || hasVisibleText {
+	if hasToolCalls || hasVisibleText || hasReasoning {
 		state.emptyTurns = 0
 	} else {
 		state.emptyTurns++
@@ -427,11 +443,16 @@ func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (sto
 		state.toolOnlyReminderSent = false
 	}
 
+	// One turn has passed; the plan-update below resets this to 0 when the model
+	// refreshes the plan this turn.
+	state.turnsSincePlanUpdate++
+
 	for _, call := range collected.ToolCalls {
 		state.totalToolCalls++
 		if call.Name == planToolName {
 			state.planEverCalled = true
 			state.toolCallsSincePlanUpdate = 0
+			state.turnsSincePlanUpdate = 0
 			// A fresh plan update opens a new stale interval.
 			state.staleReminderSent = false
 			// Record how many items remain so the completion gate knows whether
@@ -485,10 +506,12 @@ func (state *guardState) progressReminder() string {
 // number of turns completed so far).
 func (state *guardState) planReminder(turn int) string {
 	// STALE reminder takes priority: a long run without a plan update is the
-	// stronger signal. One-shot per stale interval.
-	if state.planEverCalled &&
-		!state.staleReminderSent &&
-		state.toolCallsSincePlanUpdate >= staleToolCallThreshold {
+	// stronger signal. Fires on either the tool-call streak OR a turn streak with
+	// pending items (so a plan drifting stale across many low-call turns is caught,
+	// while a fully-completed plan is left alone). One-shot per stale interval.
+	if state.planEverCalled && !state.staleReminderSent &&
+		(state.toolCallsSincePlanUpdate >= staleToolCallThreshold ||
+			(state.turnsSincePlanUpdate >= stalePlanTurnThreshold && state.planItemsPending > 0)) {
 		state.staleReminderSent = true
 		return planStaleReminder(state.toolCallsSincePlanUpdate)
 	}

internal/agent/guardrails_test.go

@@ -22,6 +22,14 @@ func textTurn(content string) []zeroruntime.StreamEvent {
 	}
 }
 
+// reasoningTurn produces live reasoning without visible assistant text.
+func reasoningTurn(content string) []zeroruntime.StreamEvent {
+	return []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventReasoning, Content: content},
+		{Type: zeroruntime.StreamEventDone},
+	}
+}
+
 // toolTurn produces a turn that calls a named tool with the given args JSON.
 func toolTurn(callID string, toolName string, args string) []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{
@@ -101,6 +109,31 @@ func TestRunResetsEmptyTurnCounterOnVisibleOutput(t *testing.T) {
 	}
 }
 
+func TestRunResetsEmptyTurnCounterOnReasoning(t *testing.T) {
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			reasoningTurn("thinking 1"),
+			reasoningTurn("thinking 2"),
+			reasoningTurn("thinking 3"),
+			textTurn("done"),
+		},
+	}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("expected reasoning-only turns to keep the run live until final answer, got %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 4 {
+		t.Fatalf("expected 4 turns, got %d", len(provider.requests))
+	}
+}
+
 func TestRunResetsEmptyTurnCounterOnToolCall(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")

internal/agent/loop.go

@@ -25,7 +25,12 @@ const maxTurnsFinalAnswerPrompt = "You have reached the tool-turn limit. Do not
 // WITH NO OUTPUT yet is re-issued on a fresh connection before giving up. Only
 // the no-output case is retried (a partial turn would duplicate), so this is a
 // safe recovery for a stalled/dead pooled connection.
-const maxStreamStallRetries = 2
+//
+// Set to 1 (not 2): each attempt can itself idle for the full stream timeout
+// (~5min) before the stall is even detected, so 2 retries left an interactive
+// session frozen for ~15min. One retry keeps the common single-hiccup recovery
+// while bounding the worst case to ~2× the idle timeout.
+const maxStreamStallRetries = 1
 
 const (
 	toolResultMetaControl       = "control"
@@ -132,6 +137,12 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	guards := newGuardState()
 	compactor := newCompactionState(options)
 
+	// Background post-edit diagnostics: files changed by mutating tools are
+	// checked off the tool-call critical path and any errors are appended as a
+	// nudge before the NEXT request (see async_diagnostics.go). nil when no
+	// FileDiagnostics callback is wired; every method no-ops on nil.
+	postEditDiagnostics := newAsyncDiagnostics(options.FileDiagnostics, options.Cwd)
+
 	// loaded tracks deferred-eligible tools the model has pulled via tool_search
 	// during THIS run. It is consulted by partitionTools each turn to expose a
 	// loaded tool's full schema; it lives only for the run (v1 within-run scope).
@@ -145,15 +156,31 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	// has already been demanded this run, so it fires at most once.
 	acceptanceRequested := false
 
+	// toolDefCache memoizes each tool's rendered JSON-schema definition across
+	// turns (a tool's advertised schema is stable for the run), so partitionTools
+	// doesn't re-run the recursive schema→map conversion for every tool every turn.
+	toolDefCache := map[string]zeroruntime.ToolDefinition{}
+
 	result := Result{Messages: copyMessages(messages)}
 	for turn := 0; turn < maxTurns; turn++ {
 		result.Turns = turn + 1
 
+		// Deliver background post-edit diagnostics from the previous turn's edits
+		// BEFORE compaction so the nudge is part of the request being budgeted.
+		// A brief wait at most (asyncDiagnosticsDrainTimeout); an unfinished check
+		// simply delivers on a later turn.
+		if nudge := postEditDiagnostics.drain(ctx); nudge != "" {
+			messages = append(messages, zeroruntime.Message{
+				Role:    zeroruntime.MessageRoleUser,
+				Content: nudge,
+			})
+		}
+
 		// Build the per-turn tool list first so proactive compaction can include
 		// the tool-definition tokens (they ride on every request) in its estimate.
 		// partitionTools depends only on registry/permissions/options/loaded, not on
 		// the messages, so computing it before compaction is safe.
-		exposed, _ := partitionTools(registry, permissionMode, options, loaded)
+		exposed, _ := partitionToolsCached(registry, permissionMode, options, loaded, toolDefCache)
 
 		// PROACTIVE compaction: if the history is approaching the model's
 		// context window, summarize the oldest middle before building the
@@ -163,6 +190,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			Messages:        copyMessages(messages),
 			Tools:           exposed,
 			ReasoningEffort: options.ReasoningEffort,
+			PromptCacheKey:  options.SessionID,
 		}
 
 		// Report the per-category context budget for this turn so a surface can
@@ -198,6 +226,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					Messages:        copyMessages(messages),
 					Tools:           exposed,
 					ReasoningEffort: options.ReasoningEffort,
+					PromptCacheKey:  options.SessionID,
 				}
 				// Pre-content connect after a context-limit compaction: route through the
 				// reconnect helper so a transient upstream hiccup here doesn't fail the
@@ -258,6 +287,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					Messages:        copyMessages(messages),
 					Tools:           exposed,
 					ReasoningEffort: options.ReasoningEffort,
+					PromptCacheKey:  options.SessionID,
 				}
 				retryStream, retryStreamErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
 				if retryStreamErr != nil {
@@ -316,6 +346,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				Messages:        copyMessages(messages),
 				Tools:           exposed,
 				ReasoningEffort: options.ReasoningEffort,
+				PromptCacheKey:  options.SessionID,
 			}
 			retryStream, retryErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
 			if retryErr != nil {
@@ -340,6 +371,16 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 		}
 
+		// Calibrate the compaction token estimator against the provider's real
+		// prompt-token count for the request we just sent, so later turns trigger
+		// compaction near true capacity instead of ~15% early on code-heavy history.
+		// Recompute the estimate HERE (not at request-build time): a reactive
+		// compaction may have replaced `messages` with a smaller set and re-sent, so
+		// this reflects the request that actually produced collected.Usage. The
+		// assistant reply is appended below, after this, so `messages` is still the
+		// sent request.
+		compactor.calibrate(estimateTokens(messages)+estimateToolDefTokens(exposed), collected.Usage.InputTokens)
+
 		// Carry the turn's terminal stop reason so a final answer cut off at the
 		// output token cap (or by a content filter) is reported as truncated. A
 		// tool-call turn normalizes to "" and clears any prior reason.
@@ -450,6 +491,19 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					continue
 				}
 			}
+			// Finalization diagnostics gate: edits from this run may still have
+			// checks in flight — the per-turn drain waits only briefly and defers
+			// to "a later turn", but a final answer means there is no later turn.
+			// Wait out the full inline-era budget once; an introduced error gives
+			// the model one more turn to see (and fix) it instead of being lost.
+			// Free for runs that never edited: an idle collector returns "".
+			if nudge := postEditDiagnostics.drainFinal(ctx); nudge != "" {
+				messages = append(messages, zeroruntime.Message{
+					Role:    zeroruntime.MessageRoleUser,
+					Content: nudge,
+				})
+				continue
+			}
 			result.FinalAnswer = collected.Text
 			result.Messages = copyMessages(messages)
 			return result, nil
@@ -475,11 +529,35 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// between tool_results breaks strict provider replay) — same after-batch
 		// rationale as turnRequestedModel above.
 		var changedFilesThisBatch []string
+		// Parallel read-ahead state: results for calls[precomputedStart:precomputedEnd]
+		// executed concurrently, consumed strictly in order below.
+		var precomputed []precomputedToolResult
+		precomputedStart, precomputedEnd := 0, 0
 		for index, call := range collected.ToolCalls {
+			// When this call starts a consecutive run of >= 2 auto-allowed read-only
+			// calls, execute the whole run concurrently now (see parallel_tools.go).
+			// The scan happens lazily at the run's first index — never ahead of a
+			// pending mutating call — so read-after-write ordering is preserved.
+			if index >= precomputedEnd {
+				runEnd := index
+				for runEnd < len(collected.ToolCalls) && parallelSafeToolCall(registry, collected.ToolCalls[runEnd], options) {
+					runEnd++
+				}
+				if runEnd-index >= 2 {
+					precomputed = executeParallelReadBatch(ctx, registry, collected.ToolCalls, index, runEnd, permissionMode, options)
+					precomputedStart, precomputedEnd = index, runEnd
+				}
+			}
 			if options.OnToolCall != nil {
 				options.OnToolCall(call)
 			}
-			toolResult, abortErr := executeToolCall(ctx, registry, call, permissionMode, options)
+			var toolResult ToolResult
+			var abortErr error
+			if index >= precomputedStart && index < precomputedEnd {
+				toolResult, abortErr = precomputed[index-precomputedStart].result, precomputed[index-precomputedStart].abortErr
+			} else {
+				toolResult, abortErr = executeToolCall(ctx, registry, call, permissionMode, options)
+			}
 			if options.OnToolResult != nil {
 				options.OnToolResult(toolResult)
 			}
@@ -542,11 +620,14 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				failureHint = toolFailureHint(call.Name, toolSchemaJSON(registry, call.Name), toolResult.Output)
 			}
 
-			// Post-edit self-correction: collect the files this successful mutating
-			// tool changed; verification runs once over the union after the batch.
-			// A read-only tool (no ChangedFiles) never contributes.
-			if options.SelfCorrect != nil && toolResult.Status == tools.StatusOK && len(toolResult.ChangedFiles) > 0 {
+			// Collect the files this successful mutating tool changed. Self-correct
+			// verification runs once over the union after the batch; background
+			// diagnostics start NOW so the language server analyzes while the rest
+			// of the batch executes. A read-only tool (no ChangedFiles) never
+			// contributes.
+			if toolResult.Status == tools.StatusOK && len(toolResult.ChangedFiles) > 0 {
 				changedFilesThisBatch = append(changedFilesThisBatch, toolResult.ChangedFiles...)
+				postEditDiagnostics.enqueue(ctx, toolResult.ChangedFiles)
 			}
 		}
 
@@ -629,6 +710,17 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	// mid-run, not a natural completion (a finished run returns via the no-tool-call
 	// success path before this). Under the headless completion gate that is INCOMPLETE,
 	// not success, so a run that loops to the turn limit isn't reported as done.
+	//
+	// The final-answer call is one more model request, so diagnostics from the
+	// LAST turn's edits get a drain too — with the finalization budget, since
+	// there is no later turn to defer to — otherwise an error introduced by
+	// the final edit would go unreported in the summary.
+	if nudge := postEditDiagnostics.drainFinal(ctx); nudge != "" {
+		messages = append(messages, zeroruntime.Message{
+			Role:    zeroruntime.MessageRoleUser,
+			Content: nudge,
+		})
+	}
 	if answer, finalMessages, finishReason := finalAnswerAfterMaxTurns(ctx, provider, messages, options); strings.TrimSpace(answer) != "" {
 		result.FinalAnswer = answer
 		result.FinishReason = finishReason
@@ -661,6 +753,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 	stream, err := streamWithReconnect(ctx, provider, zeroruntime.CompletionRequest{
 		Messages:        copyMessages(finalMessages),
 		ReasoningEffort: options.ReasoningEffort,
+		PromptCacheKey:  options.SessionID,
 	}, reconnectNoticeFor(options))
 	if err != nil {
 		return "", messages, ""
@@ -1051,6 +1144,12 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 		// Per-session file version tracker so write_file/edit_file refuse to clobber
 		// a file that changed on disk outside Zero since it was last read.
 		FileTracker: options.FileTracker,
+		// Post-edit diagnostics are NOT forwarded to the tools: they used to run
+		// synchronously inside edit_file/write_file and block every mutating tool
+		// call on the language server (≥300ms debounce floor, 10s cap). The loop
+		// now checks changed files in the background and nudges the model before
+		// its next request instead (see async_diagnostics.go). The tools' inline
+		// mechanism (tools.RunOptions.Diagnostics) stays for direct API callers.
 		// Forward the run's operator tool filters so a filter-aware tool
 		// (tool_search) never discloses or loads an operator-hidden deferred tool.
 		EnabledTools:  options.EnabledTools,
@@ -2462,6 +2561,17 @@ func permissionActionFromSandbox(action sandbox.Action) PermissionAction {
 // exposed. The exposed slice is alpha-sorted by name, matching the legacy order
 // so the inactive path is stable.
 func partitionTools(registry *tools.Registry, permissionMode PermissionMode, options Options, loaded map[string]bool) ([]zeroruntime.ToolDefinition, string) {
+	return partitionToolsCached(registry, permissionMode, options, loaded, nil)
+}
+
+// partitionToolsCached is partitionTools with an optional per-tool definition
+// cache. The partitioning itself (visibility, deferral, ordering) is recomputed
+// every call — it must be, because a tool's deferred state can flip mid-run (e.g.
+// swarm tools un-defer once a swarm is active). Only the expensive part —
+// rendering each tool's JSON-schema parameters — is memoized by tool name, since a
+// tool's advertised name/description/schema is stable for the run. defCache nil
+// disables caching (used by tests and the plain partitionTools entrypoint).
+func partitionToolsCached(registry *tools.Registry, permissionMode PermissionMode, options Options, loaded map[string]bool, defCache map[string]zeroruntime.ToolDefinition) ([]zeroruntime.ToolDefinition, string) {
 	registeredTools := registry.All()
 
 	visible := make([]tools.Tool, 0, len(registeredTools))
@@ -2505,7 +2615,7 @@ func partitionTools(registry *tools.Registry, permissionMode PermissionMode, opt
 			if tool.Name() == tools.ToolSearchToolName {
 				continue
 			}
-			definitions = append(definitions, runtimeToolDefinition(tool))
+			definitions = append(definitions, cachedRuntimeToolDefinition(defCache, tool))
 		}
 		sort.Slice(definitions, func(left int, right int) bool {
 			return definitions[left].Name < definitions[right].Name
@@ -2537,13 +2647,13 @@ func partitionTools(registry *tools.Registry, permissionMode PermissionMode, opt
 		}
 		if tools.IsDeferred(tool) {
 			if loaded[name] {
-				loadedTail = append(loadedTail, runtimeToolDefinition(tool))
+				loadedTail = append(loadedTail, cachedRuntimeToolDefinition(defCache, tool))
 			} else {
 				hiddenTools = append(hiddenTools, tool)
 			}
 			continue
 		}
-		eager = append(eager, runtimeToolDefinition(tool))
+		eager = append(eager, cachedRuntimeToolDefinition(defCache, tool))
 	}
 	sort.Slice(eager, func(left int, right int) bool {
 		return eager[left].Name < eager[right].Name
@@ -2577,6 +2687,24 @@ func partitionTools(registry *tools.Registry, permissionMode PermissionMode, opt
 	return definitions, discovery
 }
 
+// cachedRuntimeToolDefinition returns the tool's rendered definition, reusing a
+// cached render when defCache holds one for this tool name. A tool's advertised
+// definition is stable across a run, so caching skips the recursive schema→map
+// conversion (schemaToRuntimeMap) that would otherwise run for every tool on every
+// turn. tool_search is excluded by its callers (its description is dynamic), so it
+// never poisons the cache. A nil cache computes fresh.
+func cachedRuntimeToolDefinition(defCache map[string]zeroruntime.ToolDefinition, tool tools.Tool) zeroruntime.ToolDefinition {
+	if defCache == nil {
+		return runtimeToolDefinition(tool)
+	}
+	if def, ok := defCache[tool.Name()]; ok {
+		return def
+	}
+	def := runtimeToolDefinition(tool)
+	defCache[tool.Name()] = def
+	return def
+}
+
 // runtimeToolDefinition renders a tool's advertised definition (name, description,
 // JSON-schema parameters) as sent to the provider.
 func runtimeToolDefinition(tool tools.Tool) zeroruntime.ToolDefinition {

internal/agent/loop_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -1945,6 +1946,11 @@ func TestRunCommandPrefixApprovalBypassesSandboxForMatchingShellCalls(t *testing
 
 func TestRunCommandPrefixApprovalCoversSegmentedShellWithSafeTail(t *testing.T) {
 	root := t.TempDir()
+	segmentedCommand := `ps aux | head -5`
+	if runtime.GOOS == "windows" {
+		// head is MSYS-prone on Windows (#458) and no longer counts as a known-safe tail.
+		segmentedCommand = `ps aux | echo ok`
+	}
 	retryTool := &sandboxDeniedRetryTool{}
 	registry := tools.NewRegistry()
 	registry.Register(retryTool)
@@ -1958,7 +1964,7 @@ func TestRunCommandPrefixApprovalCoversSegmentedShellWithSafeTail(t *testing.T)
 			},
 			{
 				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-2", ToolName: "bash"},
-				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-2", ArgumentsFragment: `{"command":"ps aux | head -5"}`},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-2", ArgumentsFragment: fmt.Sprintf(`{"command":%q}`, segmentedCommand)},
 				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-2"},
 				{Type: zeroruntime.StreamEventDone},
 			},
@@ -2907,8 +2913,10 @@ func TestBuildSystemPromptInjectsHostShellContext(t *testing.T) {
 		t.Fatalf("expected operating system in environment block, got %q", prompt)
 	}
 	if runtime.GOOS == "windows" {
-		if !strings.Contains(prompt, "Windows cmd.exe syntax") || !strings.Contains(prompt, "cwd argument") {
-			t.Fatalf("expected Windows cmd.exe shell guidance in prompt, got %q", prompt)
+		for _, want := range []string{"Windows cmd.exe syntax", "cwd argument", "MSYS binaries", "grep", "require_escalated"} {
+			if !strings.Contains(prompt, want) {
+				t.Fatalf("expected Windows shell guidance to mention %q, got %q", want, prompt)
+			}
 		}
 	} else if !strings.Contains(prompt, "/bin/sh syntax") {
 		t.Fatalf("expected POSIX shell guidance in prompt, got %q", prompt)

internal/agent/parallel_tools.go

@@ -0,0 +1,97 @@
+package agent
+
+import (
+	"context"
+	"sync"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// Parallel read-ahead for tool batches. When a turn requests several
+// independent lookups (read_file + grep + glob is the common shape), executing
+// them one after another serializes pure I/O waits. A consecutive run of
+// auto-allowed read-only calls is executed concurrently instead; results are
+// then consumed in the original call order, so guard counters, message
+// ordering, abort semantics, and the surface's call/result event pairing are
+// byte-identical to sequential execution. Runs never span a mutating call: a
+// read that follows a write must observe the write, so eligibility is decided
+// per consecutive run, not per batch.
+
+// maxParallelReadTools bounds concurrent read-only tool executions in a turn.
+const maxParallelReadTools = 8
+
+// precomputedToolResult is one parallel read-ahead execution, keyed back to
+// its batch index by the caller.
+type precomputedToolResult struct {
+	result   ToolResult
+	abortErr error
+}
+
+// parallelSafeToolCall reports whether call may run concurrently with its
+// neighbors: the tool must exist, be side-effect-free (SideEffectRead), and be
+// auto-allowed for these args, so no interactive prompt or workspace mutation
+// is on the hot path. Loop-intercepted tools (ask_user, request_permissions)
+// and tool_search (mutates the deferred-tool set) stay sequential.
+func parallelSafeToolCall(registry *tools.Registry, call ToolCall, options Options) bool {
+	switch call.Name {
+	case "ask_user", tools.RequestPermissionsToolName, tools.ToolSearchToolName:
+		return false
+	}
+	tool, found := registry.Get(call.Name)
+	if !found || tool.Safety().SideEffect != tools.SideEffectRead {
+		return false
+	}
+	args := map[string]any{}
+	if call.Arguments != "" {
+		if err := decodeToolArguments(call.Arguments, &args); err != nil {
+			return false
+		}
+	}
+	return effectivePermission(tool, args) == tools.PermissionAllow
+}
+
+// executeParallelReadBatch runs calls[start:end] concurrently (bounded by
+// maxParallelReadTools) and returns results indexed relative to start. All
+// execution-side callbacks that can fire inside executeToolCall are serialized
+// behind one mutex: a permission prompt (a sandbox preflight can demand one
+// even for an auto-allowed read) must never appear twice at once on an
+// interactive front-end, and OnPermission event handlers append to shared
+// session-recording state without their own locking — two batched reads under
+// a granted extra root would otherwise race (the pre-batch serial loop never
+// had two callbacks in flight at once).
+func executeParallelReadBatch(ctx context.Context, registry *tools.Registry, calls []ToolCall, start, end int, permissionMode PermissionMode, options Options) []precomputedToolResult {
+	batchOptions := options
+	var callbackMutex sync.Mutex
+	if options.OnPermissionRequest != nil {
+		inner := options.OnPermissionRequest
+		batchOptions.OnPermissionRequest = func(ctx context.Context, request PermissionRequest) (PermissionDecision, error) {
+			callbackMutex.Lock()
+			defer callbackMutex.Unlock()
+			return inner(ctx, request)
+		}
+	}
+	if options.OnPermission != nil {
+		inner := options.OnPermission
+		batchOptions.OnPermission = func(event PermissionEvent) {
+			callbackMutex.Lock()
+			defer callbackMutex.Unlock()
+			inner(event)
+		}
+	}
+
+	results := make([]precomputedToolResult, end-start)
+	semaphore := make(chan struct{}, maxParallelReadTools)
+	var waitGroup sync.WaitGroup
+	for index := start; index < end; index++ {
+		waitGroup.Add(1)
+		go func(index int) {
+			defer waitGroup.Done()
+			semaphore <- struct{}{}
+			defer func() { <-semaphore }()
+			result, abortErr := executeToolCall(ctx, registry, calls[index], permissionMode, batchOptions)
+			results[index-start] = precomputedToolResult{result: result, abortErr: abortErr}
+		}(index)
+	}
+	waitGroup.Wait()
+	return results
+}

internal/agent/parallel_tools_test.go

@@ -0,0 +1,204 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// probeTool records execution overlap and ordering so tests can assert what
+// actually ran concurrently.
+type probeTool struct {
+	name       string
+	sideEffect tools.SideEffect
+	delay      time.Duration
+	// shared, when set, receives this probe's start/end entries too, so a test
+	// can observe ordering ACROSS probes (e.g. reads vs a write barrier).
+	shared *probeLog
+
+	mu        sync.Mutex
+	active    int
+	maxActive int
+	log       []string
+}
+
+// probeLog is a mutex-guarded event log shared between probes.
+type probeLog struct {
+	mu      sync.Mutex
+	entries []string
+}
+
+func (log *probeLog) append(entry string) {
+	log.mu.Lock()
+	defer log.mu.Unlock()
+	log.entries = append(log.entries, entry)
+}
+
+func (log *probeLog) snapshot() []string {
+	log.mu.Lock()
+	defer log.mu.Unlock()
+	return append([]string(nil), log.entries...)
+}
+
+func (tool *probeTool) Name() string        { return tool.name }
+func (tool *probeTool) Description() string { return "test probe tool" }
+func (tool *probeTool) Parameters() tools.Schema {
+	return tools.Schema{
+		Type:                 "object",
+		Properties:           map[string]tools.PropertySchema{"id": {Type: "string"}},
+		AdditionalProperties: false,
+	}
+}
+func (tool *probeTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tool.sideEffect, Permission: tools.PermissionAllow, Reason: "test"}
+}
+func (tool *probeTool) Run(_ context.Context, args map[string]any) tools.Result {
+	id, _ := args["id"].(string)
+	tool.mu.Lock()
+	tool.active++
+	if tool.active > tool.maxActive {
+		tool.maxActive = tool.active
+	}
+	tool.log = append(tool.log, "start:"+id)
+	tool.mu.Unlock()
+	if tool.shared != nil {
+		tool.shared.append("start:" + id)
+	}
+	time.Sleep(tool.delay)
+	if tool.shared != nil {
+		tool.shared.append("end:" + id)
+	}
+	tool.mu.Lock()
+	tool.active--
+	tool.log = append(tool.log, "end:"+id)
+	tool.mu.Unlock()
+	return tools.Result{Status: tools.StatusOK, Output: "probe " + id}
+}
+
+func probeCallEvents(callID, toolName, id string) []zeroruntime.StreamEvent {
+	return []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: callID, ToolName: toolName},
+		{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: callID, ArgumentsFragment: fmt.Sprintf(`{"id":%q}`, id)},
+		{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: callID},
+	}
+}
+
+func TestParallelSafeToolCall(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(&probeTool{name: "probe_read", sideEffect: tools.SideEffectRead})
+	registry.Register(&probeTool{name: "probe_write", sideEffect: tools.SideEffectWrite})
+
+	call := func(name, args string) ToolCall { return ToolCall{ID: "c", Name: name, Arguments: args} }
+	if !parallelSafeToolCall(registry, call("probe_read", `{"id":"a"}`), Options{}) {
+		t.Fatal("auto-allowed read tool must be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("probe_write", `{"id":"a"}`), Options{}) {
+		t.Fatal("mutating tool must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("unknown_tool", `{}`), Options{}) {
+		t.Fatal("unknown tool must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("probe_read", `{"id":`), Options{}) {
+		t.Fatal("undecodable arguments must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("ask_user", `{}`), Options{}) {
+		t.Fatal("loop-intercepted tools must stay sequential")
+	}
+}
+
+func TestRunExecutesConsecutiveReadsConcurrently(t *testing.T) {
+	probe := &probeTool{name: "probe_read", sideEffect: tools.SideEffectRead, delay: 60 * time.Millisecond}
+	registry := tools.NewRegistry()
+	registry.Register(probe)
+
+	turnOne := append(probeCallEvents("call-1", "probe_read", "a"), probeCallEvents("call-2", "probe_read", "b")...)
+	turnOne = append(turnOne, probeCallEvents("call-3", "probe_read", "c")...)
+	turnOne = append(turnOne, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			turnOne,
+			{{Type: zeroruntime.StreamEventText, Content: "done"}, {Type: zeroruntime.StreamEventDone}},
+		},
+	}
+
+	var results []ToolResult
+	_, err := Run(context.Background(), "probe", provider, Options{
+		Registry:     registry,
+		OnToolResult: func(result ToolResult) { results = append(results, result) },
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if probe.maxActive < 2 {
+		t.Fatalf("consecutive read-only calls must overlap, max concurrency was %d", probe.maxActive)
+	}
+	// Results must still be recorded in original call order.
+	if len(results) != 3 || results[0].ToolCallID != "call-1" || results[1].ToolCallID != "call-2" || results[2].ToolCallID != "call-3" {
+		t.Fatalf("tool results out of order: %#v", results)
+	}
+	messages := provider.requests[1].Messages
+	var toolOrder []string
+	for _, message := range messages {
+		if message.Role == zeroruntime.MessageRoleTool {
+			toolOrder = append(toolOrder, message.ToolCallID)
+		}
+	}
+	if len(toolOrder) != 3 || toolOrder[0] != "call-1" || toolOrder[1] != "call-2" || toolOrder[2] != "call-3" {
+		t.Fatalf("recorded tool messages out of order: %v", toolOrder)
+	}
+}
+
+func TestRunParallelReadsNeverSpanMutatingCall(t *testing.T) {
+	shared := &probeLog{}
+	read := &probeTool{name: "probe_read", sideEffect: tools.SideEffectRead, delay: 30 * time.Millisecond, shared: shared}
+	write := &probeTool{name: "probe_write", sideEffect: tools.SideEffectWrite, shared: shared}
+	registry := tools.NewRegistry()
+	registry.Register(read)
+	registry.Register(write)
+
+	turnOne := append(probeCallEvents("call-1", "probe_read", "r1"), probeCallEvents("call-2", "probe_read", "r2")...)
+	turnOne = append(turnOne, probeCallEvents("call-3", "probe_write", "w")...)
+	turnOne = append(turnOne, probeCallEvents("call-4", "probe_read", "r3")...)
+	turnOne = append(turnOne, probeCallEvents("call-5", "probe_read", "r4")...)
+	turnOne = append(turnOne, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			turnOne,
+			{{Type: zeroruntime.StreamEventText, Content: "done"}, {Type: zeroruntime.StreamEventDone}},
+		},
+	}
+
+	_, err := Run(context.Background(), "probe", provider, Options{Registry: registry})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Cross-probe ordering on the SHARED log: the write must start only after
+	// both first-batch reads finished, and both second-batch reads must start
+	// only after the write finished — batches never cross a mutating call.
+	log := shared.snapshot()
+	index := func(entry string) int {
+		for i, e := range log {
+			if e == entry {
+				return i
+			}
+		}
+		t.Fatalf("entry %q missing from shared log: %v", entry, log)
+		return -1
+	}
+	writeStart, writeEnd := index("start:w"), index("end:w")
+	if firstBatchMaxEnd := max(index("end:r1"), index("end:r2")); writeStart < firstBatchMaxEnd {
+		t.Fatalf("write started before the first read batch finished: %v", log)
+	}
+	if secondBatchMinStart := min(index("start:r3"), index("start:r4")); secondBatchMinStart < writeEnd {
+		t.Fatalf("second read batch started before the write finished: %v", log)
+	}
+	if read.maxActive < 2 {
+		t.Fatalf("reads within a batch must overlap, max concurrency was %d", read.maxActive)
+	}
+}

internal/agent/partition_cache_stable_test.go

@@ -34,7 +34,7 @@ func TestPartitionToolsActiveAppendsLoadedToolAfterEagerBlock(t *testing.T) {
 		t.Fatalf("expected read_file, tool_search and loaded alpha all exposed, got %#v", exposed)
 	}
 	// The eager tools must precede the loaded deferred tool (appended, not inserted).
-	if !(readPos < alphaPos && searchPos < alphaPos) {
+	if readPos >= alphaPos || searchPos >= alphaPos {
 		t.Fatalf("loaded deferred tool must be appended after the eager block; got read=%d search=%d alpha=%d", readPos, searchPos, alphaPos)
 	}
 	// beta was never loaded — it stays hidden.

internal/agent/partition_cache_test.go

@@ -0,0 +1,63 @@
+package agent
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// countingSchemaTool records how many times its Parameters() schema is read, so a
+// test can prove the definition cache avoids re-rendering across turns.
+type countingSchemaTool struct {
+	name  string
+	calls *int
+}
+
+func (t countingSchemaTool) Name() string        { return t.name }
+func (t countingSchemaTool) Description() string { return "counts schema reads" }
+func (t countingSchemaTool) Parameters() tools.Schema {
+	*t.calls++
+	return tools.Schema{Type: "object", AdditionalProperties: false, Properties: map[string]tools.PropertySchema{
+		"x": {Type: "string"},
+	}}
+}
+func (t countingSchemaTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionAllow}
+}
+func (t countingSchemaTool) Run(_ context.Context, _ map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// 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) {
+	calls := 0
+	registry := tools.NewRegistry()
+	registry.Register(countingSchemaTool{name: "alpha", calls: &calls})
+	registry.Register(countingSchemaTool{name: "beta", calls: &calls})
+
+	uncached, _ := partitionTools(registry, PermissionModeAuto, Options{}, map[string]bool{})
+
+	base := calls
+	cache := map[string]zeroruntime.ToolDefinition{}
+	first, _ := partitionToolsCached(registry, PermissionModeAuto, Options{}, map[string]bool{}, cache)
+	rendersAfterFirst := calls - base
+	second, _ := partitionToolsCached(registry, PermissionModeAuto, Options{}, map[string]bool{}, cache)
+	rendersAfterSecond := calls - base
+
+	if rendersAfterFirst != 2 {
+		t.Fatalf("first cached call should render 2 tools once each, rendered %d times", rendersAfterFirst)
+	}
+	if rendersAfterSecond != 2 {
+		t.Fatalf("second cached call must reuse the cache (no new schema reads), total renders = %d", rendersAfterSecond)
+	}
+	if !reflect.DeepEqual(first, second) {
+		t.Fatal("cached calls must return identical definitions")
+	}
+	if !reflect.DeepEqual(first, uncached) {
+		t.Fatal("cached output must match the uncached partitionTools output")
+	}
+}

internal/agent/plan_staleness_test.go

@@ -0,0 +1,64 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func planUpdateTurn(status string) zeroruntime.CollectedStream {
+	return zeroruntime.CollectedStream{ToolCalls: []zeroruntime.ToolCall{
+		{Name: "update_plan", Arguments: `{"plan":[{"content":"step one","status":"` + status + `"}]}`},
+	}}
+}
+
+// Turn-based staleness: a plan that goes many turns without an update — while
+// items are still pending — draws the stale reminder even though few tool calls
+// have run (the tool-call streak alone would not have tripped).
+func TestPlanReminderFiresOnTurnStaleness(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("in_progress"))
+
+	// Text-only turns advance the turn counter without adding tool calls.
+	for range stalePlanTurnThreshold {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "still working…"})
+	}
+
+	if state.toolCallsSincePlanUpdate >= staleToolCallThreshold {
+		t.Fatalf("precondition: the tool-call trigger must not fire (%d calls)", state.toolCallsSincePlanUpdate)
+	}
+	if got := state.planReminder(stalePlanTurnThreshold + 2); !strings.Contains(got, planStaleReminderMarker) {
+		t.Fatalf("expected a turn-based stale reminder, got %q", got)
+	}
+}
+
+// A fully-completed plan is NOT stale: no pending items means the turn-based
+// trigger stays quiet even after many turns, so a finished plan isn't nagged.
+func TestPlanReminderSkipsTurnStalenessWhenComplete(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("completed"))
+
+	for range stalePlanTurnThreshold + 3 {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "wrapping up…"})
+	}
+
+	if got := state.planReminder(stalePlanTurnThreshold + 5); got != "" {
+		t.Fatalf("a completed plan must not draw a stale reminder, got %q", got)
+	}
+}
+
+// A plan update resets the turn counter, so the stale reminder does not fire
+// right after a fresh update.
+func TestPlanUpdateResetsTurnStaleness(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("in_progress"))
+	for range stalePlanTurnThreshold - 1 {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "working…"})
+	}
+	// Refresh the plan — resets the turn streak.
+	state.observeTurn(planUpdateTurn("in_progress"))
+	if got := state.planReminder(stalePlanTurnThreshold + 1); got != "" {
+		t.Fatalf("a just-refreshed plan must not be stale, got %q", got)
+	}
+}

internal/agent/reconnect.go

@@ -3,9 +3,11 @@ package agent
 import (
 	"context"
 	"fmt"
+	"math/rand"
 	"strings"
 	"time"
 
+	"github.com/Gitlawb/zero/internal/errhint"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -17,10 +19,20 @@ import (
 // times. We retry ONLY the connect (not a partially-consumed stream), so no
 // already-forwarded OnText is ever duplicated.
 const (
-	maxStreamReconnects = 2
-	streamReconnectBase = 500 * time.Millisecond
+	// maxStreamReconnects is how many times the connect is re-issued after a
+	// transient disconnect. 4 (not 2): with jittered exponential backoff this rides
+	// out a multi-second network blip (~0.5s + 1s + 2s + 4s worst case) instead of
+	// dying on a 2s hiccup and re-burning every token on a restart.
+	maxStreamReconnects = 4
+	// streamReconnectMax caps a single backoff so the tail attempts don't wait
+	// minutes on a long outage.
+	streamReconnectMax = 8 * time.Second
 )
 
+// streamReconnectBase is the first-retry delay (doubled each subsequent attempt).
+// A var, not a const, so tests can shrink it to keep the exhaustion path fast.
+var streamReconnectBase = 500 * time.Millisecond
+
 // reconnectNotifier is called before each retry with the 1-based attempt number
 // and the max, so the caller can surface a "Reconnecting N/max" notice. Nil is
 // fine.
@@ -73,7 +85,7 @@ func streamWithReconnect(ctx context.Context, provider Provider, request zerorun
 		if notify != nil {
 			notify(attempt, maxStreamReconnects)
 		}
-		if waitErr := sleepWithContext(ctx, backoffFor(attempt)); waitErr != nil {
+		if waitErr := sleepWithContext(ctx, jitteredBackoff(attempt)); waitErr != nil {
 			return nil, err // ctx cancelled while waiting; surface the original error
 		}
 		stream, err = provider.StreamCompletion(ctx, request)
@@ -96,6 +108,19 @@ func shouldReconnect(ctx context.Context, err error) bool {
 	if isContextLimitError(msg) || isImageRejectionError(err) {
 		return false
 	}
+	// HTTP 5xx statuses are non-reconnectable and must be excluded BEFORE the
+	// transport substring match below — otherwise "504 Gateway Timeout" would slip
+	// through on the generic "timeout" needle. 503 already exhausted
+	// providerio.SendWithRetry (retrying is a redundant double-retry); 500/502/504
+	// are non-idempotent by providerio's rule — the completion POST may already
+	// have reached the model before the upstream/gateway gave up, so replaying the
+	// connect risks duplicate billable work. Digit-boundary matched so an
+	// incidental "504" inside a latency/id number is not mistaken for a status.
+	if errhint.HasStatusCode(msg, "500", "502", "503", "504") {
+		return false
+	}
+	// Transport-level disconnects only. A genuine transport failure (EOF, reset,
+	// refused, timeout) means no response was received, which is safe to reconnect.
 	for _, needle := range []string{
 		"eof",
 		"connection reset",
@@ -106,8 +131,6 @@ func shouldReconnect(ctx context.Context, err error) bool {
 		"timed out",
 		"temporarily unavailable",
 		"i/o timeout",
-		"503",
-		"502",
 		"server closed",
 		"unexpected end",
 	} {
@@ -118,14 +141,31 @@ func shouldReconnect(ctx context.Context, err error) bool {
 	return false
 }
 
+// backoffFor is the deterministic exponential base delay for a 1-based attempt,
+// capped at streamReconnectMax. Jitter is layered on separately (jitteredBackoff).
 func backoffFor(attempt int) time.Duration {
 	d := streamReconnectBase
 	for i := 1; i < attempt; i++ {
+		if d >= streamReconnectMax {
+			return streamReconnectMax
+		}
 		d *= 2
 	}
+	if d > streamReconnectMax {
+		d = streamReconnectMax
+	}
 	return d
 }
 
+// jitteredBackoff adds up to 50% random jitter on top of backoffFor so concurrent
+// runs (swarm members, a cron fleet) that all trip on the same outage don't
+// reconnect in lockstep and hammer a recovering endpoint. Never shorter than the
+// deterministic base, so backoff still grows attempt over attempt.
+func jitteredBackoff(attempt int) time.Duration {
+	base := backoffFor(attempt)
+	return base + time.Duration(rand.Int63n(int64(base/2)+1))
+}
+
 func sleepWithContext(ctx context.Context, d time.Duration) error {
 	timer := time.NewTimer(d)
 	defer timer.Stop()

internal/agent/reconnect_test.go

@@ -44,6 +44,10 @@ func TestStreamWithReconnectRecoversFromTransientDisconnect(t *testing.T) {
 }
 
 func TestStreamWithReconnectGivesUpAfterMax(t *testing.T) {
+	// Shrink the backoff so exhausting all retries stays fast (real base would
+	// sleep ~7.5s across 4 attempts).
+	defer func(orig time.Duration) { streamReconnectBase = orig }(streamReconnectBase)
+	streamReconnectBase = time.Millisecond
 	// Always fails with a disconnect error → exhausts retries and returns it.
 	p := &flakyProvider{failBefore: 99, failErr: errors.New("connection reset by peer")}
 	_, err := streamWithReconnect(context.Background(), p, zeroruntime.CompletionRequest{}, nil)
@@ -87,7 +91,7 @@ func TestShouldReconnectClassification(t *testing.T) {
 	ctx := context.Background()
 	disconnects := []string{
 		"unexpected EOF", "connection reset by peer", "broken pipe",
-		"i/o timeout", "503 Service Unavailable", "server closed the connection",
+		"i/o timeout", "server closed the connection", "connection refused",
 	}
 	for _, m := range disconnects {
 		if !shouldReconnect(ctx, errors.New(m)) {
@@ -97,6 +101,13 @@ func TestShouldReconnectClassification(t *testing.T) {
 	notDisconnects := []string{
 		"context length exceeded", "invalid api key", "model not found",
 		"400 bad request: unsupported parameter",
+		// HTTP 5xx statuses are handled by providerio.SendWithRetry (503) or are
+		// non-idempotent (500/502/504); the reconnect path must not double-retry
+		// them. "504 Gateway Timeout" must NOT slip through on the generic "timeout"
+		// transport signal — a gateway timeout is non-idempotent, since the upstream
+		// may have already processed the completion POST before giving up.
+		"503 Service Unavailable", "502 Bad Gateway",
+		"504 Gateway Timeout", "500 Internal Server Error",
 	}
 	for _, m := range notDisconnects {
 		if shouldReconnect(ctx, errors.New(m)) {
@@ -112,6 +123,24 @@ func TestBackoffGrows(t *testing.T) {
 	if backoffFor(2) != 2*streamReconnectBase {
 		t.Fatalf("attempt 2 backoff = %v, want %v", backoffFor(2), 2*streamReconnectBase)
 	}
+	// The exponential base is capped so late attempts don't wait minutes.
+	if got := backoffFor(20); got != streamReconnectMax {
+		t.Fatalf("attempt 20 backoff = %v, want cap %v", got, streamReconnectMax)
+	}
+}
+
+func TestJitteredBackoffStaysInBounds(t *testing.T) {
+	// Jitter never drops below the deterministic base and never exceeds base*1.5,
+	// so backoff still grows attempt over attempt while decorrelating retries.
+	for attempt := 1; attempt <= 5; attempt++ {
+		base := backoffFor(attempt)
+		for i := 0; i < 200; i++ {
+			got := jitteredBackoff(attempt)
+			if got < base || got > base+base/2 {
+				t.Fatalf("attempt %d jittered backoff %v out of [%v, %v]", attempt, got, base, base+base/2)
+			}
+		}
+	}
 }
 
 func TestReconnectNoticeRoutesThroughReasoning(t *testing.T) {

internal/agent/skills_context_test.go

@@ -7,14 +7,18 @@ import (
 )
 
 func TestSkillsContextCapsLongList(t *testing.T) {
-	skills := make([]SkillInfo, 0, 40)
-	for i := 0; i < 40; i++ {
+	// 60 skills with realistic trigger-rich descriptions (~180 chars each, under
+	// the 200-rune truncation) total well past the 4096-byte list budget, so the
+	// overflow summary must kick in while the block stays bounded.
+	longDesc := strings.Repeat("use when the request touches deployments, release notes, or verification; ", 2) + "number "
+	skills := make([]SkillInfo, 0, 60)
+	for i := 0; i < 60; i++ {
 		n := strconv.Itoa(i)
-		skills = append(skills, SkillInfo{Name: "skill-" + n, Description: "does something useful, number " + n})
+		skills = append(skills, SkillInfo{Name: "skill-" + n, Description: longDesc + n})
 	}
 	got := skillsContext(Options{Skills: skills})
-	if len(got) > 1200 {
-		t.Fatalf("skills block should stay bounded, got %d bytes:\n%s", len(got), got)
+	if len(got) > skillsContextListBudget+1000 {
+		t.Fatalf("skills block should stay bounded near the %d-byte budget, got %d bytes:\n%s", skillsContextListBudget, len(got), got)
 	}
 	if !strings.Contains(got, "more (call skill") {
 		t.Fatalf("expected an overflow summary line, got:\n%s", got)
@@ -25,6 +29,27 @@ func TestSkillsContextCapsLongList(t *testing.T) {
 	}
 }
 
+// A realistic mid-size skill set (20 skills, trigger-rich descriptions) must be
+// listed in FULL — no overflow summary. This pins the fix for the old 640-byte
+// budget, under which skills past ~#6 were invisible to the model and therefore
+// never triggered.
+func TestSkillsContextListsRealisticSetInFull(t *testing.T) {
+	desc := "Use when the user asks about deployments, release notes, or pre-merge verification runs."
+	skills := make([]SkillInfo, 0, 20)
+	for i := 0; i < 20; i++ {
+		skills = append(skills, SkillInfo{Name: "skill-" + strconv.Itoa(i), Description: desc})
+	}
+	got := skillsContext(Options{Skills: skills})
+	if strings.Contains(got, "more (call skill") {
+		t.Fatalf("20 described skills must all be listed without overflow, got:\n%s", got)
+	}
+	for i := 0; i < 20; i++ {
+		if !strings.Contains(got, "- skill-"+strconv.Itoa(i)+":") {
+			t.Fatalf("skill-%d missing from the list:\n%s", i, got)
+		}
+	}
+}
+
 func TestSkillsContext(t *testing.T) {
 	if got := skillsContext(Options{}); got != "" {
 		t.Fatalf("no skills should yield an empty section, got %q", got)

internal/agent/system_prompt.go

@@ -9,6 +9,7 @@ import (
 	"strings"
 	"unicode/utf8"
 
+	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/repomap"
 	"github.com/Gitlawb/zero/internal/workspaceseed"
 )
@@ -38,6 +39,14 @@ const fallbackSystemPrompt = "You are Zero, a terminal coding agent. Help with t
 // general-to-specific order.
 var projectContextFiles = []string{"AGENTS.md", "ZERO.md", ".zero/AGENTS.md"}
 
+// userContextFile is the per-user instruction file, resolved under
+// config.UserConfigDir()/zero/ alongside the rest of Zero's per-user config
+// (config.json, commands, specialists) so users can keep personal guidance out
+// of individual repositories.
+const userContextFile = "ZERO.md"
+
+var userConfigDirForPrompt = config.UserConfigDir
+
 // maxProjectContextBytes caps how much of a single project doc is injected so
 // a large guidelines file can't blow the context budget.
 const maxProjectContextBytes = 8 << 10 // 8 KiB per file
@@ -81,6 +90,13 @@ func buildSystemPrompt(options Options) string {
 	if seed := workspaceSeedContext(options.Cwd); seed != "" {
 		sections = append(sections, seed)
 	}
+	// User guidelines are injected before workspace/project guidelines so the
+	// project's AGENTS.md/ZERO.md is the later, more specific instruction
+	// block. See userGuidelines for the explicit precedence note carried in
+	// the section text itself.
+	if user := userGuidelines(); user != "" {
+		sections = append(sections, user)
+	}
 	if ws := workspaceContext(options.Cwd); ws != "" {
 		sections = append(sections, ws)
 	}
@@ -175,10 +191,19 @@ func specialistDelegationContext(options Options) string {
 // when no skills are installed, so a skill-less run reproduces the previous prompt
 // byte-for-byte.
 // skillsContextListBudget bounds the bytes spent listing individual skills so a
-// workspace with dozens of them can't bloat every turn's prompt. Skills past the
-// budget are summarized as a count rather than dropped — the model can still load
-// any of them by name (an unknown name returns the full list from the skill tool).
-const skillsContextListBudget = 640
+// workspace with an extreme number of them can't bloat every turn's prompt.
+// Skills past the budget are summarized as a count rather than dropped — the
+// model can still load any of them by name (an unknown name returns the full
+// list from the skill tool).
+//
+// The budget is deliberately generous — roughly 18 skills with maximally long
+// (200-rune, ASCII) descriptions, or 40+ with typical terser ones; non-ASCII
+// descriptions consume it faster since the budget counts bytes. This list is
+// the model's ONLY discovery surface, so a skill squeezed out of it effectively
+// never triggers. The previous 640-byte cap silently broke invocation for
+// anyone with more than ~6 skills — the model cannot match a request against a
+// skill it cannot see.
+const skillsContextListBudget = 4096
 
 func skillsContext(options Options) string {
 	if len(options.Skills) == 0 {
@@ -186,7 +211,7 @@ func skillsContext(options Options) string {
 	}
 	var b strings.Builder
 	b.WriteString("<available_skills>\n")
-	b.WriteString("Reusable, on-demand instruction sets you can load with the skill tool. When a request matches one, call skill with its exact name to pull in its guidance before acting — do not guess names or skip a relevant skill.\n")
+	b.WriteString("Reusable, on-demand instruction sets you can load with the skill tool. Before acting on a request, scan this list; when the request matches a skill's name or description, call skill with its exact name FIRST and follow the loaded guidance — do not guess names, do not skip a matching skill, and do not substitute your own approach for its instructions.\n")
 	listed, spent, omitted := 0, 0, 0
 	for _, info := range options.Skills {
 		name := strings.TrimSpace(info.Name)
@@ -216,9 +241,11 @@ func skillsContext(options Options) string {
 }
 
 // truncateForSkillLine keeps a skill's one-line description short so a single
-// verbose description can't dominate the skills-list budget.
+// verbose description can't dominate the skills-list budget. The cap is roomy on
+// purpose: the description carries the skill's trigger conditions ("use when…"),
+// and truncating those is what stops the model from ever invoking the skill.
 func truncateForSkillLine(desc string) string {
-	const maxDescRunes = 100
+	const maxDescRunes = 200
 	runes := []rune(desc)
 	if len(runes) <= maxDescRunes {
 		return desc
@@ -259,7 +286,7 @@ func workspaceContext(cwd string) string {
 	b.WriteString("Working directory: " + cwd + "\n")
 	b.WriteString("Operating system: " + runtime.GOOS + "\n")
 	if runtime.GOOS == "windows" {
-		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
+		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories. Do not pipe to or invoke POSIX coreutils from Git for Windows (usr\\bin head/grep/tail/cat/...): they are MSYS binaries and fail under the write-restricted sandbox; use native Zero tools (grep, read_file, list_directory, glob) or cmd.exe findstr/more instead, or sandbox_permissions require_escalated only when host-level execution is truly required.\n")
 	} else {
 		b.WriteString("Shell syntax: /bin/sh syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
 	}
@@ -312,13 +339,7 @@ func projectGuidelines(cwd, gitRoot string) string {
 		if remaining := maxProjectContextTotalBytes - totalUsed; remaining < limit {
 			limit = remaining
 		}
-		if len(content) > limit {
-			cut := limit
-			for cut > 0 && !utf8.RuneStart(content[cut]) {
-				cut--
-			}
-			content = content[:cut] + "\n… (truncated)"
-		}
+		content = truncateGuidelineContent(content, limit)
 		label := projectGuidelineLabel(match, gitRoot)
 		b.WriteString("\n\n## Project guidelines (" + label + ")\n\n" + content)
 		totalUsed += len(content)
@@ -326,6 +347,92 @@ func projectGuidelines(cwd, gitRoot string) string {
 	return b.String()
 }
 
+// userGuidelines returns the per-user ZERO.md instructions block, if present.
+// The file lives in config.UserConfigDir()/zero/ next to Zero's other
+// per-user config; the basename match is case-insensitive so a file saved as
+// zero.md still resolves on case-sensitive filesystems, mirroring the project
+// guideline loader. The section carries an explicit precedence note because
+// this is a global, personal preferences file: it is injected earlier in the
+// prompt than the project's AGENTS.md/ZERO.md (see buildSystemPrompt), and
+// the note keeps that precedence unambiguous even if a model weighs later
+// context more heavily than section order alone implies.
+func userGuidelines() string {
+	configDir, err := userConfigDirForPrompt()
+	if err != nil {
+		return ""
+	}
+	configDir = strings.TrimSpace(configDir)
+	if configDir == "" {
+		return ""
+	}
+	match := findCaseInsensitiveFile(filepath.Join(configDir, "zero"), userContextFile)
+	if match == "" {
+		return ""
+	}
+	data, err := os.ReadFile(match)
+	if err != nil {
+		return ""
+	}
+	content := strings.TrimSpace(string(data))
+	if content == "" {
+		return ""
+	}
+	content = truncateGuidelineContent(content, maxProjectContextBytes)
+	return "## User guidelines (" + filepath.Base(match) + ")\n\n" +
+		"These are the operator's personal preferences, not project policy. " +
+		"Where they conflict with a repository's project guidelines below (AGENTS.md/ZERO.md), the project guidelines take precedence.\n\n" +
+		content
+}
+
+// truncationMarker is appended to guideline content that was cut short.
+const truncationMarker = "\n… (truncated)"
+
+// truncateGuidelineContent caps content at limit bytes without splitting a
+// UTF-8 rune, appending a truncation marker when anything was cut. Space for
+// the marker is reserved before choosing the cut point. When limit is too
+// small to fit the marker itself, the marker is dropped instead, so the
+// returned string never exceeds limit for any non-negative limit.
+func truncateGuidelineContent(content string, limit int) string {
+	if len(content) <= limit {
+		return content
+	}
+	if limit <= 0 {
+		return ""
+	}
+	cut := limit - len(truncationMarker)
+	if cut < 0 {
+		cut = 0
+	}
+	for cut > 0 && !utf8.RuneStart(content[cut]) {
+		cut--
+	}
+	if truncated := content[:cut] + truncationMarker; len(truncated) <= limit {
+		return truncated
+	}
+	// limit is smaller than the marker itself: fall back to a hard,
+	// rune-safe cut at limit bytes with no marker rather than exceeding it.
+	end := limit
+	for end > 0 && !utf8.RuneStart(content[end]) {
+		end--
+	}
+	return content[:end]
+}
+
+// findCaseInsensitiveFile returns the on-disk path of the regular file in dir
+// whose name matches basename case-insensitively, or "" when absent.
+func findCaseInsensitiveFile(dir, basename string) string {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return ""
+	}
+	for _, e := range entries {
+		if !e.IsDir() && strings.EqualFold(e.Name(), basename) {
+			return filepath.Join(dir, e.Name())
+		}
+	}
+	return ""
+}
+
 // projectGuidelineDirs returns the directory chain from gitRoot to cwd
 // (inclusive), in root-to-leaf order. If gitRoot is empty or unreachable
 // from cwd, the chain collapses to [cwd].
@@ -378,23 +485,16 @@ func projectGuidelineLabel(match, gitRoot string) string {
 // Returns "" when nothing matches.
 func findProjectContextFile(dir string) string {
 	for _, name := range projectContextFiles {
-		baseLower := strings.ToLower(filepath.Base(name))
 		// Walk to the file's parent through dir with case-insensitive segment
-		// matching, then find a regular file with the same lowercased
-		// basename. This works on both case-sensitive and case-insensitive
-		// filesystems and always returns the on-disk filename.
+		// matching, then find a regular file with the same basename. This works
+		// on both case-sensitive and case-insensitive filesystems and always
+		// returns the on-disk filename.
 		parent, ok := resolveDirCaseInsensitive(filepath.Dir(filepath.Join(dir, name)), dir)
 		if !ok {
 			continue
 		}
-		entries, err := os.ReadDir(parent)
-		if err != nil {
-			continue
-		}
-		for _, e := range entries {
-			if !e.IsDir() && strings.ToLower(e.Name()) == baseLower {
-				return filepath.Join(parent, e.Name())
-			}
+		if match := findCaseInsensitiveFile(parent, filepath.Base(name)); match != "" {
+			return match
 		}
 	}
 	return ""

internal/agent/system_prompt.md

@@ -63,6 +63,10 @@ work.
   targeted diffs. 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.
+- 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.
 - 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.
 

internal/agent/system_prompt_test.go

@@ -9,6 +9,25 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 )
 
+// TestMain points userConfigDirForPrompt at an empty, package-wide temp
+// directory for every test in this package. Without this, buildSystemPrompt
+// falls back to the real config.UserConfigDir, so any developer with a
+// personal ~/.config/zero/ZERO.md would get its content folded into
+// otherwise-unrelated prompt assertions, making test runs non-deterministic
+// across contributor machines. Tests that specifically exercise user
+// guidelines stub userConfigDirForPrompt themselves via
+// withSystemPromptTestUserConfigDir(Func) and restore this default on cleanup.
+func TestMain(m *testing.M) {
+	dir, err := os.MkdirTemp("", "zero-agent-tests-*")
+	if err != nil {
+		panic(err)
+	}
+	userConfigDirForPrompt = func() (string, error) { return dir, nil }
+	code := m.Run()
+	os.RemoveAll(dir)
+	os.Exit(code)
+}
+
 func TestCoreSystemPromptIncludesCodingQualityRules(t *testing.T) {
 	prompt := strings.ToLower(buildSystemPrompt(Options{}))
 
@@ -128,6 +147,88 @@ func systemPromptTestBlock(t *testing.T, prompt, start, end string) string {
 	return body
 }
 
+func TestBuildSystemPromptIncludesUserGuidelines(t *testing.T) {
+	configDir := t.TempDir()
+	writeSystemPromptTestFile(t, configDir, "zero/ZERO.md", "  Prefer concise summaries.  \n")
+	t.Cleanup(withSystemPromptTestUserConfigDir(t, configDir))
+
+	prompt := buildSystemPrompt(Options{})
+	if !strings.Contains(prompt, "## User guidelines (ZERO.md)") {
+		t.Fatalf("expected user guidelines header, got:\n%s", prompt)
+	}
+	if !strings.Contains(prompt, "Prefer concise summaries.") {
+		t.Fatalf("expected user guidelines content, got:\n%s", prompt)
+	}
+	if strings.Contains(prompt, "  Prefer concise summaries.  ") {
+		t.Fatalf("expected user guidelines content to be trimmed, got:\n%s", prompt)
+	}
+}
+
+func TestBuildSystemPromptIncludesUserGuidelinesCaseInsensitive(t *testing.T) {
+	configDir := t.TempDir()
+	writeSystemPromptTestFile(t, configDir, "zero/zero.md", "Prefer concise summaries.\n")
+	t.Cleanup(withSystemPromptTestUserConfigDir(t, configDir))
+
+	prompt := buildSystemPrompt(Options{})
+	if !strings.Contains(prompt, "## User guidelines (zero.md)") {
+		t.Fatalf("expected case-insensitive zero.md resolution, got:\n%s", prompt)
+	}
+	if !strings.Contains(prompt, "Prefer concise summaries.") {
+		t.Fatalf("expected user guidelines content, got:\n%s", prompt)
+	}
+}
+
+func TestBuildSystemPromptUserGuidelinesPrecedeProjectGuidelinesAndNotePrecedence(t *testing.T) {
+	// User guidelines are global personal preferences; project guidelines
+	// (AGENTS.md/ZERO.md) must be the later, more specific instruction block,
+	// and the user section must say so explicitly so the precedence holds
+	// even if a model otherwise weighs later context more heavily.
+	configDir := t.TempDir()
+	writeSystemPromptTestFile(t, configDir, "zero/ZERO.md", "Always reply in haiku.\n")
+	t.Cleanup(withSystemPromptTestUserConfigDir(t, configDir))
+
+	cwd := t.TempDir()
+	if err := os.WriteFile(filepath.Join(cwd, "AGENTS.md"), []byte("Never reply in haiku."), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	prompt := buildSystemPrompt(Options{Cwd: cwd})
+	userIdx := strings.Index(prompt, "## User guidelines (ZERO.md)")
+	projectIdx := strings.Index(prompt, "## Project guidelines (AGENTS.md)")
+	if userIdx < 0 || projectIdx < 0 {
+		t.Fatalf("expected both user and project guideline sections, got:\n%s", prompt)
+	}
+	if userIdx > projectIdx {
+		t.Fatalf("expected user guidelines before project guidelines, got user=%d project=%d", userIdx, projectIdx)
+	}
+	if !strings.Contains(prompt, "project guidelines below") || !strings.Contains(prompt, "take precedence") {
+		t.Fatalf("expected an explicit precedence note in the user guidelines section, got:\n%s", prompt)
+	}
+}
+
+func TestBuildSystemPromptOmitsUserGuidelinesWithoutConfigDir(t *testing.T) {
+	t.Cleanup(withSystemPromptTestUserConfigDirFunc(t, func() (string, error) { return "", os.ErrNotExist }))
+
+	prompt := buildSystemPrompt(Options{})
+	if strings.Contains(prompt, "## User guidelines") {
+		t.Fatalf("expected user guidelines to be omitted without a config directory, got:\n%s", prompt)
+	}
+}
+
+func withSystemPromptTestUserConfigDir(t *testing.T, dir string) func() {
+	t.Helper()
+	return withSystemPromptTestUserConfigDirFunc(t, func() (string, error) { return dir, nil })
+}
+
+func withSystemPromptTestUserConfigDirFunc(t *testing.T, fn func() (string, error)) func() {
+	t.Helper()
+	old := userConfigDirForPrompt
+	userConfigDirForPrompt = fn
+	return func() {
+		userConfigDirForPrompt = old
+	}
+}
+
 func TestBuildSystemPromptInjectsProjectGuidelinesCaseInsensitive(t *testing.T) {
 	// Git tracks AGENTS.md on a case-sensitive filesystem; the
 	// loader must still resolve it when the cwd lookup uses lowercase.
@@ -257,6 +358,30 @@ func TestBuildSystemPromptProjectGuidelinesTruncatesAtTotalCap(t *testing.T) {
 	}
 }
 
+func TestTruncateGuidelineContentStaysWithinLimit(t *testing.T) {
+	limit := 64
+	content := strings.Repeat("x", limit+1)
+	got := truncateGuidelineContent(content, limit)
+	if len(got) > limit {
+		t.Fatalf("truncated result is %d bytes, want at most %d", len(got), limit)
+	}
+	if !strings.Contains(got, "truncated") {
+		t.Fatalf("expected truncation marker, got %q", got)
+	}
+}
+
+func TestTruncateGuidelineContentStaysWithinLimitForSmallBudgets(t *testing.T) {
+	// Limits at or below the truncation marker's own length can't fit the
+	// marker at all; the helper must still never exceed the requested limit.
+	content := strings.Repeat("x", 100)
+	for _, limit := range []int{0, 1, len(truncationMarker) - 1, len(truncationMarker), len(truncationMarker) + 1} {
+		got := truncateGuidelineContent(content, limit)
+		if len(got) > limit {
+			t.Fatalf("truncateGuidelineContent(_, %d) = %q (%d bytes), want at most %d bytes", limit, got, len(got), limit)
+		}
+	}
+}
+
 func TestProjectGuidelineDirsOrdersRootToLeaf(t *testing.T) {
 	root := filepath.Join("r")
 	leaf := filepath.Join(root, "a", "b")

internal/agent/types.go

@@ -285,6 +285,14 @@ type Options struct {
 	// ceiling and the autonomy gate. nil disables it entirely (the loop is
 	// byte-identical to before). One instance per run — it holds attempt state.
 	SelfCorrect *SelfCorrector
+	// FileDiagnostics, when set, checks files changed by mutating tools for
+	// error-severity language diagnostics IN THE BACKGROUND and appends any
+	// errors as a nudge before the model's next request — the model still sees
+	// an error it introduced at its next decision point, but no tool call ever
+	// blocks on the language server (the old inline path stalled every edit on
+	// a ≥300ms debounce, 10s cap). Build one with NewFileDiagnostics. nil
+	// disables post-edit diagnostics.
+	FileDiagnostics func(ctx context.Context, absPath string) string
 
 	// RequireCompletionSignal gates run completion for HEADLESS exec. Without it,
 	// any assistant turn that produces text but no tool call is accepted as the

internal/cli/app.go

@@ -12,6 +12,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/agent"
@@ -19,6 +20,7 @@ import (
 	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/localcontrol"
 	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/modelregistry"
 	"github.com/Gitlawb/zero/internal/observability"
 	"github.com/Gitlawb/zero/internal/plugins"
 	"github.com/Gitlawb/zero/internal/providerhealth"
@@ -77,6 +79,8 @@ type appDeps struct {
 	runAgentEval           func(context.Context, agentEvalOptions) (agentEvalReport, error)
 	inspectChanges         func(context.Context, zerogit.InspectOptions) (zerogit.ChangeSummary, error)
 	commitChanges          func(context.Context, zerogit.CommitOptions) (zerogit.CommitResult, error)
+	pushChanges            func(context.Context, zerogit.PushOptions) (zerogit.PushResult, error)
+	createPR               func(context.Context, zerogit.PROptions) (zerogit.PRResult, error)
 	runTUI                 func(context.Context, tui.Options) int
 	runEditor              func(string) error
 	checkUpdate            func(context.Context, update.Options) (update.Result, error)
@@ -178,6 +182,8 @@ func defaultAppDeps() appDeps {
 		runAgentEval:     defaultRunAgentEval,
 		inspectChanges:   zerogit.Inspect,
 		commitChanges:    zerogit.Commit,
+		pushChanges:      zerogit.Push,
+		createPR:         zerogit.CreatePR,
 		runTUI:           tui.Run,
 		runEditor:        openEditor,
 		checkUpdate:      update.Check,
@@ -229,6 +235,12 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	defer observability.Recover(observability.DefaultCrashDir(), "cli", stderr, &exitCode)
 	deps = fillAppDeps(deps)
 
+	// CLI runs opt into the models.dev overlay (cached live context limits and
+	// pricing on top of the curated catalog). Explicitly enabled here — and only
+	// here — so library consumers and hermetic tests are never perturbed by a
+	// cache file on the machine. The refresh itself is fired in exec/TUI startup.
+	modelregistry.EnableModelsDevOverlay()
+
 	addDirs, args, err := splitLeadingAddDirFlags(args)
 	if err != nil {
 		return writeAppError(stderr, err.Error(), 1)
@@ -510,6 +522,12 @@ func fillAppDeps(deps appDeps) appDeps {
 	if deps.commitChanges == nil {
 		deps.commitChanges = defaults.commitChanges
 	}
+	if deps.pushChanges == nil {
+		deps.pushChanges = defaults.pushChanges
+	}
+	if deps.createPR == nil {
+		deps.createPR = defaults.createPR
+	}
 	if deps.runTUI == nil {
 		deps.runTUI = defaults.runTUI
 	}
@@ -544,6 +562,11 @@ func runInteractiveTUI(stderr io.Writer, deps appDeps, permissionMode agent.Perm
 }
 
 func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode agent.PermissionMode, addDirs []string, theme string, forceSetup bool) int {
+	// Refresh the models.dev pricing/limits cache in the background when stale;
+	// the overlay is read at registry construction from the cache file, so this
+	// benefits the next run and never blocks or fails this one.
+	go func() { _ = modelregistry.RefreshModelsDevCache(context.Background()) }()
+
 	workspaceRoot, err := deps.getwd()
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve workspace: "+err.Error(), 1)
@@ -760,6 +783,14 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			Specialists:    specialistRuntime.specialists,
 			Skills:         pluginActivation.skillInfos(deps.skillsDir()),
 		},
+		// LoadSkills backs /skills and direct /<skill-name> invocation in the TUI.
+		// It resolves against the same merged set (default dir + plugin skill
+		// roots) as the skill tool and the system-prompt list, re-read per use so
+		// newly installed skills work without a restart.
+		LoadSkills: cachedSkillsLoader(func() []skills.Skill {
+			merged, _ := plugins.MergedSkillsLoaded(deps.skillsDir(), pluginActivation.skillRoots)
+			return merged
+		}),
 		PermissionMode: permissionMode,
 		Notify:         resolved.Notify,
 		KeyBindings:    resolved.KeyBindings,
@@ -1255,3 +1286,26 @@ Flags:
 `)
 	return err
 }
+
+// cachedSkillsLoader memoizes a skills loader for a short interval. The TUI
+// calls the loader from autocomplete on every "/x" keystroke; without a cache
+// each keystroke would re-read every SKILL.md body. Two seconds is fresh enough
+// that a newly installed skill still shows up "immediately" while typing.
+func cachedSkillsLoader(load func() []skills.Skill) func() []skills.Skill {
+	var (
+		mu     sync.Mutex
+		at     time.Time
+		cached []skills.Skill
+	)
+	const ttl = 2 * time.Second
+	return func() []skills.Skill {
+		mu.Lock()
+		defer mu.Unlock()
+		if cached != nil && time.Since(at) < ttl {
+			return cached
+		}
+		cached = load()
+		at = time.Now()
+		return cached
+	}
+}

internal/cli/exec.go

@@ -12,6 +12,7 @@ import (
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/errhint"
 	"github.com/Gitlawb/zero/internal/imageinput"
 	"github.com/Gitlawb/zero/internal/lsp"
 	"github.com/Gitlawb/zero/internal/modelregistry"
@@ -135,6 +136,11 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		return exitSuccess
 	}
 
+	// Refresh the models.dev pricing/limits cache in the background when stale;
+	// the overlay is read at registry construction from the cache file, so this
+	// benefits the next run and never blocks or fails this one.
+	go func() { _ = modelregistry.RefreshModelsDevCache(context.Background()) }()
+
 	// A mode seeds model/effort/max-turns/tool filters as a preset. Expand it up
 	// front — before tool-filter validation and the --list-tools branch — so a
 	// mode-injected tool filter is validated and reflected in --list-tools, and a
@@ -499,9 +505,10 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	// by default the corrector is nil, leaving the agent loop byte-identical. When
 	// on we verify with both the workspace test plan and LSP diagnostics over the
 	// changed files; the autonomy gate inside the corrector still decides whether
-	// failures auto-fix or just report. lspShutdown tears down any language-server
-	// sessions the LSP half spawned (no-op when self-correct is off).
-	selfCorrector, lspShutdown := newExecSelfCorrector(options.selfCorrect, workspaceRoot, options.autonomy)
+	// failures auto-fix or just report. fileDiagnostics (always on, lazy) gives
+	// edit_file/write_file inline error diagnostics for the file they just wrote.
+	// lspShutdown tears down any language-server sessions either half spawned.
+	selfCorrector, fileDiagnostics, lspShutdown := newExecSelfCorrector(options.selfCorrect, workspaceRoot, options.autonomy)
 	defer lspShutdown()
 	result, err := agent.Run(runCtx, agentPrompt, provider, agent.Options{
 		MaxTurns:         resolved.MaxTurns,
@@ -525,6 +532,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		PermissionMode:   permissionMode,
 		Autonomy:         options.autonomy,
 		SelfCorrect:      selfCorrector,
+		FileDiagnostics:  fileDiagnostics,
 		// Headless exec: don't accept a no-tool-call turn as "done" while work
 		// clearly remains (pending plan items / a mid-step continuation cue) —
 		// nudge to continue, and finalize as INCOMPLETE rather than false success
@@ -536,6 +544,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		EnabledTools:            options.enabledTools,
 		DisabledTools:           options.disabledTools,
 		OnText:                  writer.text,
+		OnReasoning:             writer.reasoning,
 		OnToolCall: func(call agent.ToolCall) {
 			writer.toolCall(call, registry)
 			sessionRecorder.append(sessions.EventToolCall, map[string]any{
@@ -700,23 +709,29 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 // unless a changed file's language actually has one installed on PATH. The
 // returned cleanup shuts that manager down (terminating any spawned server
 // sessions); it is a no-op when self-correct is off.
-func newExecSelfCorrector(enabled bool, workspaceRoot string, autonomy string) (*agent.SelfCorrector, func()) {
-	if !enabled {
-		return nil, func() {}
-	}
+func newExecSelfCorrector(enabled bool, workspaceRoot string, autonomy string) (*agent.SelfCorrector, func(context.Context, string) string, func()) {
+	// The manager is created regardless of --self-correct: it also backs the
+	// always-on background post-edit diagnostics (agent.NewFileDiagnostics,
+	// collected off the tool-call path by the loop), and it stays lazy — no
+	// language server is spawned unless an edited file's language actually has
+	// one installed on PATH.
 	manager := lsp.NewManager(workspaceRoot)
+	cleanup := func() {
+		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+		_ = manager.Shutdown(ctx)
+	}
+	fileDiagnostics := agent.NewFileDiagnostics(manager, workspaceRoot)
+	if !enabled {
+		return nil, fileDiagnostics, cleanup
+	}
 	corrector := agent.NewSelfCorrector(workspaceRoot, agent.NewLSPDiagnosticsChecker(manager), agent.NewProjectVerifier(workspaceRoot), agent.SelfCorrectConfig{
 		Enabled:      true,
 		IncludeTests: true,
 		IncludeLSP:   true,
 		Autonomy:     autonomy,
 	})
-	cleanup := func() {
-		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
-		defer cancel()
-		_ = manager.Shutdown(ctx)
-	}
-	return corrector, cleanup
+	return corrector, fileDiagnostics, cleanup
 }
 
 func deferredEligibleCount(registry *tools.Registry, permissionMode agent.PermissionMode, enabledTools []string, disabledTools []string) int {
@@ -949,6 +964,14 @@ func writeExecProviderError(stdout io.Writer, stderr io.Writer, format execOutpu
 	if _, err := fmt.Fprintf(stderr, "[zero] %s\n", message); err != nil {
 		return exitCrash
 	}
+	// Append a one-line next step for recognized provider failures (auth /
+	// rate-limit / connectivity / …). The classifier gates on a provider-origin
+	// marker, so non-provider codes (sandbox_error, mcp_error) never draw a hint.
+	if hint := errhint.CLIHint(errors.New(message)); hint != "" {
+		if _, err := fmt.Fprintf(stderr, "[zero] %s\n", hint); err != nil {
+			return exitCrash
+		}
+	}
 	return exitProvider
 }
 

internal/cli/exec_protocol_test.go

@@ -529,6 +529,45 @@ func TestRunExecStreamJSONRunStartUsesResolvedAPIModel(t *testing.T) {
 	}
 }
 
+func TestRunExecStreamJSONEmitsReasoningEvents(t *testing.T) {
+	t.Setenv("XDG_DATA_HOME", t.TempDir())
+	cwd := t.TempDir()
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+
+	exitCode := runWithDeps([]string{"exec", "--output-format", "stream-json", "think then answer"}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) {
+			return cwd, nil
+		},
+		resolveConfig: func(_ string, _ config.Overrides) (config.ResolvedConfig, error) {
+			return execResolvedConfig(), nil
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return reasoningExecProvider{}, nil
+		},
+	})
+
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	events := decodeJSONLines(t, stdout.String())
+	reasoningIdx, reasoning := findJSONEventIndex(t, events, "reasoning")
+	if reasoning["delta"] != "Thinking. " {
+		t.Fatalf("unexpected reasoning event: %#v", reasoning)
+	}
+	textIdx, text := findJSONEventIndex(t, events, "text")
+	if text["delta"] != "done" {
+		t.Fatalf("unexpected text event: %#v", text)
+	}
+	if reasoningIdx >= textIdx {
+		t.Fatalf("expected reasoning event before text event, got indices %d and %d", reasoningIdx, textIdx)
+	}
+	final := findJSONEvent(t, events, "final")
+	if final["text"] != "done" {
+		t.Fatalf("reasoning must not be folded into final answer: %#v", final)
+	}
+}
+
 func TestExecEventWriterTruncatesStreamJSONToolResults(t *testing.T) {
 	var stdout bytes.Buffer
 	writer := execEventWriter{
@@ -751,6 +790,17 @@ type toolCallingExecProvider struct {
 	answer     string
 }
 
+type reasoningExecProvider struct{}
+
+func (reasoningExecProvider) StreamCompletion(context.Context, zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	ch := make(chan zeroruntime.StreamEvent, 3)
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventReasoning, Content: "Thinking. "}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
 func (provider toolCallingExecProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
 	for _, message := range request.Messages {
 		if message.Role == zeroruntime.MessageRoleTool {
@@ -777,13 +827,19 @@ func (provider toolCallingExecProvider) StreamCompletion(ctx context.Context, re
 
 func findJSONEvent(t *testing.T, events []map[string]any, eventType string) map[string]any {
 	t.Helper()
-	for _, event := range events {
+	_, event := findJSONEventIndex(t, events, eventType)
+	return event
+}
+
+func findJSONEventIndex(t *testing.T, events []map[string]any, eventType string) (int, map[string]any) {
+	t.Helper()
+	for idx, event := range events {
 		if event["type"] == eventType {
-			return event
+			return idx, event
 		}
 	}
 	t.Fatalf("event %q not found in %#v", eventType, events)
-	return nil
+	return -1, nil
 }
 
 func findSessionEvent(t *testing.T, events []sessions.Event, eventType sessions.EventType) sessions.Event {

internal/cli/exec_provider_hint_test.go

@@ -0,0 +1,50 @@
+package cli
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+)
+
+// A recognized provider error in text mode prints the raw message plus a one-line
+// actionable hint; a non-provider error (or JSON mode) prints no hint.
+func TestWriteExecProviderErrorAppendsHint(t *testing.T) {
+	t.Run("provider auth error gets a hint", func(t *testing.T) {
+		var stdout, stderr bytes.Buffer
+		code := writeExecProviderError(&stdout, &stderr, execOutputText, "provider_error",
+			"auth error: your API key is missing or invalid")
+		if code != exitProvider {
+			t.Fatalf("exit code = %d, want exitProvider", code)
+		}
+		out := stderr.String()
+		if !strings.Contains(out, "auth error:") {
+			t.Fatalf("expected raw message, got %q", out)
+		}
+		if !strings.Contains(out, "zero auth") {
+			t.Fatalf("expected an actionable hint referencing `zero auth`, got %q", out)
+		}
+	})
+
+	t.Run("non-provider error gets no hint", func(t *testing.T) {
+		var stdout, stderr bytes.Buffer
+		writeExecProviderError(&stdout, &stderr, execOutputText, "sandbox_error",
+			"sandbox setup failed: permission denied")
+		out := stderr.String()
+		// Exactly one "[zero]" line — no spurious hint attached to a local error.
+		if n := strings.Count(out, "[zero]"); n != 1 {
+			t.Fatalf("expected exactly one [zero] line for a non-provider error, got %d:\n%s", n, out)
+		}
+	})
+
+	t.Run("json mode never appends a hint line", func(t *testing.T) {
+		var stdout, stderr bytes.Buffer
+		writeExecProviderError(&stdout, &stderr, execOutputJSON, "provider_error",
+			"auth error: bad key")
+		if stderr.Len() != 0 {
+			t.Fatalf("json mode must not write to stderr, got %q", stderr.String())
+		}
+		if strings.Contains(stdout.String(), "zero auth") {
+			t.Fatalf("json mode must not inject a hint into the structured payload, got %q", stdout.String())
+		}
+	})
+}

internal/cli/exec_writer.go

@@ -80,6 +80,16 @@ func (writer *execEventWriter) text(delta string) {
 	writer.writeStdout(delta)
 }
 
+func (writer *execEventWriter) reasoning(delta string) {
+	if writer.format == execOutputJSON {
+		writer.writeJSON(map[string]any{"type": "reasoning", "delta": delta})
+		return
+	}
+	if writer.format == execOutputStreamJSON {
+		writer.writeStreamJSON(streamjson.Event{Type: streamjson.EventReasoning, RunID: writer.runID, Delta: delta})
+	}
+}
+
 func (writer *execEventWriter) toolCall(call agent.ToolCall, registry *tools.Registry) {
 	if writer.format == execOutputJSON {
 		writer.writeJSON(map[string]any{

internal/cli/repoinfo.go

@@ -88,6 +88,15 @@ func formatRepoInfo(info repoinfo.Info) string {
 	fmt.Fprintf(&b, "Files:        %d\n", info.FileCount)
 	fmt.Fprintf(&b, "Directories:  %d (max depth %d)\n", info.DirectoryCount, info.MaxDepth)
 	fmt.Fprintf(&b, "Est. LOC:     %d\n", info.LOCEstimate)
+	if info.CommitCount != nil {
+		fmt.Fprintf(&b, "Commits:      %d\n", *info.CommitCount)
+	}
+	if info.BranchCount != nil {
+		fmt.Fprintf(&b, "Branches:     %d\n", *info.BranchCount)
+	}
+	if info.TagCount != nil {
+		fmt.Fprintf(&b, "Tags:         %d\n", *info.TagCount)
+	}
 	if info.PrimaryLanguage != "" {
 		fmt.Fprintf(&b, "Primary lang: %s\n", info.PrimaryLanguage)
 	}

internal/cli/repoinfo_test.go

@@ -33,7 +33,7 @@ func TestFormatRepoInfoText(t *testing.T) {
 	contrib := 3
 	info := repoInfoTestData(age, contrib)
 	out := formatRepoInfo(info)
-	for _, want := range []string{"Files", "TypeScript", "main", "GitHub Actions", "42", "3"} {
+	for _, want := range []string{"Files", "TypeScript", "main", "GitHub Actions", "42", "3", "Commits:      100", "Branches:     5", "Tags:         10"} {
 		if !strings.Contains(out, want) {
 			t.Fatalf("text output missing %q:\n%s", want, out)
 		}
@@ -113,12 +113,16 @@ func initTempGitRepo(t *testing.T) string {
 }
 
 func repoInfoTestData(age, contrib int) repoinfo.Info {
+	commits := 100
+	branches := 5
+	tags := 10
 	return repoinfo.Info{
 		FileCount: 10, DirectoryCount: 3, MaxDepth: 2, LOCEstimate: 500,
 		Languages:       []repoinfo.LangStat{{Name: "TypeScript", LOCEstimate: 400, FileCount: 4}},
 		PrimaryLanguage: "TypeScript", LanguageCount: 1,
 		WorkspaceType: "none", CICD: []string{"GitHub Actions"},
 		HasGit: true, Branch: "main", AgeDays: &age, Contributors90d: &contrib,
+		CommitCount: &commits, BranchCount: &branches, TagCount: &tags,
 	}
 }
 

internal/cli/workflows.go

@@ -41,6 +41,13 @@ type changesCommandOptions struct {
 	hasMessage   bool
 	dryRun       bool
 	maxDiffBytes int
+	remote       string
+	force        bool
+	title        string
+	body         string
+	fill         bool
+	draft        bool
+	yes          bool
 	auto         bool
 }
 
@@ -266,6 +273,10 @@ func runChanges(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 			return exitCrash
 		}
 		return exitSuccess
+	case "push":
+		return runChangesPush(args, stdout, stderr, deps)
+	case "pr", "pull-request":
+		return runChangesPR(args, stdout, stderr, deps)
 	default:
 		return writeExecUsageError(stderr, fmt.Sprintf("unknown changes command %q", command))
 	}
@@ -465,6 +476,41 @@ func parseChangesArgs(args []string, command string) (changesCommandOptions, boo
 				return options, false, err
 			}
 			options.maxDiffBytes = maxDiffBytes
+		case arg == "--remote":
+			value, next, err := nextFlagValue(args, index, arg)
+			if err != nil {
+				return options, false, err
+			}
+			options.remote = strings.TrimSpace(value)
+			index = next
+		case strings.HasPrefix(arg, "--remote="):
+			options.remote = strings.TrimSpace(strings.TrimPrefix(arg, "--remote="))
+		case arg == "--force":
+			options.force = true
+		case arg == "--title":
+			value, next, err := nextFlagValue(args, index, arg)
+			if err != nil {
+				return options, false, err
+			}
+			options.title = value
+			index = next
+		case strings.HasPrefix(arg, "--title="):
+			options.title = strings.TrimPrefix(arg, "--title=")
+		case arg == "--body":
+			value, next, err := nextFlagValue(args, index, arg)
+			if err != nil {
+				return options, false, err
+			}
+			options.body = value
+			index = next
+		case strings.HasPrefix(arg, "--body="):
+			options.body = strings.TrimPrefix(arg, "--body=")
+		case arg == "--fill":
+			options.fill = true
+		case arg == "--draft":
+			options.draft = true
+		case arg == "--yes":
+			options.yes = true
 		case arg == "-a" || arg == "--auto":
 			options.auto = true
 		case strings.HasPrefix(arg, "-"):
@@ -473,15 +519,30 @@ func parseChangesArgs(args []string, command string) (changesCommandOptions, boo
 			return options, false, execUsageError{fmt.Sprintf("unexpected changes argument %q", arg)}
 		}
 	}
+	if command != "commit" && options.message != "" {
+		return options, false, execUsageError{"--message is only valid with `zero changes commit`"}
+	}
 	if command != "commit" && (options.hasMessage || options.dryRun || options.auto) {
 		return options, false, execUsageError{"--message, --dry-run, and --auto are only valid with `zero changes commit`"}
 	}
 	if command == "commit" && options.hasMessage && options.auto {
 		return options, false, execUsageError{"cannot specify both --message and --auto"}
 	}
-	if command == "commit" && options.baseRef != "" {
+	if command != "commit" && command != "push" && options.dryRun {
+		return options, false, execUsageError{"--dry-run is only valid with commit or push"}
+	}
+	if command != "inspect" && options.baseRef != "" {
 		return options, false, execUsageError{"--base is only valid with `zero changes inspect`"}
 	}
+	if command != "push" && command != "pr" && (options.remote != "" || options.force) {
+		return options, false, execUsageError{"--remote and --force are only valid with push or pr"}
+	}
+	if command != "pr" && (options.title != "" || options.body != "" || options.fill || options.draft) {
+		return options, false, execUsageError{"--title, --body, --fill, and --draft are only valid with `zero changes pr`"}
+	}
+	if command != "push" && command != "pr" && options.yes {
+		return options, false, execUsageError{"--yes is only valid with push or pr"}
+	}
 	return options, false, nil
 }
 
@@ -759,14 +820,24 @@ func writeChangesHelp(w io.Writer) error {
 	_, err := fmt.Fprint(w, `Usage:
   zero changes inspect [flags]
   zero changes commit [flags]
+  zero changes push [flags]
+  zero changes pr [flags]
 
-Inspects local git changes and optionally creates a commit.
+Inspects, commits, pushes, and creates pull requests for local git changes.
 
 Flags:
   -C, --cwd <path>        Workspace directory
       --base <ref>        Diff against <ref>...HEAD instead of the working tree
       --diff-bytes <n>    Maximum diff bytes to include
   -m, --message <text>    Commit message for `+"`zero changes commit`"+`
+      --dry-run           Preview commit metadata / push without mutating git state
+      --remote <name>     Remote to push to (defaults to upstream tracked branch or origin)
+      --force             Use force-with-lease when pushing
+      --title <text>      PR title
+      --body <text>       PR body
+      --fill              Automatically populate PR title and body from commits
+      --draft             Create PR as a draft
+      --yes               Confirm pushing to a default/protected branch
   -a, --auto              Auto-generate commit message using LLM (use --dry-run to preview)
       --dry-run           Preview commit metadata without mutating git state
       --json              Print JSON output
@@ -775,6 +846,128 @@ Flags:
 	return err
 }
 
+func runChangesPush(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	options, help, err := parseChangesArgs(args, "push")
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if help {
+		if err := writeChangesHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	workspaceRoot, err := resolveWorkspaceRoot(options.cwd, deps)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	result, err := deps.pushChanges(context.Background(), zerogit.PushOptions{
+		Cwd:                    workspaceRoot,
+		Remote:                 options.remote,
+		Force:                  options.force,
+		DryRun:                 options.dryRun,
+		AllowPushDefaultBranch: options.yes,
+	})
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	if options.json {
+		if err := writePrettyJSON(stdout, result); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+
+	dryRunStr := ""
+	if options.dryRun {
+		dryRunStr = " (dry run)"
+	}
+	if _, err := fmt.Fprintf(stdout, "Pushed branch %s to remote %s%s\n", result.Branch, result.Remote, dryRunStr); err != nil {
+		return exitCrash
+	}
+	if result.Output != "" {
+		if _, err := fmt.Fprintln(stdout, result.Output); err != nil {
+			return exitCrash
+		}
+	}
+	return exitSuccess
+}
+
+func runChangesPR(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	options, help, err := parseChangesArgs(args, "pr")
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if help {
+		if err := writeChangesHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if !options.fill && options.title == "" {
+		return writeExecUsageError(stderr, "must provide either --fill or --title to run non-interactively")
+	}
+	workspaceRoot, err := resolveWorkspaceRoot(options.cwd, deps)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	if !options.json {
+		if _, err := fmt.Fprintln(stdout, "Pushing current branch to set upstream..."); err != nil {
+			return exitCrash
+		}
+	}
+	pushResult, err := deps.pushChanges(context.Background(), zerogit.PushOptions{
+		Cwd:                    workspaceRoot,
+		Remote:                 options.remote,
+		Force:                  options.force,
+		AllowPushDefaultBranch: options.yes,
+	})
+	if err != nil {
+		return writeExecUsageError(stderr, fmt.Sprintf("auto-push failed: %v", err))
+	}
+	if !options.json {
+		if _, err := fmt.Fprintf(stdout, "Pushed branch %s to remote %s\n", pushResult.Branch, pushResult.Remote); err != nil {
+			return exitCrash
+		}
+	}
+
+	prResult, err := deps.createPR(context.Background(), zerogit.PROptions{
+		Cwd:   workspaceRoot,
+		Fill:  options.fill,
+		Draft: options.draft,
+		Title: options.title,
+		Body:  options.body,
+	})
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	if options.json {
+		type prJSONResult struct {
+			Branch string `json:"branch"`
+			Remote string `json:"remote"`
+			Output string `json:"output"`
+		}
+		res := prJSONResult{
+			Branch: pushResult.Branch,
+			Remote: pushResult.Remote,
+			Output: strings.TrimSpace(prResult.Output),
+		}
+		if err := writePrettyJSON(stdout, res); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+
+	if _, err := fmt.Fprintln(stdout, strings.TrimSpace(prResult.Output)); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
 func generateAutoCommitMessage(ctx context.Context, provider zeroruntime.Provider, model string, summary zerogit.ChangeSummary) (string, error) {
 	var promptBuilder strings.Builder
 	promptBuilder.WriteString("Analyze the following git diff and generate a concise, conventional commit message.\n")

internal/config/resolver.go

@@ -590,9 +590,17 @@ func applyProviderEnv(cfg *FileConfig, providerKind ProviderKind, env envProfile
 	if profile.Name == "" {
 		profile.Name = string(providerKind)
 	}
+	// A non-official baseURL from the environment signals a proxy/gateway, so promote
+	// the first-party kind to its -compatible transport (which accepts a custom URL).
+	// Env-only by design: a custom baseURL in a project config.json is left as-is and
+	// rejected by Resolve, so an untrusted project can't redirect a first-party key to
+	// another host.
 	if providerKind == ProviderKindOpenAI && baseURL != "" && !isOfficialOpenAIBaseURL(baseURL) {
 		profile.ProviderKind = ProviderKindOpenAICompatible
 	}
+	if providerKind == ProviderKindAnthropic && baseURL != "" && !isOfficialAnthropicBaseURL(baseURL) {
+		profile.ProviderKind = ProviderKindAnthropicCompat
+	}
 	// When the env supplies only credentials (no baseURL) for a provider name that
 	// already exists, don't force the standard transport kind — a same-named
 	// openai-compatible/anthropic-compatible proxy or gateway must keep its kind.

internal/config/resolver_test.go

@@ -617,6 +617,35 @@ func TestResolveUsesOpenAIEnvFallback(t *testing.T) {
 	}
 }
 
+func TestResolveUsesAnthropicEnvBaseURLAsCompatible(t *testing.T) {
+	// ANTHROPIC_BASE_URL pointing at a proxy/gateway must resolve to an
+	// anthropic-compatible provider (mirroring OPENAI_BASE_URL), not the "requires
+	// official baseURL" rejection — issue #479. The env is user-controlled, so a
+	// custom URL there is a deliberate gateway choice, unlike a project config.json.
+	resolved, err := Resolve(ResolveOptions{
+		Env: map[string]string{
+			"ANTHROPIC_API_KEY":  "sk-ant-env",
+			"ANTHROPIC_BASE_URL": "https://gateway.example/anthropic",
+			"ANTHROPIC_MODEL":    "claude-custom",
+		},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if resolved.ActiveProvider != "anthropic" {
+		t.Fatalf("ActiveProvider = %q, want anthropic", resolved.ActiveProvider)
+	}
+	if resolved.Provider.ProviderKind != ProviderKindAnthropicCompat {
+		t.Fatalf("ProviderKind = %q, want anthropic-compatible", resolved.Provider.ProviderKind)
+	}
+	if resolved.Provider.BaseURL != "https://gateway.example/anthropic" {
+		t.Fatalf("BaseURL = %q, want gateway URL", resolved.Provider.BaseURL)
+	}
+	if resolved.Provider.APIKey != "sk-ant-env" || resolved.Provider.Model != "claude-custom" {
+		t.Fatalf("Provider = %#v, want env credentials/model", resolved.Provider)
+	}
+}
+
 func TestResolveUsesOpenAIAPIKeyOnlyWithDefaultModel(t *testing.T) {
 	resolved, err := Resolve(ResolveOptions{
 		Env: map[string]string{

internal/doctor/hardening.go

@@ -131,29 +131,45 @@ func doctorSandboxPolicy(cfg config.SandboxConfig) sandbox.Policy {
 // lspServersCheck reports which language servers ZERO would use are present on
 // PATH. Missing servers are not a failure — ZERO degrades to text-only edits for
 // those languages — so the worst status is WARN, and each missing server gets an
-// actionable install command keyed by its binary name.
+// actionable install command keyed by its binary name. Only the tier-1 servers
+// (lsp.CoreServerBinaries) drive the status: the long-tail servers configured
+// for breadth are listed informationally under missingOptional, because warning
+// about a missing zls on a machine with no Zig code would be permanent noise.
 func lspServersCheck(lookup func(string) (string, error)) Check {
 	if lookup == nil {
 		lookup = exec.LookPath
 	}
+	core := map[string]bool{}
+	for _, binary := range lsp.CoreServerBinaries() {
+		core[binary] = true
+	}
 	present := map[string]any{}
 	missing := map[string]any{}
+	missingOptional := map[string]any{}
 	for _, binary := range lsp.ServerBinaries() {
 		if _, err := lookup(binary); err == nil {
 			present[binary] = "on PATH"
 			continue
 		}
-		missing[binary] = lspRemedy(binary)
+		if core[binary] {
+			missing[binary] = lspRemedy(binary)
+		} else {
+			missingOptional[binary] = lspRemedy(binary)
+		}
+	}
+	details := map[string]any{"present": present}
+	if len(missingOptional) > 0 {
+		details["missingOptional"] = missingOptional
 	}
 	if len(missing) == 0 {
-		return check("lsp.servers", "LSP servers", StatusPass, "All known language servers are available on PATH.", map[string]any{
-			"present": present,
-		})
+		message := "All core language servers are available on PATH."
+		if len(missingOptional) > 0 {
+			message = fmt.Sprintf("All core language servers are available on PATH (%d optional server(s) not installed).", len(missingOptional))
+		}
+		return check("lsp.servers", "LSP servers", StatusPass, message, details)
 	}
-	return check("lsp.servers", "LSP servers", StatusWarn, fmt.Sprintf("%d language server(s) missing from PATH; affected files degrade to text-only edits.", len(missing)), map[string]any{
-		"present": present,
-		"missing": missing,
-	})
+	details["missing"] = missing
+	return check("lsp.servers", "LSP servers", StatusWarn, fmt.Sprintf("%d language server(s) missing from PATH; affected files degrade to text-only edits.", len(missing)), details)
 }
 
 // lspRemedy returns an actionable install command for a missing language-server

internal/errhint/errhint.go

@@ -0,0 +1,167 @@
+// Package errhint classifies provider/model failures into a few user-actionable
+// categories and turns them into a one-line "next step" hint.
+//
+// Provider errors already arrive with a classified string prefix from
+// providerio.ClassifiedError ("auth error:", "rate limit error:", …); lower-level
+// failures (DNS, TLS, timeouts, context-length) arrive as raw driver or library
+// messages. Classify matches both so the interactive error row (TUI) and the
+// `zero exec` provider-error path (CLI) can append one concrete next step instead
+// of dumping an identical red blob for every failure mode.
+package errhint
+
+import "strings"
+
+// Category buckets a provider/model failure into a small set of classes that each
+// map to a distinct recovery action.
+type Category int
+
+const (
+	// Unknown means the error didn't match any known signature; callers should
+	// emit no hint rather than guess.
+	Unknown Category = iota
+	Auth
+	RateLimit
+	Connectivity
+	ModelNotFound
+	ContextOverflow
+)
+
+// providerMarkers are the prefixes the provider layer attaches to every
+// provider-originated failure: providerio.ClassifiedError emits "auth error:",
+// "rate limit error:", "provider error:", and "provider request error:", and the
+// streaming paths wrap transport/decoding failures as "provider stream error:".
+// A UI surface's error can also be a *local* failure (a tool's "permission
+// denied", a "file does not exist", a config error), so Classify only proceeds
+// past this gate for messages that are recognizably from the provider — otherwise
+// a broad substring like "does not exist" would attach a bogus /model hint to an
+// unrelated local error.
+var providerMarkers = []string{
+	"auth error:",
+	"rate limit error:",
+	"provider error:",
+	"provider request error:",
+	"provider stream error:",
+}
+
+// Classify buckets err by scanning its message for known signatures. It is a
+// deliberately conservative string heuristic — the provider layer has already
+// discarded the numeric HTTP status by the time the error reaches a UI surface,
+// so the message is all we have. It first gates on a provider-origin marker (see
+// providerMarkers) so local failures never draw a provider hint, then sub-classifies.
+// Order matters: more specific signatures are tested before broader ones (e.g.
+// "context length" as overflow before the generic "timeout" as connectivity).
+func Classify(err error) Category {
+	if err == nil {
+		return Unknown
+	}
+	m := strings.ToLower(err.Error())
+	if !containsAny(m, providerMarkers...) {
+		return Unknown
+	}
+	switch {
+	case containsAny(m, "auth error:", "unauthorized", "api key", "api_key", "invalid_api_key",
+		"authentication", "permission denied", "forbidden") || containsStatusCode(m, "401", "403"):
+		return Auth
+	case containsAny(m, "rate limit", "rate_limit", "too many requests", "quota",
+		"resource_exhausted", "overloaded") || containsStatusCode(m, "429", "529"):
+		return RateLimit
+	case containsAny(m, "context length", "context window", "maximum context", "context_length_exceeded",
+		"too many tokens", "prompt is too long", "reduce the length", "maximum context length"):
+		return ContextOverflow
+	case containsAny(m, "model not found", "model_not_found", "does not exist", "unknown model",
+		"no such model", "unsupported model", "invalid model", "model is not"):
+		return ModelNotFound
+	case containsAny(m, "dial tcp", "no such host", "connection refused", "network is unreachable",
+		"i/o timeout", "context deadline exceeded", "tls handshake", "connection reset",
+		"unexpected eof", "lookup ", "timeout"):
+		return Connectivity
+	default:
+		return Unknown
+	}
+}
+
+// TUIHint returns a one-line hint referencing interactive slash commands, or ""
+// when the category is Unknown. Meant to sit under the raw error in the live
+// error row.
+func TUIHint(err error) string {
+	switch Classify(err) {
+	case Auth:
+		return "API key rejected — run /provider to re-check your credentials"
+	case RateLimit:
+		return "Rate limited — wait a moment, or switch model with /model"
+	case Connectivity:
+		return "Can't reach the provider — run /doctor --connectivity"
+	case ModelNotFound:
+		return "Model unavailable — pick another with /model"
+	case ContextOverflow:
+		return "Context window full — run /compact to free space"
+	default:
+		return ""
+	}
+}
+
+// CLIHint returns a one-line hint referencing `zero …` subcommands, or "" when the
+// category is Unknown. Meant for the non-interactive `zero exec` error path, where
+// slash commands don't apply.
+func CLIHint(err error) string {
+	switch Classify(err) {
+	case Auth:
+		return "API key rejected — run `zero auth` or set the provider's API key, then retry"
+	case RateLimit:
+		return "Rate limited — wait a moment, or switch model with --model"
+	case Connectivity:
+		return "Can't reach the provider — run `zero doctor`"
+	case ModelNotFound:
+		return "Model unavailable — run `zero doctor` or pick another with --model"
+	case ContextOverflow:
+		return "Context window full — shorten the prompt or start a fresh session"
+	default:
+		return ""
+	}
+}
+
+func containsAny(haystack string, needles ...string) bool {
+	for _, n := range needles {
+		if strings.Contains(haystack, n) {
+			return true
+		}
+	}
+	return false
+}
+
+// containsStatusCode reports whether haystack contains any of the given HTTP
+// status codes as a standalone number — not embedded in a longer digit run like
+// "completed in 4290ms" or "request id 14015" — so an incidental number can't be
+// mis-bucketed as an auth/rate-limit failure.
+func containsStatusCode(haystack string, codes ...string) bool {
+	return HasStatusCode(haystack, codes...)
+}
+
+// HasStatusCode reports whether haystack contains any of the given HTTP status
+// codes as a standalone number — not embedded in a longer digit run like
+// "completed in 4290ms" or "request id 14015". Exported so other packages (e.g.
+// the reconnect classifier) can gate on a status code without re-implementing
+// the digit-boundary check.
+func HasStatusCode(haystack string, codes ...string) bool {
+	for _, code := range codes {
+		for from := 0; ; {
+			rel := strings.Index(haystack[from:], code)
+			if rel < 0 {
+				break
+			}
+			pos := from + rel
+			beforeOK := pos == 0 || !isASCIIDigit(haystack[pos-1])
+			end := pos + len(code)
+			afterOK := end >= len(haystack) || !isASCIIDigit(haystack[end])
+			if beforeOK && afterOK {
+				return true
+			}
+			from = pos + 1
+		}
+	}
+	return false
+}
+
+func isASCIIDigit(b byte) bool {
+	return b >= '0' && b <= '9'
+}

internal/errhint/errhint_test.go

@@ -0,0 +1,94 @@
+package errhint
+
+import (
+	"errors"
+	"strings"
+	"testing"
+)
+
+func TestClassify(t *testing.T) {
+	cases := []struct {
+		name string
+		msg  string
+		want Category
+	}{
+		{"nil-ish empty", "", Unknown},
+		{"providerio auth prefix", "auth error: your API key is missing or invalid — run `zero auth`", Auth},
+		{"raw 401", "provider request error: 401 Unauthorized", Auth},
+		{"invalid api key", "provider request error: invalid_api_key: incorrect key provided", Auth},
+		{"rate limit prefix", "rate limit error: 429 too many requests", RateLimit},
+		{"overloaded", "provider error: model is overloaded, please retry", RateLimit},
+		{"resource exhausted gemini", "provider stream error: rpc error: code = ResourceExhausted desc = quota exceeded", RateLimit},
+		{"context length openai", "provider request error: this model's maximum context length is 128000 tokens", ContextOverflow},
+		{"context_length_exceeded", "provider request error: context_length_exceeded", ContextOverflow},
+		{"model not found", "provider request error: 404 the model `gpt-9` does not exist", ModelNotFound},
+		{"unknown model", "provider request error: unknown model: sonnet-99", ModelNotFound},
+		{"dns failure", "provider stream error: Post \"https://api...\": dial tcp: lookup api.foo.com: no such host", Connectivity},
+		{"connection refused", "provider stream error: dial tcp 127.0.0.1:443: connection refused", Connectivity},
+		{"deadline", "provider stream error: context deadline exceeded (Client.Timeout exceeded)", Connectivity},
+		{"provider-marked but no sub-signature", "provider error: something totally unexpected happened", Unknown},
+
+		// An incidental number that merely contains a status-code digit run must NOT
+		// be mis-bucketed (digit-boundary check), even behind a provider marker.
+		{"incidental latency number", "provider error: request completed in 4290ms then failed", Unknown},
+		{"incidental request id", "provider error: request id 14015 failed unexpectedly", Unknown},
+		{"standalone 401 still auth", "provider request error: got 401 from upstream", Auth},
+		{"standalone 429 still rate limit", "provider error: 429 returned", RateLimit},
+
+		// Local (non-provider) failures must NOT be classified — no provider marker,
+		// so no bogus recovery hint attaches to them.
+		{"local fs permission denied", "open /etc/shadow: permission denied", Unknown},
+		{"local file missing", "stat foo.go: no such file or directory", Unknown},
+		{"local model config typo", "unknown model: sonnet-99", Unknown},
+		{"tool timeout local", "tool bash timed out after 600s", Unknown},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := Classify(errors.New(tc.msg)); got != tc.want {
+				t.Fatalf("Classify(%q) = %v, want %v", tc.msg, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestClassifyNil(t *testing.T) {
+	if got := Classify(nil); got != Unknown {
+		t.Fatalf("Classify(nil) = %v, want Unknown", got)
+	}
+}
+
+// Context-overflow must win over the connectivity "timeout" catch-all: a message
+// mentioning both "context length" and a timeout is an overflow, not a network
+// problem.
+func TestContextOverflowBeatsConnectivity(t *testing.T) {
+	err := errors.New("provider request error: maximum context length exceeded; request timeout")
+	if got := Classify(err); got != ContextOverflow {
+		t.Fatalf("Classify = %v, want ContextOverflow", got)
+	}
+}
+
+func TestHintsPresentForKnownCategories(t *testing.T) {
+	known := []error{
+		errors.New("auth error: bad key"),
+		errors.New("rate limit error: 429"),
+		errors.New("provider stream error: dial tcp: no such host"),
+		errors.New("provider request error: model does not exist"),
+		errors.New("provider request error: maximum context length is 128000 tokens"),
+	}
+	for _, err := range known {
+		if h := TUIHint(err); strings.TrimSpace(h) == "" {
+			t.Fatalf("TUIHint(%q) empty, want a hint", err)
+		}
+		if h := CLIHint(err); strings.TrimSpace(h) == "" {
+			t.Fatalf("CLIHint(%q) empty, want a hint", err)
+		}
+	}
+	// Unknown yields no hint on either surface.
+	unknown := errors.New("provider error: mystery")
+	if h := TUIHint(unknown); h != "" {
+		t.Fatalf("TUIHint(unknown) = %q, want empty", h)
+	}
+	if h := CLIHint(unknown); h != "" {
+		t.Fatalf("CLIHint(unknown) = %q, want empty", h)
+	}
+}

internal/lsp/registry.go

@@ -9,26 +9,134 @@ import (
 
 // serverCommands maps a file extension to the language-server command (argv) ZERO
 // will spawn. The first element is the binary looked up on PATH; missing binaries
-// are not an error, the agent just degrades to text-only for that file.
+// are not an error, the agent just degrades to text-only for that file. Every
+// command is the language's community-standard server invoked in stdio mode, so
+// a wrong guess can never spawn something surprising — only fail the PATH check.
 var serverCommands = map[string][]string{
-	".go":  {"gopls", "serve"},
-	".ts":  {"typescript-language-server", "--stdio"},
-	".tsx": {"typescript-language-server", "--stdio"},
-	".js":  {"typescript-language-server", "--stdio"},
-	".jsx": {"typescript-language-server", "--stdio"},
-	".py":  {"pyright-langserver", "--stdio"},
-	".rs":  {"rust-analyzer"},
+	".go":     {"gopls", "serve"},
+	".ts":     {"typescript-language-server", "--stdio"},
+	".tsx":    {"typescript-language-server", "--stdio"},
+	".mts":    {"typescript-language-server", "--stdio"},
+	".cts":    {"typescript-language-server", "--stdio"},
+	".js":     {"typescript-language-server", "--stdio"},
+	".jsx":    {"typescript-language-server", "--stdio"},
+	".mjs":    {"typescript-language-server", "--stdio"},
+	".cjs":    {"typescript-language-server", "--stdio"},
+	".py":     {"pyright-langserver", "--stdio"},
+	".pyi":    {"pyright-langserver", "--stdio"},
+	".rs":     {"rust-analyzer"},
+	".c":      {"clangd"},
+	".h":      {"clangd"},
+	".cpp":    {"clangd"},
+	".cc":     {"clangd"},
+	".cxx":    {"clangd"},
+	".hpp":    {"clangd"},
+	".java":   {"jdtls"},
+	".kt":     {"kotlin-language-server"},
+	".kts":    {"kotlin-language-server"},
+	".rb":     {"ruby-lsp"},
+	".php":    {"intelephense", "--stdio"},
+	".zig":    {"zls"},
+	".lua":    {"lua-language-server"},
+	".ex":     {"elixir-ls"},
+	".exs":    {"elixir-ls"},
+	".hs":     {"haskell-language-server-wrapper", "--lsp"},
+	".swift":  {"sourcekit-lsp"},
+	".ml":     {"ocamllsp"},
+	".mli":    {"ocamllsp"},
+	".scala":  {"metals"},
+	".clj":    {"clojure-lsp"},
+	".cljs":   {"clojure-lsp"},
+	".cljc":   {"clojure-lsp"},
+	".dart":   {"dart", "language-server"},
+	".gleam":  {"gleam", "lsp"},
+	".nix":    {"nixd"},
+	".tf":     {"terraform-ls", "serve"},
+	".sh":     {"bash-language-server", "start"},
+	".bash":   {"bash-language-server", "start"},
+	".yaml":   {"yaml-language-server", "--stdio"},
+	".yml":    {"yaml-language-server", "--stdio"},
+	".json":   {"vscode-json-language-server", "--stdio"},
+	".css":    {"vscode-css-language-server", "--stdio"},
+	".scss":   {"vscode-css-language-server", "--stdio"},
+	".less":   {"vscode-css-language-server", "--stdio"},
+	".html":   {"vscode-html-language-server", "--stdio"},
+	".svelte": {"svelteserver", "--stdio"},
+	".vue":    {"vue-language-server", "--stdio"},
+	".astro":  {"astro-ls", "--stdio"},
 }
 
 // languageIDs maps a file extension to the LSP languageId used in didOpen.
 var languageIDs = map[string]string{
-	".go":  "go",
-	".ts":  "typescript",
-	".tsx": "typescriptreact",
-	".js":  "javascript",
-	".jsx": "javascriptreact",
-	".py":  "python",
-	".rs":  "rust",
+	".go":     "go",
+	".ts":     "typescript",
+	".tsx":    "typescriptreact",
+	".mts":    "typescript",
+	".cts":    "typescript",
+	".js":     "javascript",
+	".jsx":    "javascriptreact",
+	".mjs":    "javascript",
+	".cjs":    "javascript",
+	".py":     "python",
+	".pyi":    "python",
+	".rs":     "rust",
+	".c":      "c",
+	".h":      "c",
+	".cpp":    "cpp",
+	".cc":     "cpp",
+	".cxx":    "cpp",
+	".hpp":    "cpp",
+	".java":   "java",
+	".kt":     "kotlin",
+	".kts":    "kotlin",
+	".rb":     "ruby",
+	".php":    "php",
+	".zig":    "zig",
+	".lua":    "lua",
+	".ex":     "elixir",
+	".exs":    "elixir",
+	".hs":     "haskell",
+	".swift":  "swift",
+	".ml":     "ocaml",
+	".mli":    "ocaml",
+	".scala":  "scala",
+	".clj":    "clojure",
+	".cljs":   "clojure",
+	".cljc":   "clojure",
+	".dart":   "dart",
+	".gleam":  "gleam",
+	".nix":    "nix",
+	".tf":     "terraform",
+	".sh":     "shellscript",
+	".bash":   "shellscript",
+	".yaml":   "yaml",
+	".yml":    "yaml",
+	".json":   "json",
+	".css":    "css",
+	".scss":   "scss",
+	".less":   "less",
+	".html":   "html",
+	".svelte": "svelte",
+	".vue":    "vue",
+	".astro":  "astro",
+}
+
+// coreServerBinaries are the tier-1 servers for the languages agents hit most;
+// `zero doctor` treats their absence as warn-worthy. The long tail of servers
+// configured above for breadth is reported informationally only — warning about
+// a missing zls on a machine with no Zig code would be permanent noise.
+var coreServerBinaries = []string{
+	"gopls",
+	"typescript-language-server",
+	"pyright-langserver",
+	"rust-analyzer",
+}
+
+// CoreServerBinaries returns the tier-1 server binaries (sorted copy).
+func CoreServerBinaries() []string {
+	binaries := append([]string(nil), coreServerBinaries...)
+	sort.Strings(binaries)
+	return binaries
 }
 
 // ServerBinaries returns the unique set of language-server binaries ZERO may

internal/modelregistry/catalog.go

@@ -3,7 +3,6 @@ package modelregistry
 import (
 	"fmt"
 	"strings"
-	"sync"
 )
 
 const DefaultModelID = "gpt-4.1"
@@ -43,20 +42,8 @@ type ListOptions struct {
 	Capability        ModelCapability
 }
 
-// defaultRegistry builds the default registry exactly once and caches the result.
-// The default catalog is static, every accessor returns clones, and a built
-// Registry is never mutated (concurrent reads of its maps are safe), so a single
-// shared instance is correct — and it avoids re-cloning all entries and
-// recompiling every fuzzy-match regex on the per-frame TUI paths (title chip,
-// context gauge, pickers) that call DefaultRegistry each render.
-var defaultRegistry = sync.OnceValues(func() (Registry, error) {
-	return NewRegistry(DefaultModelEntries())
-})
-
-// DefaultRegistry returns the shared, immutable default model registry, building
-// it on first use.
 func DefaultRegistry() (Registry, error) {
-	return defaultRegistry()
+	return NewRegistry(DefaultModelEntries())
 }
 
 func DefaultModelEntries() []ModelEntry {
@@ -76,6 +63,10 @@ func DefaultModelEntries() []ModelEntry {
 		googleModel("gemini-2.5-flash-lite", "Gemini 2.5 Flash-Lite", "gemini-2.5-flash-lite", ModelStatusActive, []string{"google:gemini-2.5-flash-lite", "gemini-flash-lite"}, ContextLimits{ContextWindow: 1_048_576, MaxOutputTokens: 65_536}, ModelCost{InputPerMillion: 0.1, CachedInputPerMillion: 0.01, OutputPerMillion: 0.4}, []ModelCapability{ModelCapabilityVision, ModelCapabilityJSONMode, ModelCapabilityReasoning, ModelCapabilityLongContext}, standardReasoningEfforts(), "Google low-cost Flash model for background routing and summaries."),
 	}
 	decorateModelDepth(entries)
+	// Overlay volatile facts (context limits, base pricing) from a cached
+	// models.dev snapshot when one is present and fresh — see modelsdev.go.
+	// Identity fields (ids, aliases, patterns, deprecations) stay curated.
+	entries = applyModelsDevOverrides(entries, cachedModelsDevProviders())
 	return cloneModelEntries(entries)
 }
 
@@ -89,13 +80,11 @@ func decorateModelDepth(entries []ModelEntry) {
 		patterns      []string
 		deprecation   *DeprecationRule
 		upgradeTarget string
-		fastVariant   string
 	}{
 		"claude-sonnet-4.5": {
 			defaultEffort: ReasoningEffortMedium,
 			patterns:      []string{`(?i)^sonnet[^a-z0-9]*4[.\s]?5$`},
 			upgradeTarget: "claude-opus-4.1",
-			fastVariant:   "claude-haiku-4.5",
 		},
 		"claude-opus-4.1": {
 			defaultEffort: ReasoningEffortHigh,
@@ -109,7 +98,6 @@ func decorateModelDepth(entries []ModelEntry) {
 		"gemini-2.5-pro": {
 			defaultEffort: ReasoningEffortMedium,
 			patterns:      []string{`(?i)^gemini[^a-z0-9]*pro$`},
-			fastVariant:   "gemini-2.5-flash",
 		},
 		"gemini-2.5-flash": {
 			defaultEffort: ReasoningEffortLow,
@@ -118,12 +106,6 @@ func decorateModelDepth(entries []ModelEntry) {
 		"gemini-2.5-flash-lite": {
 			upgradeTarget: "gemini-2.5-flash",
 		},
-		"gpt-4.1": {
-			fastVariant: "gpt-4.1-mini",
-		},
-		"gpt-4o": {
-			fastVariant: "gpt-4o-mini",
-		},
 		"gpt-4.1-mini": {
 			upgradeTarget: "gpt-4.1",
 		},
@@ -165,9 +147,6 @@ func decorateModelDepth(entries []ModelEntry) {
 		if extra.upgradeTarget != "" {
 			entries[index].UpgradeTargetID = extra.upgradeTarget
 		}
-		if extra.fastVariant != "" {
-			entries[index].FastVariantID = extra.fastVariant
-		}
 	}
 }
 

internal/modelregistry/models.go

@@ -110,14 +110,7 @@ type ModelEntry struct {
 	// during mid-run escalation. Empty means no escalation target (e.g. top-tier
 	// models). Resolved and availability-checked by Registry.UpgradeTarget.
 	UpgradeTargetID string
-	// FastVariantID, when set, names a genuinely faster same-family model this one
-	// can drop down to (e.g. claude-sonnet-4.5 -> claude-haiku-4.5). Empty means no
-	// fast variant. It is the single source of truth for the /fast command and the
-	// "f" fast-mode marker: Registry.FastVariant walks it forward and BaseVariant
-	// walks it in reverse, so a model is "in fast mode" iff another model names it
-	// here. Resolved and availability-checked like UpgradeTargetID.
-	FastVariantID string
-	Description   string
+	Description     string
 }
 
 // DeprecationRule describes how a deprecated model is phased out and what to use
@@ -375,18 +368,6 @@ func NewRegistry(entries []ModelEntry) (Registry, error) {
 			return Registry{}, fmt.Errorf("model %q upgrade target %q does not resolve to a known model", entry.ID, targetID)
 		}
 	}
-	// A non-empty FastVariantID must resolve to a known model, mirroring the
-	// UpgradeTargetID check — otherwise /fast and the "f" fast-mode marker would
-	// silently vanish at runtime on a catalog typo instead of failing loudly here.
-	for _, entry := range registry.models {
-		fastID := strings.TrimSpace(entry.FastVariantID)
-		if fastID == "" {
-			continue
-		}
-		if _, ok := registry.Get(fastID); !ok {
-			return Registry{}, fmt.Errorf("model %q fast variant %q does not resolve to a known model", entry.ID, fastID)
-		}
-	}
 	return registry, nil
 }
 

internal/modelregistry/modelsdev.go

@@ -0,0 +1,269 @@
+package modelregistry
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// Live models.dev overlay for the curated catalog. The hand-maintained
+// DefaultModelEntries list is the source of truth for identity — ids, aliases,
+// match patterns, deprecations, escalation targets — but its VOLATILE facts
+// (context window, max output tokens, per-million pricing) go stale between
+// releases. When a cached snapshot of https://models.dev/api.json is present,
+// those fields are refreshed from it at registry construction; everything else
+// stays curated. The overlay never adds models (an auto-added entry would lack
+// aliases, match patterns, and provider wiring) and never touches the network
+// on the registry hot path — fetching happens only in the explicit background
+// refresh, cached to disk with a TTL.
+
+const (
+	modelsDevDefaultURL = "https://models.dev/api.json"
+	// modelsDevRefreshAfter is how old the cache may get before a background
+	// refresh re-fetches it.
+	modelsDevRefreshAfter = 24 * time.Hour
+	// modelsDevMaxAge is the oldest cache still applied as an overlay. Beyond
+	// this the curated catalog (updated with the binary) is likely fresher than
+	// the snapshot, so a stale file is ignored rather than trusted.
+	modelsDevMaxAge      = 7 * 24 * time.Hour
+	modelsDevFetchLimit  = 32 << 20 // 32MiB guard on the response body
+	modelsDevFetchWindow = 15 * time.Second
+)
+
+// modelsDevModel is the subset of a models.dev model record the overlay uses.
+type modelsDevModel struct {
+	Limit struct {
+		Context int `json:"context"`
+		Output  int `json:"output"`
+	} `json:"limit"`
+	Cost struct {
+		Input      float64 `json:"input"`
+		Output     float64 `json:"output"`
+		CacheRead  float64 `json:"cache_read"`
+		CacheWrite float64 `json:"cache_write"`
+	} `json:"cost"`
+}
+
+// modelsDevProvider matches one provider object in api.json.
+type modelsDevProvider struct {
+	Models map[string]modelsDevModel `json:"models"`
+}
+
+// parseModelsDev decodes an api.json document into provider-slug -> api-model
+// -> record. api.json is a top-level object keyed by provider id.
+func parseModelsDev(data []byte) (map[string]map[string]modelsDevModel, error) {
+	var doc map[string]modelsDevProvider
+	if err := json.Unmarshal(data, &doc); err != nil {
+		return nil, fmt.Errorf("modelregistry: parse models.dev: %w", err)
+	}
+	providers := make(map[string]map[string]modelsDevModel, len(doc))
+	for slug, provider := range doc {
+		if len(provider.Models) == 0 {
+			continue
+		}
+		providers[slug] = provider.Models
+	}
+	if len(providers) == 0 {
+		return nil, fmt.Errorf("modelregistry: models.dev document has no providers")
+	}
+	return providers, nil
+}
+
+// modelsDevSlugs maps an entry's provider kind to the models.dev provider ids
+// worth checking. Only first-party slugs are used: router entries on models.dev
+// carry generic or stale numbers.
+func modelsDevSlugs(kind ProviderKind) []string {
+	switch kind {
+	case ProviderAnthropic:
+		return []string{"anthropic"}
+	case ProviderOpenAI:
+		return []string{"openai"}
+	case ProviderGoogle:
+		return []string{"google", "google-vertex"}
+	default:
+		return nil
+	}
+}
+
+// applyModelsDevOverrides refreshes each curated entry's context limits and
+// base pricing from the snapshot, when the snapshot knows the model. Tiered
+// pricing is never touched: models.dev has no tier data, and mixing a live
+// base rate with curated tiers would misprice the tier boundaries.
+func applyModelsDevOverrides(entries []ModelEntry, providers map[string]map[string]modelsDevModel) []ModelEntry {
+	if len(providers) == 0 {
+		return entries
+	}
+	for i := range entries {
+		entry := &entries[i]
+		var record modelsDevModel
+		found := false
+		for _, slug := range modelsDevSlugs(entry.Provider) {
+			if models, ok := providers[slug]; ok {
+				if candidate, ok := models[strings.TrimSpace(entry.APIModel)]; ok {
+					record = candidate
+					found = true
+					break
+				}
+			}
+		}
+		if !found {
+			continue
+		}
+		if record.Limit.Context > 0 {
+			entry.ContextLimits.ContextWindow = record.Limit.Context
+		}
+		if record.Limit.Output > 0 {
+			entry.ContextLimits.MaxOutputTokens = record.Limit.Output
+		}
+		if len(entry.Cost.Tiers) == 0 && record.Cost.Input > 0 && record.Cost.Output > 0 {
+			entry.Cost.InputPerMillion = record.Cost.Input
+			entry.Cost.OutputPerMillion = record.Cost.Output
+			if record.Cost.CacheRead > 0 {
+				entry.Cost.CachedInputPerMillion = record.Cost.CacheRead
+			}
+			if record.Cost.CacheWrite > 0 {
+				entry.Cost.CacheWritePerMillion = record.Cost.CacheWrite
+			}
+			entry.Cost.Source = "models.dev/api.json (cached)"
+		}
+	}
+	return entries
+}
+
+// modelsDevCachePath returns the on-disk cache location. ZERO_MODELS_CACHE_PATH
+// overrides it (used by tests and unusual setups).
+func modelsDevCachePath() (string, error) {
+	if override := strings.TrimSpace(os.Getenv("ZERO_MODELS_CACHE_PATH")); override != "" {
+		return override, nil
+	}
+	base, err := os.UserCacheDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(base, "zero", "modelsdev.json"), nil
+}
+
+var (
+	modelsDevOnce    sync.Once
+	modelsDevCached  map[string]map[string]modelsDevModel
+	modelsDevEnabled atomic.Bool
+)
+
+// EnableModelsDevOverlay opts the process into applying the cached models.dev
+// snapshot on top of the curated catalog. The CLI entrypoint calls it; library
+// consumers and tests that never do get the curated catalog byte-identical to
+// before, so hermetic tests can't be perturbed by a cache file on the machine.
+// ZERO_DISABLE_MODELS_FETCH disables both the overlay and the fetch.
+func EnableModelsDevOverlay() {
+	modelsDevEnabled.Store(true)
+}
+
+// cachedModelsDevProviders loads the cached snapshot once per process. Not
+// enabled, missing, stale (> modelsDevMaxAge), or malformed all yield nil and
+// the curated catalog is used untouched. Read once deliberately:
+// DefaultRegistry is called on hot paths (pickers, cost views) and must not
+// re-stat the file every time; a background refresh benefits the NEXT process.
+func cachedModelsDevProviders() map[string]map[string]modelsDevModel {
+	if !modelsDevEnabled.Load() || strings.TrimSpace(os.Getenv("ZERO_DISABLE_MODELS_FETCH")) != "" {
+		return nil
+	}
+	modelsDevOnce.Do(func() {
+		path, err := modelsDevCachePath()
+		if err != nil {
+			return
+		}
+		info, err := os.Stat(path)
+		if err != nil || time.Since(info.ModTime()) > modelsDevMaxAge {
+			return
+		}
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return
+		}
+		providers, err := parseModelsDev(data)
+		if err != nil {
+			return
+		}
+		modelsDevCached = providers
+	})
+	return modelsDevCached
+}
+
+// resetModelsDevCacheForTest clears the process-level cache memoization and
+// disables the overlay.
+func resetModelsDevCacheForTest() {
+	modelsDevOnce = sync.Once{}
+	modelsDevCached = nil
+	modelsDevEnabled.Store(false)
+}
+
+// RefreshModelsDevCache fetches models.dev/api.json into the on-disk cache
+// when the cache is missing or older than modelsDevRefreshAfter. It is safe to
+// call fire-and-forget from startup (use a goroutine); it never affects the
+// current process's registry (see cachedModelsDevProviders). Disabled entirely
+// by ZERO_DISABLE_MODELS_FETCH. The URL can be overridden with ZERO_MODELS_URL.
+func RefreshModelsDevCache(ctx context.Context) error {
+	if strings.TrimSpace(os.Getenv("ZERO_DISABLE_MODELS_FETCH")) != "" {
+		return nil
+	}
+	path, err := modelsDevCachePath()
+	if err != nil {
+		return err
+	}
+	if info, err := os.Stat(path); err == nil && time.Since(info.ModTime()) < modelsDevRefreshAfter {
+		return nil
+	}
+
+	url := strings.TrimSpace(os.Getenv("ZERO_MODELS_URL"))
+	if url == "" {
+		url = modelsDevDefaultURL
+	}
+	fetchCtx, cancel := context.WithTimeout(ctx, modelsDevFetchWindow)
+	defer cancel()
+	request, err := http.NewRequestWithContext(fetchCtx, http.MethodGet, url, nil)
+	if err != nil {
+		return err
+	}
+	request.Header.Set("User-Agent", "zero-models-refresh/0.1")
+	response, err := http.DefaultClient.Do(request)
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode != http.StatusOK {
+		return fmt.Errorf("modelregistry: models.dev fetch: HTTP %d", response.StatusCode)
+	}
+	data, err := io.ReadAll(io.LimitReader(response.Body, modelsDevFetchLimit))
+	if err != nil {
+		return err
+	}
+	// Validate before persisting: a bad body must never clobber a good cache.
+	if _, err := parseModelsDev(data); err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return err
+	}
+	temp, err := os.CreateTemp(filepath.Dir(path), "modelsdev-*.json")
+	if err != nil {
+		return err
+	}
+	if _, err := temp.Write(data); err != nil {
+		_ = temp.Close()
+		_ = os.Remove(temp.Name())
+		return err
+	}
+	if err := temp.Close(); err != nil {
+		_ = os.Remove(temp.Name())
+		return err
+	}
+	return os.Rename(temp.Name(), path)
+}

internal/modelregistry/modelsdev_test.go

@@ -0,0 +1,228 @@
+package modelregistry
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+const sampleModelsDev = `{
+  "anthropic": {
+    "id": "anthropic",
+    "models": {
+      "claude-sonnet-4-5-20250929": {
+        "limit": {"context": 1000000, "output": 64000},
+        "cost": {"input": 3.5, "output": 17.5, "cache_read": 0.35, "cache_write": 4.4}
+      }
+    }
+  },
+  "google": {
+    "id": "google",
+    "models": {
+      "gemini-2.5-pro": {
+        "limit": {"context": 2097152, "output": 65536},
+        "cost": {"input": 9.99, "output": 9.99}
+      }
+    }
+  }
+}`
+
+func TestParseModelsDev(t *testing.T) {
+	providers, err := parseModelsDev([]byte(sampleModelsDev))
+	if err != nil {
+		t.Fatal(err)
+	}
+	record, ok := providers["anthropic"]["claude-sonnet-4-5-20250929"]
+	if !ok {
+		t.Fatal("expected anthropic sonnet record")
+	}
+	if record.Limit.Context != 1_000_000 || record.Cost.Input != 3.5 {
+		t.Fatalf("unexpected record: %+v", record)
+	}
+	if _, err := parseModelsDev([]byte(`{}`)); err == nil {
+		t.Fatal("empty document must be rejected")
+	}
+	if _, err := parseModelsDev([]byte(`not json`)); err == nil {
+		t.Fatal("malformed document must be rejected")
+	}
+}
+
+func TestApplyModelsDevOverrides(t *testing.T) {
+	// Point the cache at a non-existent file so DefaultModelEntries returns the
+	// pure curated catalog, then apply the sample snapshot explicitly.
+	t.Setenv("ZERO_MODELS_CACHE_PATH", filepath.Join(t.TempDir(), "absent.json"))
+	resetModelsDevCacheForTest()
+	t.Cleanup(resetModelsDevCacheForTest)
+
+	providers, err := parseModelsDev([]byte(sampleModelsDev))
+	if err != nil {
+		t.Fatal(err)
+	}
+	entries := applyModelsDevOverrides(DefaultModelEntries(), providers)
+
+	var sonnet, geminiPro, opus ModelEntry
+	for _, entry := range entries {
+		switch entry.ID {
+		case "claude-sonnet-4.5":
+			sonnet = entry
+		case "gemini-2.5-pro":
+			geminiPro = entry
+		case "claude-opus-4.1":
+			opus = entry
+		}
+	}
+
+	// Known model: limits and base pricing refreshed from the snapshot.
+	if sonnet.ContextLimits.ContextWindow != 1_000_000 || sonnet.ContextLimits.MaxOutputTokens != 64_000 {
+		t.Fatalf("sonnet limits not overridden: %+v", sonnet.ContextLimits)
+	}
+	if sonnet.Cost.InputPerMillion != 3.5 || sonnet.Cost.OutputPerMillion != 17.5 || sonnet.Cost.CacheWritePerMillion != 4.4 {
+		t.Fatalf("sonnet cost not overridden: %+v", sonnet.Cost)
+	}
+	if sonnet.Cost.Source != "models.dev/api.json (cached)" {
+		t.Fatalf("sonnet cost source not marked: %q", sonnet.Cost.Source)
+	}
+
+	// Tiered pricing is curated: limits refresh, cost must NOT (gemini-2.5-pro
+	// has curated tiers and the snapshot's flat 9.99 would misprice them).
+	if geminiPro.ContextLimits.ContextWindow != 2_097_152 {
+		t.Fatalf("gemini limits not overridden: %+v", geminiPro.ContextLimits)
+	}
+	if geminiPro.Cost.InputPerMillion == 9.99 || len(geminiPro.Cost.Tiers) == 0 {
+		t.Fatalf("tiered cost must stay curated: %+v", geminiPro.Cost)
+	}
+
+	// Model absent from the snapshot: untouched.
+	if opus.ContextLimits.ContextWindow != 200_000 || opus.Cost.InputPerMillion != 15 {
+		t.Fatalf("opus must be untouched: %+v %+v", opus.ContextLimits, opus.Cost)
+	}
+}
+
+func TestRefreshModelsDevCacheFetchesAndCaches(t *testing.T) {
+	hits := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hits++
+		_, _ = w.Write([]byte(sampleModelsDev))
+	}))
+	defer server.Close()
+
+	cachePath := filepath.Join(t.TempDir(), "modelsdev.json")
+	t.Setenv("ZERO_MODELS_CACHE_PATH", cachePath)
+	t.Setenv("ZERO_MODELS_URL", server.URL)
+	t.Setenv("ZERO_DISABLE_MODELS_FETCH", "")
+
+	if err := RefreshModelsDevCache(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	if hits != 1 {
+		t.Fatalf("expected 1 fetch, got %d", hits)
+	}
+	if _, err := os.Stat(cachePath); err != nil {
+		t.Fatalf("cache file missing: %v", err)
+	}
+	// Fresh cache: second call must not re-fetch.
+	if err := RefreshModelsDevCache(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	if hits != 1 {
+		t.Fatalf("fresh cache must skip the fetch, got %d hits", hits)
+	}
+}
+
+func TestRefreshModelsDevCacheRejectsBadBodyWithoutClobbering(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte("not json"))
+	}))
+	defer server.Close()
+
+	cachePath := filepath.Join(t.TempDir(), "modelsdev.json")
+	if err := os.WriteFile(cachePath, []byte(sampleModelsDev), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	stale := time.Now().Add(-48 * time.Hour)
+	if err := os.Chtimes(cachePath, stale, stale); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("ZERO_MODELS_CACHE_PATH", cachePath)
+	t.Setenv("ZERO_MODELS_URL", server.URL)
+
+	if err := RefreshModelsDevCache(t.Context()); err == nil {
+		t.Fatal("bad body must return an error")
+	}
+	content, err := os.ReadFile(cachePath)
+	if err != nil || string(content) != sampleModelsDev {
+		t.Fatal("bad fetch must not clobber the existing cache")
+	}
+}
+
+func TestRefreshModelsDevCacheDisabledByEnv(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatal("fetch must not happen when disabled")
+	}))
+	defer server.Close()
+	t.Setenv("ZERO_MODELS_CACHE_PATH", filepath.Join(t.TempDir(), "modelsdev.json"))
+	t.Setenv("ZERO_MODELS_URL", server.URL)
+	t.Setenv("ZERO_DISABLE_MODELS_FETCH", "1")
+	if err := RefreshModelsDevCache(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestCachedModelsDevProvidersIgnoresStaleCache(t *testing.T) {
+	cachePath := filepath.Join(t.TempDir(), "modelsdev.json")
+	if err := os.WriteFile(cachePath, []byte(sampleModelsDev), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	stale := time.Now().Add(-modelsDevMaxAge - time.Hour)
+	if err := os.Chtimes(cachePath, stale, stale); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("ZERO_MODELS_CACHE_PATH", cachePath)
+	resetModelsDevCacheForTest()
+	t.Cleanup(resetModelsDevCacheForTest)
+	EnableModelsDevOverlay()
+
+	if providers := cachedModelsDevProviders(); providers != nil {
+		t.Fatal("stale cache must be ignored")
+	}
+}
+
+func TestCachedModelsDevProvidersRequiresOptIn(t *testing.T) {
+	cachePath := filepath.Join(t.TempDir(), "modelsdev.json")
+	if err := os.WriteFile(cachePath, []byte(sampleModelsDev), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("ZERO_MODELS_CACHE_PATH", cachePath)
+	resetModelsDevCacheForTest()
+	t.Cleanup(resetModelsDevCacheForTest)
+
+	// Without EnableModelsDevOverlay a fresh, valid cache must still be ignored:
+	// library consumers and hermetic tests get the pure curated catalog.
+	if providers := cachedModelsDevProviders(); providers != nil {
+		t.Fatal("overlay must be opt-in")
+	}
+}
+
+func TestDefaultModelEntriesAppliesFreshCache(t *testing.T) {
+	cachePath := filepath.Join(t.TempDir(), "modelsdev.json")
+	if err := os.WriteFile(cachePath, []byte(sampleModelsDev), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("ZERO_MODELS_CACHE_PATH", cachePath)
+	resetModelsDevCacheForTest()
+	t.Cleanup(resetModelsDevCacheForTest)
+	EnableModelsDevOverlay()
+
+	for _, entry := range DefaultModelEntries() {
+		if entry.ID == "claude-sonnet-4.5" {
+			if entry.ContextLimits.ContextWindow != 1_000_000 {
+				t.Fatalf("fresh cache must overlay the registry: %+v", entry.ContextLimits)
+			}
+			return
+		}
+	}
+	t.Fatal("claude-sonnet-4.5 not found")
+}

internal/modelregistry/variants.go

@@ -1,75 +0,0 @@
-package modelregistry
-
-import "strings"
-
-// The base<->fast model mapping is data, not logic: it lives entirely in each
-// ModelEntry's FastVariantID field (populated in the catalog, validated in
-// NewRegistry). Adding a pair is a one-line catalog change — never a code change
-// here. Both directions derive from that one field: FastVariant walks a model to
-// the fast id it names, BaseVariant scans for the model that names id as its fast
-// variant. This is why "fast mode" is encoded by model identity and needs no flag.
-//
-// Only models with an unambiguous 1:1 fast counterpart carry a FastVariantID. A
-// model with no clean counterpart (e.g. a top-tier Opus, where both Sonnet and
-// Haiku would be candidates) is intentionally left unset, so /fast reports "no
-// fast mode available" for it rather than guessing.
-
-// FastVariant resolves id to its configured faster same-family counterpart. It
-// resolves id to a concrete entry, reads that entry's FastVariantID, and returns
-// the target only when it resolves and is not deprecated. Returns (_, false) when
-// id is unknown, has no fast variant, or the variant is unavailable/deprecated.
-// Mirrors Registry.UpgradeTarget.
-func (registry Registry) FastVariant(id string) (ModelEntry, bool) {
-	source, ok := registry.Resolve(id)
-	if !ok {
-		return ModelEntry{}, false
-	}
-	fastID := strings.TrimSpace(source.FastVariantID)
-	if fastID == "" {
-		return ModelEntry{}, false
-	}
-	target, ok := registry.Resolve(fastID)
-	if !ok || target.Status == ModelStatusDeprecated {
-		return ModelEntry{}, false
-	}
-	return target, true
-}
-
-// BaseVariant resolves id to the base model that declares id as its fast variant
-// — the reverse of FastVariant. A non-empty result therefore means id IS a fast
-// model: this is exactly how the /fast command derives "currently in fast mode"
-// without storing any flag. Same resolution and availability rules as FastVariant.
-func (registry Registry) BaseVariant(id string) (ModelEntry, bool) {
-	source, ok := registry.Resolve(id)
-	if !ok {
-		return ModelEntry{}, false
-	}
-	for _, entry := range registry.models {
-		fastID := strings.TrimSpace(entry.FastVariantID)
-		if fastID == "" {
-			continue
-		}
-		target, ok := registry.Resolve(fastID)
-		if !ok || target.ID != source.ID {
-			continue
-		}
-		if entry.Status == ModelStatusDeprecated {
-			return ModelEntry{}, false
-		}
-		base, ok := registry.Resolve(entry.ID)
-		if !ok {
-			return ModelEntry{}, false
-		}
-		return base, true
-	}
-	return ModelEntry{}, false
-}
-
-// HasFastVariant reports whether id has an available fast variant — i.e. whether
-// `/fast on` would switch to a faster model from here. It is the predicate the
-// TUI uses to show the "f" fast-mode marker on a model (in the model/provider
-// pickers and the active-model indicator).
-func (registry Registry) HasFastVariant(id string) bool {
-	_, ok := registry.FastVariant(id)
-	return ok
-}

internal/modelregistry/variants_test.go

@@ -1,145 +0,0 @@
-package modelregistry
-
-import "testing"
-
-func testRegistry(t *testing.T) Registry {
-	t.Helper()
-	registry, err := DefaultRegistry()
-	if err != nil {
-		t.Fatalf("DefaultRegistry() error = %v", err)
-	}
-	return registry
-}
-
-// expectedFastPairs is the known-good base->fast mapping. It is a literal
-// regression guard: if the catalog silently drops or mis-wires a FastVariantID,
-// the forward/reverse assertions below fail loudly rather than making /fast a
-// quiet no-op. Keep in sync with the FastVariantID fields in catalog.go.
-var expectedFastPairs = map[string]string{
-	"claude-sonnet-4.5": "claude-haiku-4.5",
-	"gpt-4.1":           "gpt-4.1-mini",
-	"gpt-4o":            "gpt-4o-mini",
-	"gemini-2.5-pro":    "gemini-2.5-flash",
-}
-
-// Every FastVariantID declared in the catalog must resolve to a real, non-self,
-// non-deprecated model. NewRegistry already rejects an unresolvable id at build
-// time; this additionally guards the runtime availability contract that /fast and
-// the "f" marker depend on.
-func TestCatalogFastVariantsResolve(t *testing.T) {
-	registry := testRegistry(t)
-	for _, entry := range registry.List(ListOptions{IncludeDeprecated: true}) {
-		if entry.FastVariantID == "" {
-			continue
-		}
-		target, ok := registry.Resolve(entry.FastVariantID)
-		if !ok {
-			t.Errorf("model %q fast variant %q does not resolve", entry.ID, entry.FastVariantID)
-			continue
-		}
-		if target.ID == entry.ID {
-			t.Errorf("model %q names itself as its fast variant", entry.ID)
-		}
-		if target.Status == ModelStatusDeprecated {
-			t.Errorf("model %q fast variant %q is deprecated", entry.ID, entry.FastVariantID)
-		}
-	}
-}
-
-// No model may be both a base and a fast, or the two directions become ambiguous
-// (a model would be "currently fast" AND advertise a fast variant). Derived from
-// the registry so a bad catalog edit — e.g. chaining mini->nano — is caught.
-func TestFastVariantPairsUnambiguous(t *testing.T) {
-	registry := testRegistry(t)
-	role := map[string]string{} // canonical id -> "base" | "fast"
-	assign := func(id, r string) {
-		entry, ok := registry.Resolve(id)
-		if !ok {
-			return
-		}
-		if prev, seen := role[entry.ID]; seen && prev != r {
-			t.Fatalf("model %q is declared as both a base and a fast variant", entry.ID)
-		}
-		role[entry.ID] = r
-	}
-	for _, entry := range registry.List(ListOptions{IncludeDeprecated: true}) {
-		if entry.FastVariantID == "" {
-			continue
-		}
-		assign(entry.ID, "base")
-		assign(entry.FastVariantID, "fast")
-	}
-}
-
-// The known pairs must be present and wired in both directions, and neither end
-// may leak into the other role (a base is not itself fast; a fast has no fast).
-func TestFastVariantAndBaseVariant(t *testing.T) {
-	registry := testRegistry(t)
-
-	for base, fast := range expectedFastPairs {
-		// base -> fast (forward), and the fast id has NO fast variant of its own.
-		if got, ok := registry.FastVariant(base); !ok {
-			t.Errorf("FastVariant(%q) = not found, want the fast variant", base)
-		} else if canon, _ := registry.Resolve(fast); got.ID != canon.ID {
-			t.Errorf("FastVariant(%q).ID = %q, want %q", base, got.ID, canon.ID)
-		}
-		if _, ok := registry.FastVariant(fast); ok {
-			t.Errorf("FastVariant(%q) resolved, but a fast model should have no fast variant", fast)
-		}
-
-		// fast -> base (reverse), and the base id is NOT itself a fast variant.
-		if got, ok := registry.BaseVariant(fast); !ok {
-			t.Errorf("BaseVariant(%q) = not found, want the base variant", fast)
-		} else if canon, _ := registry.Resolve(base); got.ID != canon.ID {
-			t.Errorf("BaseVariant(%q).ID = %q, want %q", fast, got.ID, canon.ID)
-		}
-		if _, ok := registry.BaseVariant(base); ok {
-			t.Errorf("BaseVariant(%q) resolved, but a base model is not a fast variant", base)
-		}
-	}
-}
-
-// HasFastVariant is the TUI marker predicate: true only for base models with an
-// available fast variant, false for the fast models themselves, unpaired models,
-// and unknown ids.
-func TestHasFastVariant(t *testing.T) {
-	registry := testRegistry(t)
-	for base, fast := range expectedFastPairs {
-		if !registry.HasFastVariant(base) {
-			t.Errorf("HasFastVariant(%q) = false, want true", base)
-		}
-		if registry.HasFastVariant(fast) {
-			t.Errorf("HasFastVariant(%q) = true, want false (a fast model has no fast variant)", fast)
-		}
-	}
-	for _, id := range []string{"claude-opus-4.1", "definitely-not-a-real-model", ""} {
-		if registry.HasFastVariant(id) {
-			t.Errorf("HasFastVariant(%q) = true, want false", id)
-		}
-	}
-}
-
-func TestFastVariantAbsentAndUnknown(t *testing.T) {
-	registry := testRegistry(t)
-
-	// A real model with no configured fast/base variant reports none in both
-	// directions (a valid, handled case — /fast will say "no fast mode available").
-	const noVariant = "claude-opus-4.1"
-	if _, ok := registry.Resolve(noVariant); !ok {
-		t.Skipf("%q not in registry; skipping no-variant case", noVariant)
-	}
-	if _, ok := registry.FastVariant(noVariant); ok {
-		t.Errorf("FastVariant(%q) resolved, want none", noVariant)
-	}
-	if _, ok := registry.BaseVariant(noVariant); ok {
-		t.Errorf("BaseVariant(%q) resolved, want none", noVariant)
-	}
-
-	// An unknown id resolves to nothing in either direction.
-	if _, ok := registry.FastVariant("definitely-not-a-real-model"); ok {
-		t.Error("FastVariant(unknown) resolved, want none")
-	}
-	if _, ok := registry.BaseVariant("definitely-not-a-real-model"); ok {
-		t.Error("BaseVariant(unknown) resolved, want none")
-	}
-}

internal/providers/anthropic/cache_breakpoints_test.go

@@ -0,0 +1,96 @@
+package anthropic
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// The final two messages must carry a cache_control breakpoint on their last
+// cacheable block so the conversation transcript — not just system prompt and
+// tools — is a prompt-cache hit on the next turn. Earlier messages must stay
+// unmarked (Anthropic caps breakpoints at 4 per request: system, tools, and
+// these two).
+func TestAnthropicRequestMarksLastTwoMessagesForCaching(t *testing.T) {
+	var gotBody map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := json.NewDecoder(r.Body).Decode(&gotBody); err != nil {
+			t.Fatalf("decode request body: %v", err)
+		}
+		writeSSEEvent(w, "message_stop", `{"type":"message_stop"}`)
+	}))
+	defer server.Close()
+
+	provider, err := New(Options{BaseURL: server.URL + "/", Model: "claude-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: "first"},
+			{
+				Role:      zeroruntime.MessageRoleAssistant,
+				Content:   "calling a tool",
+				ToolCalls: []zeroruntime.ToolCall{{ID: "toolu_1", Name: "grep", Arguments: `{}`}},
+			},
+			{Role: zeroruntime.MessageRoleTool, Content: "result", ToolCallID: "toolu_1"},
+		},
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion returned error: %v", err)
+	}
+	drain(stream)
+
+	messages := gotBody["messages"].([]any)
+	if len(messages) != 3 {
+		t.Fatalf("expected 3 wire messages, got %d: %#v", len(messages), messages)
+	}
+	lastBlockHasCache := func(message any) bool {
+		blocks, ok := message.(map[string]any)["content"].([]any)
+		if !ok || len(blocks) == 0 {
+			return false
+		}
+		last := blocks[len(blocks)-1].(map[string]any)
+		control, ok := last["cache_control"].(map[string]any)
+		return ok && control["type"] == "ephemeral"
+	}
+	if lastBlockHasCache(messages[0]) {
+		t.Fatalf("first message must not carry a breakpoint: %#v", messages[0])
+	}
+	if !lastBlockHasCache(messages[1]) {
+		t.Fatalf("second-to-last message must carry a breakpoint: %#v", messages[1])
+	}
+	if !lastBlockHasCache(messages[2]) {
+		t.Fatalf("last message must carry a breakpoint: %#v", messages[2])
+	}
+}
+
+// A string-content message must be converted to a block array so it can carry
+// the breakpoint, and thinking blocks must never carry cache_control (the API
+// rejects them) — the marker goes on the last cacheable block instead.
+func TestApplyMessageCacheBreakpointsSkipsThinkingBlocks(t *testing.T) {
+	messages := []anthropicMessage{
+		{Role: "user", Content: "plain string"},
+		{Role: "assistant", Content: []map[string]any{
+			{"type": "text", "text": "answer"},
+			{"type": "thinking", "thinking": "hmm", "signature": "sig"},
+		}},
+	}
+	applyMessageCacheBreakpoints(messages)
+
+	userBlocks := contentBlocks(messages[0].Content)
+	if len(userBlocks) != 1 || userBlocks[0]["cache_control"] == nil {
+		t.Fatalf("string content must become a marked block: %#v", messages[0].Content)
+	}
+	assistantBlocks := contentBlocks(messages[1].Content)
+	if assistantBlocks[1]["cache_control"] != nil {
+		t.Fatalf("thinking block must not be marked: %#v", assistantBlocks[1])
+	}
+	if assistantBlocks[0]["cache_control"] == nil {
+		t.Fatalf("text block must carry the breakpoint instead: %#v", assistantBlocks[0])
+	}
+}

internal/providers/anthropic/provider.go

@@ -405,9 +405,35 @@ func (provider *Provider) anthropicRequest(request zeroruntime.CompletionRequest
 		}
 		mapped.Tools[len(mapped.Tools)-1].CacheControl = &cacheControl{Type: cacheEphemeral}
 	}
+	applyMessageCacheBreakpoints(mapped.Messages)
 	return mapped, nil
 }
 
+// applyMessageCacheBreakpoints marks the last content block of the final two
+// messages with cache_control so the conversation transcript is cached
+// turn-over-turn, not just the system prompt and tool definitions. Two message
+// breakpoints (not one) keep the previous turn's prefix a cache hit while the
+// newest suffix is being written. Anthropic allows at most 4 breakpoints per
+// request: system + tools use two above, these use the remaining two. Thinking
+// blocks cannot carry cache_control, so the marker goes on the last block that
+// can.
+func applyMessageCacheBreakpoints(messages []anthropicMessage) {
+	marked := 0
+	for i := len(messages) - 1; i >= 0 && marked < 2; i-- {
+		blocks := contentBlocks(messages[i].Content)
+		for j := len(blocks) - 1; j >= 0; j-- {
+			blockType, _ := blocks[j]["type"].(string)
+			if blockType == "thinking" || blockType == "redacted_thinking" {
+				continue
+			}
+			blocks[j]["cache_control"] = map[string]any{"type": cacheEphemeral}
+			messages[i].Content = blocks
+			marked++
+			break
+		}
+	}
+}
+
 func mapMessages(messages []zeroruntime.Message) (string, []anthropicMessage, error) {
 	systemParts := []string{}
 	mapped := []anthropicMessage{}

internal/providers/factory.go

@@ -24,7 +24,7 @@ type Options struct {
 	ModelRegistry *modelregistry.Registry
 	// OAuthResolver, when set, lets the provider authenticate model calls with an
 	// OAuth bearer token (preferred over the API key). nil => API-key auth only.
-	// Applied to the OpenAI and Anthropic providers.
+	// Applied to the OpenAI, Anthropic, and Google (Gemini) providers.
 	OAuthResolver providerio.TokenResolver
 	// OAuthLoginKey is the credential-store key OAuthResolver bound to (empty when
 	// there is no OAuth login). It is the SAME selection the bearer resolver made,
@@ -91,6 +91,7 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 			AuthScheme:      profile.AuthScheme,
 			AuthHeaderValue: profile.AuthHeaderValue,
 			CustomHeaders:   profile.CustomHeaders,
+			OAuthResolver:   options.OAuthResolver,
 			MaxTokens:       resolved.maxOutputTokens,
 			HTTPClient:      options.HTTPClient,
 			UserAgent:       options.UserAgent,

internal/providers/gemini/provider.go

@@ -49,6 +49,10 @@ type Options struct {
 	CustomHeaders   map[string]string
 	HTTPClient      *http.Client
 	UserAgent       string
+	// OAuthResolver, when set, supplies an OAuth bearer credential per request and
+	// is retried once with a forced token refresh after an upstream 401 (matching
+	// the OpenAI and Anthropic providers). Nil falls back to plain API-key auth.
+	OAuthResolver providerio.TokenResolver
 	// StreamIdleTimeout aborts the stream if no data arrives for this long.
 	// When unset, Zero uses providerio.ResolveStreamIdleTimeout — the
 	// ZERO_STREAM_IDLE_TIMEOUT override or providerio.DefaultStreamIdleTimeout.
@@ -67,6 +71,7 @@ type Provider struct {
 	customHeaders     map[string]string
 	httpClient        *http.Client
 	userAgent         string
+	oauthResolver     providerio.TokenResolver
 	streamIdleTimeout time.Duration
 }
 
@@ -95,6 +100,7 @@ func New(options Options) (*Provider, error) {
 		customHeaders:     providerio.CopyHeaders(options.CustomHeaders),
 		httpClient:        providerio.HTTPClient(options.HTTPClient),
 		userAgent:         options.UserAgent,
+		oauthResolver:     options.OAuthResolver,
 		streamIdleTimeout: providerio.ResolveStreamIdleTimeout(options.StreamIdleTimeout),
 	}, nil
 }
@@ -127,20 +133,22 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 	streamCtx, cancelStream := context.WithCancel(ctx)
 	defer cancelStream()
 
-	response, err := providerio.SendWithRetry(streamCtx, provider.httpClient, http.MethodPost, provider.streamURL(), body, func(request *http.Request) {
-		request.Header.Set("Content-Type", "application/json")
-		if provider.userAgent != "" {
-			request.Header.Set("User-Agent", provider.userAgent)
-		}
-		providerio.ApplyAuthHeaders(request, providerio.AuthHeaders{
+	response, err := providerio.SendWithAuthRetry(streamCtx, provider.httpClient, http.MethodPost, provider.streamURL(), body,
+		providerio.AuthHeaders{
 			APIKey:            provider.apiKey,
 			DefaultAuthHeader: "x-goog-api-key",
 			AuthHeader:        provider.authHeader,
 			AuthScheme:        provider.authScheme,
 			AuthHeaderValue:   provider.authHeaderValue,
 			CustomHeaders:     provider.customHeaders,
-		})
-	}, 0)
+		},
+		provider.oauthResolver,
+		func(request *http.Request) {
+			request.Header.Set("Content-Type", "application/json")
+			if provider.userAgent != "" {
+				request.Header.Set("User-Agent", provider.userAgent)
+			}
+		}, 0)
 	if err != nil {
 		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventError, Error: provider.redact("provider stream error: " + err.Error())})
 		return

internal/providers/gemini/provider_test.go

@@ -370,6 +370,57 @@ func TestStreamCompletionClassifiesHTTPAndPromptBlockErrors(t *testing.T) {
 	}
 }
 
+// A 401 with an OAuth resolver is retried once with a force-refreshed token; the
+// replayed request carries the refreshed bearer and succeeds.
+func TestStreamCompletionRetries401WithRefreshedToken(t *testing.T) {
+	var attempts int
+	var secondAuth string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		if attempts == 1 {
+			http.Error(w, `{"error":{"message":"token expired"}}`, http.StatusUnauthorized)
+			return
+		}
+		secondAuth = r.Header.Get("Authorization")
+		writeSSE(w, `{}`)
+	}))
+	defer server.Close()
+
+	var forceRefreshOnRetry bool
+	resolver := func(ctx context.Context, forceRefresh bool) (string, string, bool, error) {
+		if forceRefresh {
+			forceRefreshOnRetry = true
+			return "Authorization", "Bearer refreshed", true, nil
+		}
+		return "Authorization", "Bearer stale", true, nil
+	}
+
+	provider, err := New(Options{
+		APIKey:        "sk-google",
+		BaseURL:       server.URL,
+		Model:         "gemini-test",
+		OAuthResolver: resolver,
+	})
+	if err != nil {
+		t.Fatalf("New returned error: %v", err)
+	}
+	stream, err := provider.StreamCompletion(context.Background(), validRequest())
+	if err != nil {
+		t.Fatalf("StreamCompletion returned error: %v", err)
+	}
+	drain(stream)
+
+	if attempts != 2 {
+		t.Fatalf("attempts = %d, want 2 (initial 401 + one refreshed retry)", attempts)
+	}
+	if !forceRefreshOnRetry {
+		t.Fatalf("resolver was not called with forceRefresh on the retry")
+	}
+	if secondAuth != "Bearer refreshed" {
+		t.Fatalf("retry Authorization = %q, want refreshed bearer", secondAuth)
+	}
+}
+
 func TestStreamCompletionEmitsStreamErrorObject(t *testing.T) {
 	provider := newTestProviderWithKey(t, "sk-google", func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"error":{"code":429,"message":"stream failed sk-google","status":"RESOURCE_EXHAUSTED"}}`)

internal/providers/openai/provider.go

@@ -9,6 +9,7 @@ import (
 	"io"
 	"net/http"
 	"net/url"
+	"os"
 	"strings"
 	"time"
 
@@ -283,10 +284,10 @@ func (provider *Provider) emitChunk(
 	events chan<- zeroruntime.StreamEvent,
 ) {
 	for _, choice := range chunk.Choices {
-		if choice.Delta.ReasoningContent != "" {
+		if reasoning := choice.Delta.reasoningText(); reasoning != "" {
 			sendEvent(ctx, events, zeroruntime.StreamEvent{
 				Type:    zeroruntime.StreamEventReasoning,
-				Content: choice.Delta.ReasoningContent,
+				Content: reasoning,
 			})
 		}
 		if choice.Delta.Content != "" {
@@ -317,6 +318,13 @@ func (provider *Provider) emitChunk(
 	}
 }
 
+func (delta streamDelta) reasoningText() string {
+	if delta.ReasoningContent != "" {
+		return delta.ReasoningContent
+	}
+	return delta.Reasoning
+}
+
 // mapFinishReason maps OpenAI's finish_reason onto the runtime's normalized
 // terminal reasons. A normal finish ("stop"/"tool_calls"/"") returns "".
 func mapFinishReason(reason string) string {
@@ -414,6 +422,12 @@ func (provider *Provider) openAIRequest(request zeroruntime.CompletionRequest) c
 	if effort := openAIReasoningEffort(request.ReasoningEffort); effort != "" {
 		mapped.ReasoningEffort = effort
 	}
+	// prompt_cache_key is a documented OpenAI parameter; compatible servers
+	// ignore unknown fields, but a strict endpoint that rejects it can be
+	// accommodated with ZERO_DISABLE_PROMPT_CACHE_KEY=1.
+	if key := strings.TrimSpace(request.PromptCacheKey); key != "" && !promptCacheKeyDisabled() {
+		mapped.PromptCacheKey = key
+	}
 	if len(request.Tools) > 0 {
 		mapped.Tools = make([]toolDefinition, 0, len(request.Tools))
 		for _, tool := range request.Tools {
@@ -430,6 +444,15 @@ func (provider *Provider) openAIRequest(request zeroruntime.CompletionRequest) c
 	return mapped
 }
 
+// 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
+// explicitly-disabled toggle never flips the behavior it names.
+func promptCacheKeyDisabled() bool {
+	value := strings.TrimSpace(os.Getenv("ZERO_DISABLE_PROMPT_CACHE_KEY"))
+	return value != "" && value != "0" && !strings.EqualFold(value, "false")
+}
+
 // openAIReasoningEffort normalizes a requested effort to a value the OpenAI chat
 // completions API accepts, or "" to omit the field. "none" (and anything else)
 // is dropped rather than risking a 400 on an unrecognized enum.

internal/providers/openai/provider_test.go

@@ -227,6 +227,39 @@ func TestStreamCompletionEmitsReasoningContentDeltas(t *testing.T) {
 	}
 }
 
+func TestStreamCompletionEmitsReasoningAliasDeltas(t *testing.T) {
+	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"reasoning":"Thinking. "}}]}`)
+		writeSSE(w, `{"choices":[{"delta":{"reasoning":"Answering now."}}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning)
+	if len(reasoning) != 2 {
+		t.Fatalf("reasoning events = %#v, want two reasoning deltas", reasoning)
+	}
+	if reasoning[0].Content != "Thinking. " || reasoning[1].Content != "Answering now." {
+		t.Fatalf("unexpected reasoning events: %#v", reasoning)
+	}
+	if text := eventsOfType(events, zeroruntime.StreamEventText); len(text) != 0 {
+		t.Fatalf("reasoning must not emit text events, got %#v", text)
+	}
+}
+
+func TestStreamCompletionPrefersReasoningContentOverAlias(t *testing.T) {
+	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"reasoning_content":"standard","reasoning":"alias"}}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning)
+	if len(reasoning) != 1 || reasoning[0].Content != "standard" {
+		t.Fatalf("reasoning events = %#v, want standard reasoning_content", reasoning)
+	}
+}
+
 func TestStreamCompletionEmitsReasoningBeforeRegularContent(t *testing.T) {
 	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"choices":[{"delta":{"reasoning_content":"Thinking. ","content":"Answer."}}]}`)
@@ -1226,3 +1259,61 @@ func TestOpenAIRequestEmptyContentHandling(t *testing.T) {
 		t.Fatalf("assistant tool calls dropped: %#v", req.Messages[1])
 	}
 }
+
+// TestOpenAIRequestPromptCacheKey locks in prompt_cache_key forwarding: a
+// session-carrying request serializes the key so the backend can route to a
+// replica holding the cached prefix, a keyless request omits the field
+// entirely (strict servers see byte-identical requests to before), and
+// ZERO_DISABLE_PROMPT_CACHE_KEY suppresses it for endpoints that reject it.
+func TestOpenAIRequestPromptCacheKey(t *testing.T) {
+	provider, err := New(Options{Model: "gpt-test"})
+	if err != nil {
+		t.Fatalf("New returned error: %v", err)
+	}
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "hi"}}
+
+	req := provider.openAIRequest(zeroruntime.CompletionRequest{
+		Messages:       messages,
+		PromptCacheKey: "sess_123",
+	})
+	if req.PromptCacheKey != "sess_123" {
+		t.Fatalf("PromptCacheKey = %q, want sess_123", req.PromptCacheKey)
+	}
+	data, err := json.Marshal(req)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if !strings.Contains(string(data), `"prompt_cache_key":"sess_123"`) {
+		t.Fatalf("prompt_cache_key not serialized: %s", data)
+	}
+
+	req = provider.openAIRequest(zeroruntime.CompletionRequest{Messages: messages})
+	if data, err = json.Marshal(req); err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if strings.Contains(string(data), "prompt_cache_key") {
+		t.Fatalf("keyless request must omit prompt_cache_key: %s", data)
+	}
+
+	t.Setenv("ZERO_DISABLE_PROMPT_CACHE_KEY", "1")
+	req = provider.openAIRequest(zeroruntime.CompletionRequest{
+		Messages:       messages,
+		PromptCacheKey: "sess_123",
+	})
+	if req.PromptCacheKey != "" {
+		t.Fatalf("kill switch ignored; PromptCacheKey = %q", req.PromptCacheKey)
+	}
+
+	// Explicitly-falsy kill switch values must NOT disable forwarding — only
+	// truthy values flip the toggle (same parsing as ZERO_FORMAT_ON_WRITE).
+	for _, value := range []string{"0", "false", "FALSE"} {
+		t.Setenv("ZERO_DISABLE_PROMPT_CACHE_KEY", value)
+		req = provider.openAIRequest(zeroruntime.CompletionRequest{
+			Messages:       messages,
+			PromptCacheKey: "sess_123",
+		})
+		if req.PromptCacheKey != "sess_123" {
+			t.Fatalf("ZERO_DISABLE_PROMPT_CACHE_KEY=%q must be a no-op; PromptCacheKey = %q", value, req.PromptCacheKey)
+		}
+	}
+}

internal/providers/openai/types.go

@@ -8,6 +8,11 @@ type chatCompletionRequest struct {
 	ReasoningEffort     string           `json:"reasoning_effort,omitempty"`
 	Stream              bool             `json:"stream"`
 	StreamOptions       *streamOptions   `json:"stream_options,omitempty"`
+	// PromptCacheKey asks the backend to route the request to a replica that
+	// already holds this conversation's prefix in its prompt cache (the OpenAI
+	// `prompt_cache_key` parameter). Omitted when the caller carries no session
+	// identity or when ZERO_DISABLE_PROMPT_CACHE_KEY is set.
+	PromptCacheKey string `json:"prompt_cache_key,omitempty"`
 }
 
 // streamOptions requests the final usage chunk on a streaming response. Without
@@ -78,6 +83,7 @@ type streamChoice struct {
 type streamDelta struct {
 	Content          string                `json:"content"`
 	ReasoningContent string                `json:"reasoning_content"`
+	Reasoning        string                `json:"reasoning"`
 	ToolCalls        []streamToolCallDelta `json:"tool_calls"`
 }
 

internal/providers/providerio/retry.go

@@ -23,12 +23,18 @@ import (
 // (billable) work. Only the INITIAL request is ever in scope; once the response
 // body starts streaming it is never re-issued.
 
-const defaultMaxRetryAttempts = 3
+const defaultMaxRetryAttempts = 6
 
 // maxBackoff caps a single backoff wait so a hostile or buggy Retry-After can't
 // stall the agent for minutes.
 const maxBackoff = 30 * time.Second
 
+// retryBackoffBase is the first wait when the server supplied no Retry-After.
+// Rate-limit windows are measured in seconds, not milliseconds: retrying a 429
+// after 400ms almost always burns the attempt while still limited, so the
+// schedule is 2s, 4s, 8s, 16s, then maxBackoff. A var so tests can shrink it.
+var retryBackoffBase = 2 * time.Second
+
 // SendWithRetry issues an HTTP request, retrying ONLY the safe-to-replay server
 // responses (429 and 503, see ShouldRetryStatus) up to maxAttempts — backing off
 // between tries and honoring a server Retry-After header and context
@@ -110,17 +116,11 @@ func ShouldRetryStatus(code int) bool {
 }
 
 // Backoff waits before retry attempt N (1-based), returning false if the context
-// is cancelled during the wait. The wait is attempt*400ms unless the server
-// supplied a (positive) Retry-After, and is capped at maxBackoff.
+// is cancelled during the wait. A server-supplied (positive) Retry-After wins;
+// otherwise the wait doubles from retryBackoffBase per attempt. Either way the
+// wait is capped at maxBackoff.
 func Backoff(ctx context.Context, attempt int, retryAfter time.Duration) bool {
-	wait := time.Duration(attempt) * 400 * time.Millisecond
-	if retryAfter > 0 {
-		wait = retryAfter
-	}
-	if wait > maxBackoff {
-		wait = maxBackoff
-	}
-	timer := time.NewTimer(wait)
+	timer := time.NewTimer(backoffWait(attempt, retryAfter))
 	defer timer.Stop()
 	select {
 	case <-ctx.Done():
@@ -130,6 +130,27 @@ func Backoff(ctx context.Context, attempt int, retryAfter time.Duration) bool {
 	}
 }
 
+// backoffWait computes the wait before retry attempt N (1-based): Retry-After
+// when supplied, else exponential from retryBackoffBase, both capped at
+// maxBackoff. The exponent is clamped so a large attempt count cannot overflow.
+func backoffWait(attempt int, retryAfter time.Duration) time.Duration {
+	wait := retryAfter
+	if wait <= 0 {
+		exponent := attempt - 1
+		if exponent > 5 {
+			exponent = 5
+		}
+		if exponent < 0 {
+			exponent = 0
+		}
+		wait = retryBackoffBase * time.Duration(1<<exponent)
+	}
+	if wait > maxBackoff {
+		wait = maxBackoff
+	}
+	return wait
+}
+
 // RetryAfter parses a response's Retry-After header (delay-seconds or an HTTP
 // date) into a positive duration, or 0 when absent/unparseable. The result is
 // capped at maxBackoff by Backoff.

internal/providers/providerio/retry_test.go

@@ -94,7 +94,40 @@ func TestBackoffWaitsThenReturnsTrue(t *testing.T) {
 	}
 }
 
+// shrinkBackoff makes retry waits negligible for the duration of a test.
+func shrinkBackoff(t *testing.T) {
+	t.Helper()
+	saved := retryBackoffBase
+	retryBackoffBase = time.Millisecond
+	t.Cleanup(func() { retryBackoffBase = saved })
+}
+
+func TestBackoffWaitSchedule(t *testing.T) {
+	// Without Retry-After the wait doubles per attempt from 2s and caps at 30s;
+	// a supplied Retry-After wins but is capped too.
+	cases := []struct {
+		attempt    int
+		retryAfter time.Duration
+		want       time.Duration
+	}{
+		{1, 0, 2 * time.Second},
+		{2, 0, 4 * time.Second},
+		{3, 0, 8 * time.Second},
+		{4, 0, 16 * time.Second},
+		{5, 0, 30 * time.Second},  // 32s capped
+		{50, 0, 30 * time.Second}, // clamped exponent, no overflow
+		{1, 7 * time.Second, 7 * time.Second},
+		{1, 5 * time.Minute, 30 * time.Second}, // hostile Retry-After capped
+	}
+	for _, c := range cases {
+		if got := backoffWait(c.attempt, c.retryAfter); got != c.want {
+			t.Errorf("backoffWait(%d, %v) = %v, want %v", c.attempt, c.retryAfter, got, c.want)
+		}
+	}
+}
+
 func TestSendWithRetryRetriesThenSucceeds(t *testing.T) {
+	shrinkBackoff(t)
 	var hits int32
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if atomic.AddInt32(&hits, 1) == 1 {
@@ -140,6 +173,7 @@ func TestSendWithRetryReturnsNonRetryableImmediately(t *testing.T) {
 }
 
 func TestSendWithRetryReturnsLastResponseAfterMaxAttempts(t *testing.T) {
+	shrinkBackoff(t)
 	var hits int32
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		atomic.AddInt32(&hits, 1)

internal/repoinfo/repoinfo.go

@@ -54,6 +54,9 @@ type Info struct {
 	AgeDays               *int       `json:"ageDays,omitempty"`
 	Contributors90d       *int       `json:"contributors90d,omitempty"`
 	CommitVelocity30d     *int       `json:"commitVelocity30d,omitempty"`
+	BranchCount           *int       `json:"branchCount,omitempty"`
+	CommitCount           *int       `json:"commitCount,omitempty"`
+	TagCount              *int       `json:"tagCount,omitempty"`
 }
 
 // RunGit runs a git subcommand in dir and returns raw stdout. A non-zero exit
@@ -229,6 +232,43 @@ func Collect(ctx context.Context, opts Options) (Info, error) {
 			info.CommitVelocity30d = &n
 		}
 	}
+	if count, err := run(ctx, dir, "rev-list", "--count", "HEAD"); err == nil {
+		if n, perr := strconv.Atoi(strings.TrimSpace(count)); perr == nil {
+			info.CommitCount = &n
+		}
+	}
+	if branches, err := run(ctx, dir, "branch", "-a"); err == nil {
+		set := map[string]bool{}
+		for _, line := range strings.Split(branches, "\n") {
+			line = strings.TrimSpace(line)
+			if line == "" || strings.Contains(line, "->") {
+				continue
+			}
+			// Strip leading '* ' (current branch) or '+ ' (checked out in another worktree).
+			line = strings.TrimPrefix(line, "* ")
+			line = strings.TrimPrefix(line, "+ ")
+			line = strings.TrimSpace(line)
+			if line == "" || strings.HasPrefix(line, "(") {
+				continue
+			}
+			// If it's a remote branch, extract the branch name.
+			// E.g., remotes/origin/main -> main
+			// E.g., remotes/origin/feature/xyz -> feature/xyz
+			if strings.HasPrefix(line, "remotes/") {
+				line = strings.TrimPrefix(line, "remotes/")
+				if slashIdx := strings.Index(line, "/"); slashIdx != -1 {
+					line = line[slashIdx+1:]
+				}
+			}
+			set[line] = true
+		}
+		n := len(set)
+		info.BranchCount = &n
+	}
+	if tags, err := run(ctx, dir, "tag"); err == nil {
+		n := countNonEmptyLines(tags)
+		info.TagCount = &n
+	}
 
 	return info, nil
 }
@@ -328,3 +368,13 @@ func defaultRunGit(ctx context.Context, dir string, args ...string) (string, err
 	}
 	return stdout.String(), nil
 }
+
+func countNonEmptyLines(s string) int {
+	count := 0
+	for _, line := range strings.Split(s, "\n") {
+		if strings.TrimSpace(line) != "" {
+			count++
+		}
+	}
+	return count
+}

internal/repoinfo/repoinfo_test.go

@@ -43,6 +43,8 @@ func TestCollectCoreMetrics(t *testing.T) {
 			"remote":    "https://github.com/x/y.git\n",
 			"log":       "0\n",
 			"rev-list":  "12\n",
+			"branch":    "  main\n  remotes/origin/HEAD -> origin/main\n  remotes/origin/main\n  remotes/origin/feature\n",
+			"tag":       "v0.1.0\nv0.2.0\n",
 		}, nil),
 	})
 	if err != nil {
@@ -85,6 +87,15 @@ func TestCollectCoreMetrics(t *testing.T) {
 	if info.CommitVelocity30d == nil || *info.CommitVelocity30d != 12 {
 		t.Fatalf("CommitVelocity30d=%v want 12", info.CommitVelocity30d)
 	}
+	if info.CommitCount == nil || *info.CommitCount != 12 {
+		t.Fatalf("CommitCount=%v want 12", info.CommitCount)
+	}
+	if info.BranchCount == nil || *info.BranchCount != 2 {
+		t.Fatalf("BranchCount=%v want 2", info.BranchCount)
+	}
+	if info.TagCount == nil || *info.TagCount != 2 {
+		t.Fatalf("TagCount=%v want 2", info.TagCount)
+	}
 }
 
 func TestCollectContributorsUnique(t *testing.T) {
@@ -252,12 +263,13 @@ func TestCollectNotGitRepo(t *testing.T) {
 }
 
 func TestCollectOnlyReadOnlyLocalSubcommands(t *testing.T) {
-	allowed := map[string]bool{"ls-tree": true, "rev-parse": true, "remote": true, "log": true, "rev-list": true}
+	allowed := map[string]bool{"ls-tree": true, "rev-parse": true, "remote": true, "log": true, "rev-list": true, "branch": true, "tag": true}
 	var used []string
 	_, err := Collect(context.Background(), Options{
 		Now: time.Unix(0, 0),
 		RunGit: fakeGit(t, map[string]string{
 			"ls-tree": lsTree, "rev-parse": "main\n", "remote": "u\n", "log": "0\n", "rev-list": "1\n",
+			"branch": "main\n", "tag": "v0.1.0\n",
 		}, &used),
 	})
 	if err != nil {

internal/sandbox/manager.go

@@ -3,8 +3,9 @@ package sandbox
 import (
 	"errors"
 	"os"
-	"os/exec"
+	"path/filepath"
 	"runtime"
+	"strings"
 )
 
 // errWindowsSandboxNotInitialized is returned only to a caller that explicitly
@@ -121,9 +122,55 @@ func (manager SandboxManager) Backend() Backend {
 	return manager.backend
 }
 
+// isExecutable checks whether a file is executable. On Unix, this checks the
+// execute permission bits. On Windows, Go's os.FileMode does not set execute
+// bits, so we check for common executable extensions instead.
+func isExecutable(fi os.FileInfo) bool {
+	if !fi.Mode().IsRegular() {
+		return false
+	}
+	if runtime.GOOS == "windows" {
+		name := strings.ToLower(fi.Name())
+		return strings.HasSuffix(name, ".exe") ||
+			strings.HasSuffix(name, ".com") ||
+			strings.HasSuffix(name, ".bat") ||
+			strings.HasSuffix(name, ".cmd")
+	}
+	return fi.Mode()&0o111 != 0
+}
+
+// lookupExecutable checks whether a named binary exists and is executable,
+// using os.Stat instead of exec.LookPath. exec.LookPath internally uses the
+// faccessat2 syscall, which is blocked by Android's seccomp filter (SIGSYS).
+func lookupExecutable(name string) (string, error) {
+	if !strings.Contains(name, string(os.PathSeparator)) {
+		path := os.Getenv("PATH")
+		for _, dir := range strings.Split(path, string(os.PathListSeparator)) {
+			if dir == "" {
+				continue // skip empty entries (e.g. trailing colon)
+			}
+			candidate := filepath.Join(dir, name)
+			if fi, err := os.Stat(candidate); err == nil {
+				if isExecutable(fi) {
+					return candidate, nil
+				}
+			}
+		}
+		return "", errors.New("executable file not found in $PATH")
+	}
+	fi, err := os.Stat(name)
+	if err != nil {
+		return "", err
+	}
+	if isExecutable(fi) {
+		return name, nil
+	}
+	return "", errors.New("executable file not found")
+}
+
 func selectPlatformBackend(goos string, lookup func(string) (string, error)) Backend {
 	if lookup == nil {
-		lookup = exec.LookPath
+		lookup = lookupExecutable
 	}
 	switch goos {
 	case "linux":

internal/sandbox/runner_windows_integration_test.go

@@ -190,6 +190,67 @@ func TestWindowsUnelevatedRealSandboxSmoke(t *testing.T) {
 	}
 }
 
+// TestWindowsRestrictedTokenNestedPipeCapture pins the fix in
+// broadenWindowsRestrictedTokenDefaultDacl (windows_token_windows.go): a
+// nested subprocess spawned FROM WITHIN the sandboxed process, with its
+// output captured via a pipe, must succeed. Before the fix, the sandboxed
+// process's own attempt to create a pipe for such a subprocess failed with
+// ERROR_ACCESS_DENIED (Win32 error 5) — the WRITE_RESTRICTED token's extra
+// access check has no restricted-SID match against a freshly created pipe's
+// default security descriptor. This is exactly the pattern any tool that
+// shells out internally and captures output hits (`gh` invoking `git`, for
+// one; cmd.exe's own FOR /F does the identical CreatePipe+CreateProcess
+// dance internally, so it reproduces the bug with no external dependency
+// beyond cmd.exe itself).
+func TestWindowsRestrictedTokenNestedPipeCapture(t *testing.T) {
+	if os.Getenv("ZERO_SANDBOX_REAL_SMOKE") != "1" {
+		t.Skip("set ZERO_SANDBOX_REAL_SMOKE=1 to run real Windows sandbox smoke tests")
+	}
+	runnerExe := realSmokeExecutable(t, "ZERO_WINDOWS_COMMAND_RUNNER_EXE", WindowsSandboxCommandRunnerName)
+
+	root := t.TempDir()
+	sandboxHome := filepath.Join(root, ".zero-sandbox")
+	profile := PermissionProfile{
+		FileSystem: FileSystemPolicy{
+			Kind:                 FileSystemRestricted,
+			ReadRoots:            []string{root},
+			WriteRoots:           []WritableRoot{{Root: root, ProtectedMetadataNames: []string{".git", ".zero", ".agents"}}},
+			IncludePlatformRoots: true,
+			AllowTemp:            true,
+		},
+		Network: NetworkPolicy{Mode: NetworkDeny},
+	}
+	config := WindowsSandboxCommandArgsOptions{
+		SandboxHome:       sandboxHome,
+		CommandCWD:        root,
+		WorkspaceRoots:    []string{root},
+		PermissionProfile: profile,
+		SandboxLevel:      WindowsSandboxLevelUnelevated,
+	}
+
+	marker := filepath.Join(root, "nested-pipe-marker.txt")
+	script := filepath.Join(root, "nested-pipe.cmd")
+	// FOR /F drives cmd.exe's own internal CreatePipe+CreateProcess capture
+	// of the quoted command's output — the exact mechanism this fix targets.
+	// The full path to whoami.exe sidesteps PATH resolution inside the
+	// sandboxed process's minimal environment.
+	scriptBody := "for /F %%i in ('C:\\Windows\\System32\\whoami.exe') do echo %%i> " + marker + "\r\n"
+	if err := os.WriteFile(script, []byte(scriptBody), 0o644); err != nil {
+		t.Fatalf("write nested-pipe script: %v", err)
+	}
+
+	runWindowsRealSmokeCommand(t, runnerExe, config, []string{
+		"cmd.exe", "/d", "/s", "/c", script,
+	}, 0)
+	captured, err := os.ReadFile(marker)
+	if err != nil {
+		t.Fatalf("read nested-pipe marker: %v", err)
+	}
+	if strings.TrimSpace(string(captured)) == "" {
+		t.Fatalf("nested-pipe marker is empty; FOR /F failed to capture the subprocess's output")
+	}
+}
+
 func realSmokeExecutable(t *testing.T, envKey string, fallbackName string) string {
 	t.Helper()
 	if path := os.Getenv(envKey); path != "" {

internal/sandbox/windows_process_windows.go

@@ -87,11 +87,20 @@ func windowsCommandLine(args []string) (string, error) {
 	if len(args) == 0 {
 		return "", errorsNewWindowsProcess("missing command")
 	}
-	out := make([]string, 0, len(args))
 	for _, arg := range args {
 		if strings.ContainsRune(arg, 0) {
 			return "", errorsNewWindowsProcess("command argument contains NUL")
 		}
+	}
+	// Zero's cmd.exe shell invocation needs its final element (the shell
+	// command text) reaching cmd.exe raw, not escaped like a normal argv
+	// element: see WindowsShellCommandLine for why. Every other command this
+	// runner might launch is escaped normally below.
+	if commandLine, ok := windowsShellCommandLineFromArgs(args); ok {
+		return commandLine, nil
+	}
+	out := make([]string, 0, len(args))
+	for _, arg := range args {
 		out = append(out, syscall.EscapeArg(arg))
 	}
 	return strings.Join(out, " "), nil

internal/sandbox/windows_runner.go

@@ -105,6 +105,56 @@ const (
 	WindowsSandboxLevelDisabled   WindowsSandboxLevel = "disabled"
 )
 
+// WindowsShellArgs returns cmd.exe's argv (after the executable name) for
+// running commandText as a shell command on Windows: no wrapping quotes and
+// no /S, so commandText reaches cmd.exe's own /C remainder parsing exactly as
+// written. See WindowsShellCommandLine for why that matters. Used both as
+// CommandSpec.Args for the plain (unwrapped) path and, via
+// windowsShellCommandLineFromArgs, to recognize this same shape once it has
+// round-tripped through the sandboxed runner's CLI argv.
+func WindowsShellArgs(commandText string) []string {
+	return []string{"/d", "/c", commandText}
+}
+
+// WindowsShellCommandLine builds the raw command line a Windows CreateProcess
+// call should receive to run commandText as a shell command: "cmd.exe /d /c "
+// followed by commandText completely unescaped.
+//
+// cmd.exe's own /C remainder parsing is not CommandLineToArgvW-compatible: it
+// applies its own quote-stripping heuristic (documented in `cmd /?`) whenever
+// the remainder starts with a literal quote, and that heuristic does not
+// understand backslash-escaped inner quotes the way a normal argv consumer
+// would. A command like `python -c "print(15 / 3)"` passed through the
+// standard exec.Cmd Args mechanism gets wrapped in an outer pair of quotes
+// (because it contains spaces) with its own quotes escaped as \" — exactly
+// the shape that heuristic corrupts, stripping the outer quotes but leaving
+// the literal backslashes behind. Passing commandText through raw, exactly as
+// the model wrote it, means cmd.exe parses it the same way it would if typed
+// directly at an interactive prompt: a leading quoted executable path like
+// `"C:\Program Files\foo.exe" arg` is preserved correctly by cmd.exe's own
+// quote-preserving special case, and a command with no leading quote at all,
+// like the python example above, is never touched by the stripping heuristic
+// in the first place.
+func WindowsShellCommandLine(commandText string) string {
+	return "cmd.exe /d /c " + commandText
+}
+
+// windowsShellCommandLineFromArgs reports whether args is exactly
+// ["cmd.exe"]+WindowsShellArgs(text) for some text and, if so, returns
+// WindowsShellCommandLine's result for that text. The sandboxed runner path
+// only has the deserialized argv (WindowsSandboxCommandConfig.Command) to
+// work from, not the original commandText, so it recovers the raw command
+// line by recognizing the shape instead.
+func windowsShellCommandLineFromArgs(args []string) (string, bool) {
+	if len(args) != 4 {
+		return "", false
+	}
+	if !strings.EqualFold(args[0], "cmd.exe") || args[1] != "/d" || args[2] != "/c" {
+		return "", false
+	}
+	return WindowsShellCommandLine(args[3]), true
+}
+
 type WindowsSandboxCommandArgsOptions struct {
 	SandboxHome       string
 	CommandCWD        string

internal/sandbox/windows_runner_test.go

@@ -147,3 +147,56 @@ func TestRunWindowsSandboxCommandRunnerRejectsInvalidArgs(t *testing.T) {
 		t.Fatalf("stderr = %q, want runner name", stderr.String())
 	}
 }
+
+func TestWindowsShellArgsHasNoSAndKeepsCommandTextWhole(t *testing.T) {
+	args := WindowsShellArgs(`python -c "print(15 / 3)"`)
+	want := []string{"/d", "/c", `python -c "print(15 / 3)"`}
+	if len(args) != len(want) {
+		t.Fatalf("args = %#v, want %#v", args, want)
+	}
+	for index := range want {
+		if args[index] != want[index] {
+			t.Fatalf("args = %#v, want %#v", args, want)
+		}
+	}
+}
+
+func TestWindowsShellCommandLineLeavesCommandTextUnescaped(t *testing.T) {
+	// The whole point of this function: commandText's own quotes must reach
+	// cmd.exe exactly as written, not re-quoted or backslash-escaped the way
+	// a normal argv-escaping function (like syscall.EscapeArg) would treat an
+	// argument containing spaces and embedded quotes.
+	commandText := `python -c "print(15 / 3)"`
+	got := WindowsShellCommandLine(commandText)
+	want := `cmd.exe /d /c python -c "print(15 / 3)"`
+	if got != want {
+		t.Fatalf("WindowsShellCommandLine = %q, want %q", got, want)
+	}
+}
+
+func TestWindowsShellCommandLineFromArgsRecognizesTheShellShape(t *testing.T) {
+	commandLine, ok := windowsShellCommandLineFromArgs([]string{"cmd.exe", "/d", "/c", `echo "hi"`})
+	if !ok {
+		t.Fatal("windowsShellCommandLineFromArgs did not recognize the shell shape")
+	}
+	if want := `cmd.exe /d /c echo "hi"`; commandLine != want {
+		t.Fatalf("commandLine = %q, want %q", commandLine, want)
+	}
+}
+
+func TestWindowsShellCommandLineFromArgsRejectsOtherShapes(t *testing.T) {
+	cases := [][]string{
+		nil,
+		{"cmd.exe"},
+		{"cmd.exe", "/d", "/c"},
+		{"cmd.exe", "/d", "/c", "echo hi", "extra"},
+		{"git.exe", "/d", "/c", "echo hi"},
+		{"cmd.exe", "/s", "/c", "echo hi"},
+		{"cmd.exe", "/d", "/k", "echo hi"},
+	}
+	for _, args := range cases {
+		if _, ok := windowsShellCommandLineFromArgs(args); ok {
+			t.Fatalf("windowsShellCommandLineFromArgs(%#v) matched, want no match", args)
+		}
+	}
+}

internal/sandbox/windows_token_windows.go

@@ -21,6 +21,11 @@ const (
 
 var procCreateRestrictedToken = windows.NewLazySystemDLL("advapi32.dll").NewProc("CreateRestrictedToken")
 
+// procSetEntriesInAclW merges new ACEs into an existing ACL. Not wrapped by
+// x/sys/windows (its internal setEntriesInAcl is unexported), so it is
+// declared directly the same way procCreateRestrictedToken is.
+var procSetEntriesInAclW = windows.NewLazySystemDLL("advapi32.dll").NewProc("SetEntriesInAclW")
+
 type windowsLocalSID struct {
 	sid *windows.SID
 }
@@ -122,9 +127,139 @@ func createWindowsRestrictedTokenFromBase(base windows.Token, capabilitySIDs []w
 		_ = restricted.Close()
 		return 0, err
 	}
+	if err := broadenWindowsRestrictedTokenDefaultDacl(restricted, sidFromBytes(logonSID)); err != nil {
+		_ = restricted.Close()
+		return 0, err
+	}
+	runtime.KeepAlive(logonSID)
 	return restricted, nil
 }
 
+// broadenWindowsRestrictedTokenDefaultDacl adds the token's own logon SID to
+// its default DACL (the DACL applied to any kernel object the token's process
+// creates WITHOUT an explicit security descriptor, e.g. anonymous pipes,
+// events, mutexes).
+//
+// A WRITE_RESTRICTED token requires a WRITE-type access check to match BOTH
+// the normal enabled-SID grant AND a SEPARATE grant to one of the token's
+// restricted SIDs. CreateRestrictedToken does not touch the default DACL it
+// inherits from the base token, so newly created objects only carry ACEs for
+// the base token's normal identity (owner, SYSTEM, Administrators) — none of
+// which are in the restricted-SID list — so the second check fails. This
+// breaks anything that creates a pipe/handle for itself with default
+// security, which is exactly what any tool using a language runtime's
+// "spawn a subprocess and capture its output" primitive does (Go's os/exec,
+// and by extension `gh`, and many other CLI tools that shell out
+// internally): CreatePipe() succeeds, but the creating process's own token
+// then fails to actually read/write the handle it just made.
+//
+// The logon SID is already unconditionally present in this token's
+// restricted-SID list (see createWindowsRestrictedTokenFromBase), so adding
+// it here — and ONLY it — to the default DACL is the minimal change that
+// satisfies the WRITE_RESTRICTED extra check for self-created objects. It
+// does not touch file-system access: NTFS write-jailing is enforced by the
+// explicit ACL grants applied to workspace paths (windows_acl.go), which are
+// created WITH an explicit security descriptor and never consult the token's
+// default DACL. The exposure is bounded to the same logon session (other
+// processes running as the same signed-in user could, in principle, open a
+// NAMED kernel object the sandboxed process creates with default security;
+// anonymous pipes — the actual object type this fixes — have no name and are
+// only reachable via an inherited handle, so this is a no-op for them).
+func broadenWindowsRestrictedTokenDefaultDacl(token windows.Token, logonSID *windows.SID) error {
+	oldDacl, oldDaclBuf, err := windowsTokenDefaultDacl(token)
+	if err != nil {
+		return fmt.Errorf("read token default DACL: %w", err)
+	}
+	// The second (WRITE-type) access check only needs a read/write match; a
+	// full GENERIC_ALL grant (which also implies DELETE, WRITE_DAC, and
+	// WRITE_OWNER) is broader than the pipe/event/mutex use case this exists
+	// for actually requires.
+	access := []windows.EXPLICIT_ACCESS{{
+		AccessPermissions: windows.GENERIC_READ | windows.GENERIC_WRITE,
+		AccessMode:        windows.GRANT_ACCESS,
+		Inheritance:       windows.NO_INHERITANCE,
+		Trustee: windows.TRUSTEE{
+			MultipleTrusteeOperation: windows.NO_MULTIPLE_TRUSTEE,
+			TrusteeForm:              windows.TRUSTEE_IS_SID,
+			TrusteeType:              windows.TRUSTEE_IS_USER,
+			TrusteeValue:             windows.TrusteeValueFromSID(logonSID),
+		},
+	}}
+	newDacl, err := setEntriesInACL(access, oldDacl)
+	// oldDacl points INTO oldDaclBuf (see windowsTokenDefaultDacl) and is
+	// dereferenced by native code inside setEntriesInACL, above. Nothing in Go
+	// references oldDaclBuf once windowsTokenDefaultDacl returned, so without
+	// this KeepAlive it is eligible for GC the moment an allocation in this
+	// function (the EXPLICIT_ACCESS literal, TrusteeValueFromSID) hits a
+	// safepoint, before the native call ever reads it - a real, if narrow,
+	// use-after-free window in security-boundary code.
+	runtime.KeepAlive(oldDaclBuf)
+	if err != nil {
+		return fmt.Errorf("merge token default DACL: %w", err)
+	}
+	defer func() { _, _ = windows.LocalFree(windows.Handle(unsafe.Pointer(newDacl))) }()
+	if err := windowsSetTokenDefaultDacl(token, newDacl); err != nil {
+		return fmt.Errorf("set token default DACL: %w", err)
+	}
+	return nil
+}
+
+type windowsTokenDefaultDaclInfo struct {
+	DefaultDacl *windows.ACL
+}
+
+// windowsTokenDefaultDacl returns the token's current default DACL along with
+// the buffer backing it. TOKEN_DEFAULT_DACL embeds the ACL data inside the
+// same allocation GetTokenInformation fills, so the returned *windows.ACL is
+// only valid for as long as the returned buffer stays reachable: the caller
+// must runtime.KeepAlive(buf) until every native call that dereferences the
+// ACL has completed. A KeepAlive inside this function, ending at its own
+// return, would not protect any use by the caller after that.
+func windowsTokenDefaultDacl(token windows.Token) (*windows.ACL, []byte, error) {
+	var size uint32
+	err := windows.GetTokenInformation(token, windows.TokenDefaultDacl, nil, 0, &size)
+	if err == nil || err != windows.ERROR_INSUFFICIENT_BUFFER {
+		return nil, nil, fmt.Errorf("size TokenDefaultDacl: %w", err)
+	}
+	buf := make([]byte, size)
+	if err := windows.GetTokenInformation(token, windows.TokenDefaultDacl, &buf[0], size, &size); err != nil {
+		return nil, nil, fmt.Errorf("get TokenDefaultDacl: %w", err)
+	}
+	info := (*windowsTokenDefaultDaclInfo)(unsafe.Pointer(&buf[0]))
+	return info.DefaultDacl, buf, nil
+}
+
+func windowsSetTokenDefaultDacl(token windows.Token, dacl *windows.ACL) error {
+	info := windowsTokenDefaultDaclInfo{DefaultDacl: dacl}
+	err := windows.SetTokenInformation(token, windows.TokenDefaultDacl, (*byte)(unsafe.Pointer(&info)), uint32(unsafe.Sizeof(info)))
+	runtime.KeepAlive(info)
+	return err
+}
+
+// setEntriesInACL merges entries into oldACL via Win32 SetEntriesInAclW,
+// returning a newly allocated ACL the caller must free with windows.LocalFree.
+func setEntriesInACL(entries []windows.EXPLICIT_ACCESS, oldACL *windows.ACL) (*windows.ACL, error) {
+	if len(entries) == 0 {
+		return nil, errors.New("setEntriesInACL requires at least one entry")
+	}
+	var newACL *windows.ACL
+	// SetEntriesInAclW reports failure directly in its return value (ERROR_SUCCESS
+	// on success, a Win32 error code otherwise); unlike many Win32 APIs it does not
+	// set the thread's last-error, so the syscall trio's second/third return
+	// values (which surface GetLastError) are not meaningful here.
+	ret, _, _ := procSetEntriesInAclW.Call(
+		uintptr(len(entries)),
+		uintptr(unsafe.Pointer(&entries[0])),
+		uintptr(unsafe.Pointer(oldACL)),
+		uintptr(unsafe.Pointer(&newACL)),
+	)
+	runtime.KeepAlive(entries)
+	if ret != 0 {
+		return nil, fmt.Errorf("SetEntriesInAclW: %w", syscall.Errno(ret))
+	}
+	return newACL, nil
+}
+
 func copyWindowsLogonSID(token windows.Token) ([]byte, error) {
 	groups, err := token.GetTokenGroups()
 	if err == nil {

internal/sessions/append_events_test.go

@@ -0,0 +1,233 @@
+package sessions
+
+import (
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestStoreAppendEventsBatchesSequencesAndMetadata(t *testing.T) {
+	store := NewStore(StoreOptions{RootDir: t.TempDir(), Now: sequenceClock([]time.Time{
+		time.Date(2026, 6, 4, 15, 0, 0, 0, time.UTC),
+		time.Date(2026, 6, 4, 15, 0, 1, 0, time.UTC),
+		time.Date(2026, 6, 4, 15, 0, 2, 0, time.UTC),
+		time.Date(2026, 6, 4, 15, 0, 3, 0, time.UTC),
+	})})
+	session, err := store.Create(CreateInput{SessionID: "batch"})
+	if err != nil {
+		t.Fatalf("Create returned error: %v", err)
+	}
+
+	appended, err := store.AppendEvents(session.SessionID, []AppendEventInput{
+		{Type: EventMessage, Payload: map[string]any{"role": "user", "content": "one"}},
+		{Type: EventToolCall, Payload: map[string]any{"id": "call_1", "name": "read_file"}},
+		{Type: EventToolResult, Payload: map[string]any{"toolCallId": "call_1", "status": "ok"}},
+	})
+	if err != nil {
+		t.Fatalf("AppendEvents returned error: %v", err)
+	}
+	if len(appended) != 3 {
+		t.Fatalf("expected 3 appended events, got %#v", appended)
+	}
+	for index, event := range appended {
+		wantSequence := index + 1
+		if event.Sequence != wantSequence || event.ID != fmt.Sprintf("batch:%d", wantSequence) {
+			t.Fatalf("event identity mismatch at %d: %#v", index, event)
+		}
+	}
+	if appended[0].CreatedAt != "2026-06-04T15:00:01Z" || appended[2].CreatedAt != "2026-06-04T15:00:03Z" {
+		t.Fatalf("unexpected event timestamps: %#v", appended)
+	}
+
+	loaded, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	if loaded == nil || loaded.EventCount != 3 || loaded.LastEventType != EventToolResult || loaded.UpdatedAt != appended[2].CreatedAt {
+		t.Fatalf("metadata not updated from final batch event: %#v", loaded)
+	}
+	events, err := store.ReadEvents(session.SessionID)
+	if err != nil {
+		t.Fatalf("ReadEvents returned error: %v", err)
+	}
+	if !reflect.DeepEqual(eventTypesForTest(events), []EventType{EventMessage, EventToolCall, EventToolResult}) {
+		t.Fatalf("unexpected event types: %#v", events)
+	}
+}
+
+func TestStoreAppendEventsEmptyBatchDoesNotRewriteMetadata(t *testing.T) {
+	store := NewStore(StoreOptions{RootDir: t.TempDir(), Now: fixedClock("2026-06-04T15:10:00Z")})
+	session, err := store.Create(CreateInput{SessionID: "empty_batch", Title: "unchanged"})
+	if err != nil {
+		t.Fatalf("Create returned error: %v", err)
+	}
+
+	appended, err := store.AppendEvents(session.SessionID, nil)
+	if err != nil {
+		t.Fatalf("AppendEvents empty returned error: %v", err)
+	}
+	if len(appended) != 0 {
+		t.Fatalf("empty AppendEvents returned %#v", appended)
+	}
+	loaded, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	if loaded == nil || !reflect.DeepEqual(*loaded, session) {
+		t.Fatalf("empty AppendEvents rewrote metadata: before=%#v after=%#v", session, loaded)
+	}
+}
+
+func TestStoreAppendEventsInvalidPayloadDoesNotPartiallyAppend(t *testing.T) {
+	store := NewStore(StoreOptions{RootDir: t.TempDir(), Now: fixedClock("2026-06-04T15:20:00Z")})
+	session, err := store.Create(CreateInput{SessionID: "invalid_batch"})
+	if err != nil {
+		t.Fatalf("Create returned error: %v", err)
+	}
+
+	_, err = store.AppendEvents(session.SessionID, []AppendEventInput{
+		{Type: EventMessage, Payload: map[string]any{"content": "valid first"}},
+		{Type: EventMessage, Payload: json.RawMessage(`{"broken"`)},
+	})
+	if err == nil || !strings.Contains(err.Error(), "invalid raw JSON payload") {
+		t.Fatalf("expected invalid raw payload error, got %v", err)
+	}
+	events, err := store.ReadEvents(session.SessionID)
+	if err != nil {
+		t.Fatalf("ReadEvents returned error: %v", err)
+	}
+	if len(events) != 0 {
+		t.Fatalf("invalid batch should not append partial events: %#v", events)
+	}
+	loaded, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	if loaded == nil || loaded.EventCount != 0 || loaded.LastEventType != "" {
+		t.Fatalf("invalid batch should not update metadata: %#v", loaded)
+	}
+}
+
+func TestStoreAppendEventsSequencesAfterMetadataLag(t *testing.T) {
+	store := NewStore(StoreOptions{RootDir: t.TempDir(), Now: fixedClock("2026-06-04T15:30:00Z")})
+	session, err := store.Create(CreateInput{SessionID: "lagging"})
+	if err != nil {
+		t.Fatalf("Create returned error: %v", err)
+	}
+	if _, err := store.AppendEvent(session.SessionID, AppendEventInput{Type: EventMessage, Payload: map[string]any{"content": "already durable"}}); err != nil {
+		t.Fatalf("AppendEvent returned error: %v", err)
+	}
+	lagging, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	lagging.EventCount = 0
+	lagging.LastEventType = ""
+	if err := store.writeMetadata(*lagging); err != nil {
+		t.Fatalf("writeMetadata returned error: %v", err)
+	}
+
+	appended, err := store.AppendEvents(session.SessionID, []AppendEventInput{
+		{Type: EventToolCall, Payload: map[string]any{"id": "call"}},
+		{Type: EventToolResult, Payload: map[string]any{"toolCallId": "call"}},
+	})
+	if err != nil {
+		t.Fatalf("AppendEvents returned error: %v", err)
+	}
+	if len(appended) != 2 || appended[0].Sequence != 2 || appended[1].Sequence != 3 {
+		t.Fatalf("batch did not sequence after durable log tail: %#v", appended)
+	}
+	loaded, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	if loaded == nil || loaded.EventCount != 3 || loaded.LastEventType != EventToolResult {
+		t.Fatalf("metadata not repaired by batch append: %#v", loaded)
+	}
+}
+
+func TestStoreAppendEventsSerializesConcurrentBatches(t *testing.T) {
+	store := NewStore(StoreOptions{RootDir: t.TempDir(), Now: fixedClock("2026-06-04T15:40:00Z")})
+	session, err := store.Create(CreateInput{SessionID: "concurrent_batches"})
+	if err != nil {
+		t.Fatalf("Create returned error: %v", err)
+	}
+
+	const writers = 8
+	const perBatch = 3
+	var wg sync.WaitGroup
+	errs := make(chan error, writers)
+	for writer := 0; writer < writers; writer++ {
+		wg.Add(1)
+		go func(writer int) {
+			defer wg.Done()
+			appended, err := store.AppendEvents(session.SessionID, []AppendEventInput{
+				{Type: EventMessage, Payload: map[string]int{"writer": writer, "index": 0}},
+				{Type: EventMessage, Payload: map[string]int{"writer": writer, "index": 1}},
+				{Type: EventMessage, Payload: map[string]int{"writer": writer, "index": 2}},
+			})
+			if err != nil {
+				errs <- err
+				return
+			}
+			if len(appended) != perBatch {
+				errs <- fmt.Errorf("writer %d appended %d events", writer, len(appended))
+				return
+			}
+			for index := 1; index < len(appended); index++ {
+				if appended[index].Sequence != appended[index-1].Sequence+1 {
+					errs <- fmt.Errorf("writer %d batch was interleaved: %#v", writer, appended)
+					return
+				}
+			}
+			errs <- nil
+		}(writer)
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		if err != nil {
+			t.Fatalf("AppendEvents returned error: %v", err)
+		}
+	}
+
+	events, err := store.ReadEvents(session.SessionID)
+	if err != nil {
+		t.Fatalf("ReadEvents returned error: %v", err)
+	}
+	total := writers * perBatch
+	if len(events) != total {
+		t.Fatalf("expected %d events, got %d", total, len(events))
+	}
+	seen := map[int]bool{}
+	for _, event := range events {
+		if seen[event.Sequence] {
+			t.Fatalf("duplicate sequence %d in %#v", event.Sequence, events)
+		}
+		seen[event.Sequence] = true
+	}
+	for sequence := 1; sequence <= total; sequence++ {
+		if !seen[sequence] {
+			t.Fatalf("missing sequence %d in %#v", sequence, events)
+		}
+	}
+	loaded, err := store.Get(session.SessionID)
+	if err != nil {
+		t.Fatalf("Get returned error: %v", err)
+	}
+	if loaded == nil || loaded.EventCount != total || loaded.LastEventType != EventMessage {
+		t.Fatalf("metadata not updated after concurrent batches: %#v", loaded)
+	}
+}
+
+func eventTypesForTest(events []Event) []EventType {
+	types := make([]EventType, 0, len(events))
+	for _, event := range events {
+		types = append(types, event.Type)
+	}
+	return types
+}

internal/sessions/store.go

@@ -153,6 +153,11 @@ type AppendEventInput struct {
 	Payload any
 }
 
+type preparedAppendEvent struct {
+	Type    EventType
+	Payload json.RawMessage
+}
+
 type RecordSpecInput struct {
 	SpecID              string
 	SpecFilePath        string
@@ -426,7 +431,7 @@ func (store *Store) Fork(parentSessionID string, input ForkInput) (Metadata, err
 	if err != nil {
 		return Metadata{}, err
 	}
-	copied := 0
+	copyInputs := []AppendEventInput{}
 	for _, event := range events {
 		// Do NOT copy usage accounting into the fork. It already counted against the
 		// parent, and a usage report that aggregates the parent and the fork would
@@ -435,11 +440,12 @@ func (store *Store) Fork(parentSessionID string, input ForkInput) (Metadata, err
 		if event.Type == EventUsage {
 			continue
 		}
-		if _, err := store.AppendEvent(fork.SessionID, AppendEventInput{Type: event.Type, Payload: event.Payload}); err != nil {
-			return Metadata{}, err
-		}
-		copied++
+		copyInputs = append(copyInputs, AppendEventInput{Type: event.Type, Payload: event.Payload})
 	}
+	if _, err := store.AppendEvents(fork.SessionID, copyInputs); err != nil {
+		return Metadata{}, err
+	}
+	copied := len(copyInputs)
 	// Copy the parent's content-addressed checkpoint blobs into the fork so the
 	// copied EventSessionCheckpoint events resolve to real blobs and a rewind on
 	// the fork can restore file content (otherwise rewind reads missing blobs
@@ -513,19 +519,34 @@ func (store *Store) RecordSpec(sessionID string, input RecordSpecInput) (Metadat
 }
 
 func (store *Store) AppendEvent(sessionID string, input AppendEventInput) (Event, error) {
-	if !ValidSessionID(sessionID) {
-		return Event{}, fmt.Errorf("invalid zero session id %q", sessionID)
-	}
-	if strings.TrimSpace(string(input.Type)) == "" {
-		return Event{}, fmt.Errorf("zero session event type is required")
-	}
-	unlock, err := store.lockSession(sessionID)
+	events, err := store.AppendEvents(sessionID, []AppendEventInput{input})
 	if err != nil {
 		return Event{}, err
 	}
+	return events[0], nil
+}
+
+// AppendEvents appends a batch of events atomically under one session lock and
+// durability pass. The returned events have contiguous sequences in input order.
+// An empty batch is a no-op and does not rewrite metadata.
+func (store *Store) AppendEvents(sessionID string, inputs []AppendEventInput) ([]Event, error) {
+	if !ValidSessionID(sessionID) {
+		return nil, fmt.Errorf("invalid zero session id %q", sessionID)
+	}
+	prepared, err := prepareAppendEventInputs(inputs)
+	if err != nil {
+		return nil, err
+	}
+	if len(prepared) == 0 {
+		return []Event{}, nil
+	}
+	unlock, err := store.lockSession(sessionID)
+	if err != nil {
+		return nil, err
+	}
 	defer unlock()
 
-	return store.appendEventLocked(sessionID, input)
+	return store.appendPreparedEventsLocked(sessionID, prepared)
 }
 
 // AppendEventUnlessExists appends input only when exists(currentEvents) is false,
@@ -572,14 +593,27 @@ func (store *Store) AppendEventUnlessExists(sessionID string, input AppendEventI
 // the single lock they already hold, instead of re-locking (which would deadlock
 // on the non-reentrant in-process mutex).
 func (store *Store) appendEventLocked(sessionID string, input AppendEventInput) (Event, error) {
-	session, err := store.readMetadata(sessionID)
+	prepared, err := prepareAppendEventInputs([]AppendEventInput{input})
 	if err != nil {
 		return Event{}, err
 	}
-	payload, err := rawPayload(input.Payload)
+	events, err := store.appendPreparedEventsLocked(sessionID, prepared)
 	if err != nil {
 		return Event{}, err
 	}
+	return events[0], nil
+}
+
+// appendPreparedEventsLocked appends pre-validated events WITHOUT acquiring the
+// session lock. The caller MUST already hold store.lockSession(sessionID).
+func (store *Store) appendPreparedEventsLocked(sessionID string, inputs []preparedAppendEvent) ([]Event, error) {
+	if len(inputs) == 0 {
+		return []Event{}, nil
+	}
+	session, err := store.readMetadata(sessionID)
+	if err != nil {
+		return nil, err
+	}
 	sequence := session.EventCount + 1
 	// The events log is the source of truth for sequencing. metadata.EventCount is
 	// persisted separately (after the event line below), so a crash between the two
@@ -590,46 +624,54 @@ func (store *Store) appendEventLocked(sessionID string, input AppendEventInput)
 	if logSeq, err := store.lastEventSequence(sessionID); err == nil && logSeq >= sequence {
 		sequence = logSeq + 1
 	}
-	timestamp := store.timestamp()
-	event := Event{
-		ID:        fmt.Sprintf("%s:%d", sessionID, sequence),
-		SessionID: sessionID,
-		Sequence:  sequence,
-		Type:      input.Type,
-		CreatedAt: timestamp,
-		Payload:   payload,
-	}
-	data, err := json.Marshal(event)
-	if err != nil {
-		return Event{}, fmt.Errorf("encode zero session event: %w", err)
+	events := make([]Event, 0, len(inputs))
+	var batch bytes.Buffer
+	for index, input := range inputs {
+		timestamp := store.timestamp()
+		event := Event{
+			ID:        fmt.Sprintf("%s:%d", sessionID, sequence+index),
+			SessionID: sessionID,
+			Sequence:  sequence + index,
+			Type:      input.Type,
+			CreatedAt: timestamp,
+			Payload:   input.Payload,
+		}
+		data, err := json.Marshal(event)
+		if err != nil {
+			return nil, fmt.Errorf("encode zero session event: %w", err)
+		}
+		batch.Write(data)
+		batch.WriteByte('\n')
+		events = append(events, event)
 	}
 	file, err := os.OpenFile(store.eventsPath(sessionID), os.O_WRONLY|os.O_APPEND|os.O_CREATE, 0o600)
 	if err != nil {
-		return Event{}, fmt.Errorf("append zero session event: %w", err)
+		return nil, fmt.Errorf("append zero session event: %w", err)
 	}
-	if _, err := file.Write(append(data, '\n')); err != nil {
+	if _, err := file.Write(batch.Bytes()); err != nil {
 		_ = file.Close()
-		return Event{}, fmt.Errorf("append zero session event: %w", err)
+		return nil, fmt.Errorf("append zero session event: %w", err)
 	}
-	// fsync the event before reporting success: the derived metadata.json IS
+	// fsync the events before reporting success: the derived metadata.json IS
 	// fsync'd (writeMetadata), so without this a crash after the metadata flush
 	// but before the events.jsonl page reaches disk leaves EventCount ahead of the
-	// durable log — silently losing the just-appended event (incl. the checkpoint
+	// durable log — silently losing just-appended events (incl. checkpoints
 	// that /rewind targets). Make the log at least as durable as its metadata. (AUDIT-M12)
 	if err := file.Sync(); err != nil {
 		_ = file.Close()
-		return Event{}, fmt.Errorf("sync zero session event: %w", err)
+		return nil, fmt.Errorf("sync zero session event: %w", err)
 	}
 	if err := file.Close(); err != nil {
-		return Event{}, fmt.Errorf("close zero session event file: %w", err)
+		return nil, fmt.Errorf("close zero session event file: %w", err)
 	}
-	session.UpdatedAt = timestamp
-	session.EventCount = sequence
-	session.LastEventType = input.Type
+	last := events[len(events)-1]
+	session.UpdatedAt = last.CreatedAt
+	session.EventCount = last.Sequence
+	session.LastEventType = last.Type
 	if err := store.writeMetadata(session); err != nil {
-		return Event{}, err
+		return nil, err
 	}
-	return event, nil
+	return events, nil
 }
 
 // UpdateTitle replaces a session's Title and returns the updated metadata. It is
@@ -964,6 +1006,24 @@ func rawPayload(payload any) (json.RawMessage, error) {
 	return data, nil
 }
 
+func prepareAppendEventInputs(inputs []AppendEventInput) ([]preparedAppendEvent, error) {
+	prepared := make([]preparedAppendEvent, 0, len(inputs))
+	for _, input := range inputs {
+		if strings.TrimSpace(string(input.Type)) == "" {
+			return nil, fmt.Errorf("zero session event type is required")
+		}
+		payload, err := rawPayload(input.Payload)
+		if err != nil {
+			return nil, err
+		}
+		prepared = append(prepared, preparedAppendEvent{
+			Type:    input.Type,
+			Payload: payload,
+		})
+	}
+	return prepared, nil
+}
+
 func envValue(env map[string]string, key string) string {
 	if env != nil {
 		return env[key]

internal/specialist/exec.go

@@ -25,6 +25,7 @@ import (
 const (
 	sessionTagSpecialist     = "specialist"
 	promptFileThresholdBytes = 4 * 1024
+	maxSpecialistDepth       = 8 // hard cap to prevent infinite recursion/resource exhaustion via Task
 )
 
 const SessionTagSpecialist = sessionTagSpecialist
@@ -218,6 +219,13 @@ func (executor Executor) Run(ctx context.Context, params TaskParameters, options
 	if options.CurrentDepth < 0 {
 		return ExecResult{}, fmt.Errorf("current depth cannot be negative")
 	}
+	// >= (not >): this Run call spawns a CHILD at options.CurrentDepth+1, so a
+	// parent already AT the cap must still be rejected here rather than being
+	// allowed to launch one more level before the child's own next call trips
+	// the guard.
+	if options.CurrentDepth >= maxSpecialistDepth {
+		return ExecResult{}, fmt.Errorf("spawning a specialist at depth %d would exceed maximum nesting depth %d", options.CurrentDepth+1, maxSpecialistDepth)
+	}
 	if strings.TrimSpace(params.Prompt) == "" {
 		return ExecResult{}, fmt.Errorf("specialist prompt is required")
 	}

internal/specialist/exec_test.go

@@ -1,6 +1,7 @@
 package specialist
 
 import (
+	"context"
 	"os"
 	"path/filepath"
 	"reflect"
@@ -336,6 +337,39 @@ func TestBuildArgsRejectsInvalidInputs(t *testing.T) {
 	}
 }
 
+func TestRunRejectsDepthExceedingMax(t *testing.T) {
+	_, err := (Executor{}).Run(context.Background(), TaskParameters{
+		Prompt: "hi",
+	}, TaskRunOptions{CurrentDepth: maxSpecialistDepth + 1})
+	if err == nil || !strings.Contains(err.Error(), "depth") {
+		t.Fatalf("Run error = %v, want depth error", err)
+	}
+}
+
+// TestRunRejectsDepthAtMax covers the boundary: a parent already AT the cap
+// must be rejected too, since this Run call would launch a child one level
+// past it (--depth CurrentDepth+1). Only checking ">" here would let that
+// child start before the guard ever caught it. The guard sits before the
+// fresh/resume branch in Run, so both call shapes must be proven to reject
+// here rather than falling through to runFresh/runResume.
+func TestRunRejectsDepthAtMax(t *testing.T) {
+	tests := []struct {
+		name   string
+		params TaskParameters
+	}{
+		{name: "fresh", params: TaskParameters{Prompt: "hi"}},
+		{name: "resume", params: TaskParameters{Prompt: "hi", Resume: "child_session"}},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := (Executor{}).Run(context.Background(), tc.params, TaskRunOptions{CurrentDepth: maxSpecialistDepth})
+			if err == nil || !strings.Contains(err.Error(), "depth") {
+				t.Fatalf("Run error = %v, want depth error", err)
+			}
+		})
+	}
+}
+
 func TestBuildArgsRejectsInvalidSessionIDs(t *testing.T) {
 	_, err := (Executor{NewSessionID: func() (string, error) { return "../escape", nil }}).BuildArgs(BuildArgsInput{
 		Prompt: "hi",

internal/streamjson/streamjson.go

@@ -21,6 +21,7 @@ type EventType string
 const (
 	EventRunStart           EventType = "run_start"
 	EventText               EventType = "text"
+	EventReasoning          EventType = "reasoning"
 	EventToolCall           EventType = "tool_call"
 	EventPermission         EventType = "permission"
 	EventPermissionRequest  EventType = "permission_request"

internal/tools/ask_user.go

@@ -272,15 +272,36 @@ func coerceAskUserOptions(optionsValue, descsValue any) (labels []string, descri
 // FormatAskUserAnswers renders question/answer pairs into a clear, model-readable
 // block. Missing answers are surfaced explicitly so the model never silently
 // treats an unanswered question as answered.
+//
+// It distinguishes two shapes of "empty" the model would otherwise conflate: a
+// wholesale dismissal (the user closed the prompt without answering ANYTHING) is
+// flagged up front as a skip, so the model doesn't invent a default; a single
+// blank field amid other answers is marked "(left blank)".
 func FormatAskUserAnswers(questions []AskUserQuestion, answers []string) string {
-	lines := make([]string, 0, len(questions)*3)
+	anyAnswered := false
+	for index := range questions {
+		if index < len(answers) && strings.TrimSpace(answers[index]) != "" {
+			anyAnswered = true
+			break
+		}
+	}
+
+	lines := make([]string, 0, len(questions)*3+2)
+	if len(questions) > 0 && !anyAnswered {
+		lines = append(lines, "[note] The user dismissed this prompt without answering. Treat this as a skip: do not assume a default — ask again more specifically, or proceed only if the intent is already unambiguous.")
+		lines = append(lines, "")
+	}
 	for index, question := range questions {
 		answer := ""
 		if index < len(answers) {
 			answer = strings.TrimSpace(answers[index])
 		}
 		if answer == "" {
-			answer = "(no answer provided)"
+			if anyAnswered {
+				answer = "(left blank)"
+			} else {
+				answer = "(skipped)"
+			}
 		}
 		lines = append(lines, fmt.Sprintf("%d. [question] %s", index+1, question.Question))
 		lines = append(lines, "[answer] "+answer)

internal/tools/ask_user_format_test.go

@@ -0,0 +1,34 @@
+package tools
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestFormatAskUserAnswersDistinguishesSkipFromBlank(t *testing.T) {
+	questions := []AskUserQuestion{
+		{Question: "Which database?"},
+		{Question: "Migrate existing data?"},
+	}
+
+	// Wholesale dismissal (nothing answered): flagged as a skip up front.
+	dismissed := FormatAskUserAnswers(questions, []string{"", ""})
+	if !strings.Contains(dismissed, "dismissed this prompt") || !strings.Contains(dismissed, "Treat this as a skip") {
+		t.Fatalf("all-empty answers must be flagged as a skip, got:\n%s", dismissed)
+	}
+	if strings.Contains(dismissed, "left blank") {
+		t.Fatalf("a full dismissal should read as skipped, not left-blank:\n%s", dismissed)
+	}
+	if !strings.Contains(dismissed, "(skipped)") {
+		t.Fatalf("dismissed questions should render (skipped):\n%s", dismissed)
+	}
+
+	// Partial answer: the blank one is "left blank", no wholesale-skip note.
+	partial := FormatAskUserAnswers(questions, []string{"postgres", ""})
+	if strings.Contains(partial, "dismissed this prompt") {
+		t.Fatalf("a partial answer is not a dismissal:\n%s", partial)
+	}
+	if !strings.Contains(partial, "postgres") || !strings.Contains(partial, "(left blank)") {
+		t.Fatalf("partial answers should show the answer and mark the blank one left-blank:\n%s", partial)
+	}
+}

internal/tools/bash.go

@@ -1,7 +1,6 @@
 package tools
 
 import (
-	"bytes"
 	"context"
 	"errors"
 	"fmt"
@@ -84,7 +83,12 @@ func (tool bashTool) run(ctx context.Context, args map[string]any, engine *zeroS
 	if err != nil {
 		return errorResult("Error: Invalid arguments for bash: " + err.Error())
 	}
-	if issue := detectShellCommandIssue(commandText, runtime.GOOS); issue != nil {
+	// Resolve the command engine before the MSYS preflight check so an
+	// approved require_escalated call (commandEngine == nil, truly
+	// unsandboxed) can actually bypass the MSYS guard instead of being
+	// hard-blocked by the same check it was meant to escalate past.
+	commandEngine := commandEngineForSandboxPermissions(engine, sandboxPermissions)
+	if issue := detectShellCommandIssue(commandText, runtime.GOOS); issue != nil && !msysGuardBypassed(issue, commandEngine) {
 		return shellIssueBlockResult(*issue)
 	}
 
@@ -107,7 +111,6 @@ func (tool bashTool) run(ctx context.Context, args map[string]any, engine *zeroS
 		"cwd":        relativeCwd,
 		"timeout_ms": strconv.Itoa(timeoutMS),
 	}
-	commandEngine := commandEngineForSandboxPermissions(engine, sandboxPermissions)
 	if commandEngine == nil && sandboxPermissions == SandboxPermissionsRequireEscalated {
 		meta["sandbox_permissions"] = string(SandboxPermissionsRequireEscalated)
 	}
@@ -120,14 +123,17 @@ func (tool bashTool) run(ctx context.Context, args map[string]any, engine *zeroS
 			Meta:   meta,
 		}
 	}
-	// Release any plan-scoped resources once the command has finished running.
 	defer plan.Cleanup()
 	addSandboxMeta(meta, plan)
 
-	var stdout bytes.Buffer
-	var stderr bytes.Buffer
-	command.Stdout = &stdout
-	command.Stderr = &stderr
+	// Bound the capture so a command with runaway output (`cat huge.log`, `yes`)
+	// can't grow Zero's memory before truncation: only the head+tail each stream
+	// will ever surface to the model are retained, the middle is discarded as it
+	// streams, and the true size is counted for the truncation marker.
+	stdout := newBoundedBuffer(bashCaptureBudgetBytes, bashCaptureBudgetBytes)
+	stderr := newBoundedBuffer(bashCaptureBudgetBytes, bashCaptureBudgetBytes)
+	command.Stdout = stdout
+	command.Stderr = stderr
 
 	// Kill the shell as a process group on timeout and bound the post-kill I/O
 	// wait, so a backgrounded child cannot outlive the command or hang Run().
@@ -139,7 +145,12 @@ func (tool bashTool) run(ctx context.Context, args map[string]any, engine *zeroS
 	err = command.Run()
 	exitCode := commandExitCode(err)
 	meta["exit_code"] = strconv.Itoa(exitCode)
-	stderrText := appendSandboxBlocks(stderr.String(), monitor.Stop())
+	stdoutText := stdout.retained()
+	stderrRetained := stderr.retained()
+	stderrText := appendSandboxBlocks(stderrRetained, monitor.Stop())
+	// Sandbox blocks are extra model-visible stderr bytes appended after capture;
+	// count them toward the true total so the budget/marker stay accurate.
+	stderrTotal := stderr.total + (len(stderrText) - len(stderrRetained))
 
 	if errors.Is(commandCtx.Err(), context.DeadlineExceeded) {
 		return Result{
@@ -156,26 +167,31 @@ func (tool bashTool) run(ctx context.Context, args map[string]any, engine *zeroS
 				Meta:   meta,
 			}
 		}
-		markLikelySandboxDenial(meta, plan, exitCode, stdout.String(), stderrText)
+		markLikelySandboxDenial(meta, plan, exitCode, stdoutText, stderrText)
+		outText, errText, truncated := budgetBashCapture(stdoutText, stdout.total, stderrText, stderrTotal, meta)
 		return Result{
-			Status: StatusError,
-			Output: formatBashOutputWithShellHint(commandText, stdout.String(), stderrText, exitCode, meta),
-			Meta:   meta,
+			Status:    StatusError,
+			Output:    formatBashOutputWithShellHint(outText, errText, exitCode, meta),
+			Truncated: truncated,
+			Meta:      meta,
 		}
 	}
 
-	markLikelySandboxDenial(meta, plan, exitCode, stdout.String(), stderrText)
+	markLikelySandboxDenial(meta, plan, exitCode, stdoutText, stderrText)
+	outText, errText, truncated := budgetBashCapture(stdoutText, stdout.total, stderrText, stderrTotal, meta)
 	if meta[SandboxLikelyDeniedMeta] == "true" {
 		return Result{
-			Status: StatusError,
-			Output: formatBashOutputWithShellHint(commandText, stdout.String(), stderrText, exitCode, meta),
-			Meta:   meta,
+			Status:    StatusError,
+			Output:    formatBashOutputWithShellHint(outText, errText, exitCode, meta),
+			Truncated: truncated,
+			Meta:      meta,
 		}
 	}
 	return Result{
-		Status: StatusOK,
-		Output: formatBashOutput(stdout.String(), stderrText, exitCode),
-		Meta:   meta,
+		Status:    StatusOK,
+		Output:    formatBashOutput(outText, errText, exitCode),
+		Truncated: truncated,
+		Meta:      meta,
 	}
 }
 
@@ -186,6 +202,18 @@ func commandEngineForSandboxPermissions(engine *zeroSandbox.Engine, sandboxPermi
 	return engine
 }
 
+// msysGuardBypassed reports whether a windows_msys_sandbox preflight issue no
+// longer applies because sandbox_permissions: require_escalated actually
+// resolved to unsandboxed, host-level execution (commandEngine is nil only
+// then, per commandEngineForSandboxPermissions). MSYS coreutils fail because
+// of the sandbox's restricted token/handles, so running outside it removes
+// the failure mode this guard exists for. Any other issue kind (e.g. a real
+// cmd.exe syntax problem) still blocks regardless of escalation, since
+// running outside the sandbox does not fix a syntax error.
+func msysGuardBypassed(issue *shellIssue, commandEngine *zeroSandbox.Engine) bool {
+	return issue.Kind == windowsMsysSandboxKind && commandEngine == nil
+}
+
 // appendSandboxBlocks appends a <sandbox_blocks> block listing the denials
 // the sandbox log monitor captured, so the model can see what was blocked. With no
 // blocks the stderr is returned unchanged.
@@ -221,6 +249,14 @@ func shellIssueBlockResult(issue shellIssue) Result {
 	}
 }
 
+// buildBashCommand returns the exec.Cmd and the sandbox plan for running
+// commandText. On Windows, when the command is not wrapped by the sandbox
+// engine (plan.Wrapped == false), it also overrides the child's raw command
+// line so commandText reaches cmd.exe unescaped; see
+// zeroSandbox.WindowsShellCommandLine for why that matters. The wrapped case
+// gets the same treatment inside the sandboxed runner process itself
+// (internal/sandbox/windows_process_windows.go), since that command line is
+// built there, not here.
 func buildBashCommand(ctx context.Context, commandText string, absoluteCwd string, engine *zeroSandbox.Engine) (*exec.Cmd, zeroSandbox.CommandPlan, error) {
 	spec := zeroSandbox.CommandSpec{
 		Name: shellExecutable(),
@@ -228,7 +264,11 @@ func buildBashCommand(ctx context.Context, commandText string, absoluteCwd strin
 		Dir:  absoluteCwd,
 	}
 	if engine != nil {
-		return engine.CommandContext(ctx, spec)
+		command, plan, err := engine.CommandContext(ctx, spec)
+		if err == nil {
+			applyWindowsShellCommandLine(command, commandText, plan.Wrapped)
+		}
+		return command, plan, err
 	}
 	plan := zeroSandbox.CommandPlan{
 		Backend: zeroSandbox.Backend{
@@ -242,6 +282,7 @@ func buildBashCommand(ctx context.Context, commandText string, absoluteCwd strin
 	}
 	command := exec.CommandContext(ctx, spec.Name, spec.Args...)
 	command.Dir = spec.Dir
+	applyWindowsShellCommandLine(command, commandText, plan.Wrapped)
 	return command, plan, nil
 }
 
@@ -302,7 +343,7 @@ func shellExecutable() string {
 
 func shellArguments(command string) []string {
 	if runtime.GOOS == "windows" {
-		return []string{"/d", "/s", "/c", command}
+		return zeroSandbox.WindowsShellArgs(command)
 	}
 	return []string{"-c", command}
 }
@@ -345,9 +386,178 @@ func formatBashOutput(stdout string, stderr string, exitCode int) string {
 	return strings.Join(parts, "\n")
 }
 
-func formatBashOutputWithShellHint(command string, stdout string, stderr string, exitCode int, meta map[string]string) string {
+// bashOutputBudgetBytes caps each of stdout/stderr shown to the model (32 KiB
+// ≈ 8k tokens per stream). bash is the one tool that can emit unbounded output
+// (`cat large.log`, `find /`, verbose test runs); a chatty build log at the old
+// 96 KiB budget injected ~24k tokens per call and was then re-billed on every
+// following turn. Head+tail truncation keeps both the start and the end of an
+// oversized stream, since build/test failures usually surface at the tail, and
+// the full captured text is spilled to disk so nothing is lost — the model
+// greps/reads the spill instead of re-running the command.
+const bashOutputBudgetBytes = 32 * 1024
+
+// bashCaptureBudgetBytes bounds the head and tail retained IN MEMORY per
+// stream during capture. Deliberately larger than the emit budget: the extra
+// retained bytes never reach the transcript, but they make the spill file —
+// the model's recovery path — cover 6× more of the output.
+const bashCaptureBudgetBytes = 96 * 1024
+
+// budgetBashOutput truncates stdout and stderr to bashOutputBudgetBytes each,
+// keeping the head and tail of anything larger, and records raw/emitted byte
+// counts plus a truncated flag in meta (mirroring outputBudgetMeta's shape for
+// the read/search tools). Detection that needs the full output (sandbox-denial
+// scanning) must run on the raw strings before this is applied.
+func budgetBashOutput(stdout string, stderr string, meta map[string]string) (string, string, bool) {
+	return budgetBashCapture(stdout, len(stdout), stderr, len(stderr), meta)
+}
+
+// budgetBashCapture is budgetBashOutput for the streaming-capture path: outTotal
+// and errTotal are the true byte counts (from boundedBuffer.total), which may
+// exceed the retained strings when the middle was dropped during capture. Meta's
+// raw_bytes therefore reflects everything the command produced, not just what was
+// kept in memory.
+func budgetBashCapture(out string, outTotal int, errStr string, errTotal int, meta map[string]string) (string, string, bool) {
+	outText, outRaw, outTrunc := truncateHeadTailWithTotal(out, outTotal, bashOutputBudgetBytes)
+	errText, errRaw, errTrunc := truncateHeadTailWithTotal(errStr, errTotal, bashOutputBudgetBytes)
+	truncated := outTrunc || errTrunc
+	if truncated {
+		if spillPath := spillBashStreams(out, outTotal, errStr, errTotal); spillPath != "" {
+			hint := "\n[zero] captured output saved to " + spillPath + " (grep or read_file it instead of re-running)"
+			if errTrunc {
+				errText += hint
+			} else {
+				outText += hint
+			}
+			if meta != nil {
+				meta["spill_path"] = spillPath
+			}
+		}
+	}
+	if meta != nil {
+		emitted := len(outText) + len(errText)
+		meta["raw_bytes"] = strconv.Itoa(outRaw + errRaw)
+		meta["emitted_bytes"] = strconv.Itoa(emitted)
+		meta["estimated_tokens"] = strconv.Itoa(estimatedTokensFromBytes(emitted))
+		if truncated {
+			meta["truncated"] = "true"
+		}
+	}
+	return outText, errText, truncated
+}
+
+// spillBashStreams writes the retained (capture-bounded) stdout and stderr to
+// the spill directory as one sectioned file and returns its path, or "" when
+// spilling fails. The spill holds up to bashCaptureBudgetBytes of head and
+// tail per stream — everything zero kept in memory, which is more than the
+// emit budget shows the model but not necessarily the whole output. When the
+// capture itself dropped the middle of a stream (total exceeds the retained
+// bytes), a gap marker is inserted at the head/tail junction so the spilled
+// log never reads as contiguous when it is not.
+func spillBashStreams(stdout string, stdoutTotal int, stderr string, stderrTotal int) string {
+	stdout = sectionWithCaptureGap(stdout, stdoutTotal)
+	stderr = sectionWithCaptureGap(stderr, stderrTotal)
+	var combined strings.Builder
+	combined.Grow(len(stdout) + len(stderr) + 64)
+	combined.WriteString("### stdout\n")
+	combined.WriteString(stdout)
+	if !strings.HasSuffix(stdout, "\n") {
+		combined.WriteString("\n")
+	}
+	combined.WriteString("### stderr\n")
+	combined.WriteString(stderr)
+	return spillTruncatedOutput("bash", combined.String())
+}
+
+// sectionWithCaptureGap marks the point where boundedBuffer dropped the middle
+// of a stream. When total exceeds the retained bytes, the retained text is the
+// frozen head (bashCaptureBudgetBytes, always full once overflow happened)
+// followed immediately by the rolling tail — the junction sits at the head cap,
+// snapped back to a rune boundary.
+func sectionWithCaptureGap(text string, total int) string {
+	if total <= len(text) || len(text) <= bashCaptureBudgetBytes {
+		return text
+	}
+	head := utf8Prefix(text, bashCaptureBudgetBytes)
+	marker := fmt.Sprintf("\n[zero] capture gap: %d bytes omitted from the middle of this stream\n", total-len(text))
+	return head + marker + text[len(head):]
+}
+
+// boundedBuffer is an io.Writer that retains at most headCap bytes from the start
+// and tailCap bytes from the end of a stream while counting the total written, so
+// a command emitting unbounded output (`cat huge.log`, `yes`) cannot grow Zero's
+// memory: the middle is discarded as it arrives instead of buffered whole and then
+// truncated. total records the full size for the truncation marker even though the
+// middle is never held. Not safe for concurrent writes; exec drives Stdout and
+// Stderr from separate goroutines, so each stream gets its own buffer.
+type boundedBuffer struct {
+	head    []byte
+	headCap int
+	tail    []byte
+	tailCap int
+	total   int
+}
+
+func newBoundedBuffer(headCap, tailCap int) *boundedBuffer {
+	return &boundedBuffer{headCap: headCap, tailCap: tailCap}
+}
+
+func (b *boundedBuffer) Write(p []byte) (int, error) {
+	n := len(p)
+	b.total += n
+	// Fill the head until it reaches headCap; the head is written once and frozen.
+	if len(b.head) < b.headCap {
+		take := b.headCap - len(b.head)
+		if take > len(p) {
+			take = len(p)
+		}
+		b.head = append(b.head, p[:take]...)
+		p = p[take:]
+	}
+	// Anything past the head feeds a rolling tail that keeps only the last tailCap
+	// bytes. Compact once it grows past 2×tailCap so the append stays amortized O(1)
+	// while memory never exceeds ~2×tailCap.
+	if len(p) > 0 && b.tailCap > 0 {
+		b.tail = append(b.tail, p...)
+		if len(b.tail) > 2*b.tailCap {
+			b.tail = append(b.tail[:0], b.tail[len(b.tail)-b.tailCap:]...)
+		}
+	}
+	return n, nil
+}
+
+// retained returns the kept head+tail bytes (marker-less) as a string. The junction
+// between head and tail lands in the middle, which the display budget trims away;
+// callers that need the true size read total separately.
+func (b *boundedBuffer) retained() string {
+	if len(b.tail) > b.tailCap {
+		// Not yet compacted since the last overflow; expose only the last tailCap.
+		return string(b.head) + string(b.tail[len(b.tail)-b.tailCap:])
+	}
+	return string(b.head) + string(b.tail)
+}
+
+// truncateHeadTailWithTotal head+tail-truncates value to maxBytes, using total —
+// the full original byte count — for the "N bytes omitted" marker and the raw
+// count. total may exceed len(value) when the middle was already discarded during
+// bounded capture (boundedBuffer): value then holds only the retained head+tail,
+// and this trims it to the display budget while still reporting the true total.
+func truncateHeadTailWithTotal(value string, total, maxBytes int) (string, int, bool) {
+	if maxBytes <= 0 || total <= maxBytes {
+		return value, total, false
+	}
+	marker := fmt.Sprintf("\n[zero] output truncated: %d bytes omitted from the middle — redirect to a file and read_file a range for the full text\n", total-maxBytes)
+	budget := maxBytes - len(marker)
+	if budget < 0 {
+		budget = 0
+	}
+	head := budget / 2
+	tail := budget - head
+	return utf8Prefix(value, head) + marker + utf8Suffix(value, tail), total, true
+}
+
+func formatBashOutputWithShellHint(stdout string, stderr string, exitCode int, meta map[string]string) string {
 	output := formatBashOutput(stdout, stderr, exitCode)
-	if issue := detectShellOutputIssue(command, stdout+"\n"+stderr, runtime.GOOS); issue != nil {
+	if issue := detectShellOutputIssue(stdout+"\n"+stderr, runtime.GOOS); issue != nil {
 		meta["shell_issue"] = issue.Kind
 		output = appendShellIssueHint(output, *issue)
 	}

internal/tools/bash_budget_test.go

@@ -0,0 +1,183 @@
+package tools
+
+import (
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// setTestTempDir points every platform's temp-dir lookup at a per-test
+// directory so spill files never land in the real temp dir: os.TempDir reads
+// TMPDIR on Unix but TMP/TEMP on Windows.
+func setTestTempDir(t *testing.T) {
+	t.Helper()
+	dir := t.TempDir()
+	t.Setenv("TMPDIR", dir)
+	t.Setenv("TMP", dir)
+	t.Setenv("TEMP", dir)
+}
+
+// Small output passes through untouched and records raw==emitted, no truncated flag.
+func TestBudgetBashOutputSmallPassesThrough(t *testing.T) {
+	meta := map[string]string{}
+	out, errStr, truncated := budgetBashOutput("hello\n", "warn\n", meta)
+	if out != "hello\n" || errStr != "warn\n" {
+		t.Fatalf("small output altered: out=%q err=%q", out, errStr)
+	}
+	if truncated {
+		t.Fatalf("small output must report truncated=false")
+	}
+	if meta["truncated"] == "true" {
+		t.Fatalf("small output must not be flagged truncated: %v", meta)
+	}
+	if meta["raw_bytes"] != strconv.Itoa(len("hello\n")+len("warn\n")) {
+		t.Fatalf("raw_bytes wrong: %v", meta)
+	}
+}
+
+// Oversized stdout is truncated head+tail: both the first and last lines survive,
+// the middle is dropped behind a marker, meta is flagged, and the captured text
+// is spilled to a re-readable file.
+func TestBudgetBashOutputTruncatesHeadAndTail(t *testing.T) {
+	setTestTempDir(t)
+	head := "FIRST_LINE_MARKER\n"
+	tail := "\nLAST_LINE_MARKER"
+	big := head + strings.Repeat("x", bashOutputBudgetBytes) + tail
+
+	meta := map[string]string{}
+	out, _, truncated := budgetBashOutput(big, "", meta)
+
+	if !truncated {
+		t.Fatalf("oversized output must report truncated=true")
+	}
+	if !strings.Contains(out, "FIRST_LINE_MARKER") {
+		t.Fatalf("head lost after truncation")
+	}
+	if !strings.Contains(out, "LAST_LINE_MARKER") {
+		t.Fatalf("tail lost after truncation (failures live at the tail)")
+	}
+	if !strings.Contains(out, "output truncated") {
+		t.Fatalf("expected a truncation marker, got:\n%s", out[:200])
+	}
+	// The spill hint (path length varies) is the only allowed overage past the budget.
+	if len(out) > bashOutputBudgetBytes+512 {
+		t.Fatalf("emitted %d bytes exceeds budget %d", len(out), bashOutputBudgetBytes)
+	}
+	if !strings.Contains(out, "captured output saved to ") {
+		t.Fatalf("truncated bash output must carry a spill hint:\n%s", out[len(out)-300:])
+	}
+	if meta["spill_path"] == "" {
+		t.Fatal("spill path missing from meta")
+	}
+	content, err := os.ReadFile(meta["spill_path"])
+	if err != nil {
+		t.Fatalf("spill file unreadable: %v", err)
+	}
+	if !strings.Contains(string(content), "### stdout") || !strings.Contains(string(content), strings.Repeat("x", 1024)) {
+		t.Fatalf("spill must hold the sectioned captured streams (got %d bytes)", len(content))
+	}
+	if meta["truncated"] != "true" {
+		t.Fatalf("expected truncated=true, got %v", meta)
+	}
+	if meta["raw_bytes"] != strconv.Itoa(len(big)) {
+		t.Fatalf("raw_bytes = %s, want %d", meta["raw_bytes"], len(big))
+	}
+	if got, _ := strconv.Atoi(meta["emitted_bytes"]); got != len(out) {
+		t.Fatalf("emitted_bytes = %s, want %d", meta["emitted_bytes"], len(out))
+	}
+}
+
+// boundedBuffer must keep memory bounded to head+tail while counting every byte,
+// so a runaway command can't OOM Zero before truncation runs. It keeps the very
+// first bytes (head) and the very last bytes (tail), discarding the middle.
+func TestBoundedBufferKeepsHeadAndTailBounded(t *testing.T) {
+	b := newBoundedBuffer(8, 8)
+	total := 0
+	for i := 0; i < 1000; i++ {
+		chunk := []byte(fmt.Sprintf("%05d.", i)) // 6 bytes each
+		n, err := b.Write(chunk)
+		if err != nil || n != len(chunk) {
+			t.Fatalf("Write = (%d, %v), want (%d, nil)", n, err, len(chunk))
+		}
+		total += len(chunk)
+	}
+	if b.total != total {
+		t.Fatalf("total = %d, want %d (every byte must be counted)", b.total, total)
+	}
+	// Retained memory is bounded regardless of how much streamed through.
+	if len(b.head) > 8 {
+		t.Fatalf("head grew past its cap: %d", len(b.head))
+	}
+	if len(b.tail) > 2*8 {
+		t.Fatalf("tail grew past 2×cap (compaction failed): %d", len(b.tail))
+	}
+	r := b.retained()
+	if len(r) > 8+8 {
+		t.Fatalf("retained %d bytes, want <= headCap+tailCap", len(r))
+	}
+	if !strings.HasPrefix(r, "00000.") {
+		t.Fatalf("retained head lost, got prefix %q", r[:min(6, len(r))])
+	}
+	if !strings.HasSuffix(r, "00999.") {
+		t.Fatalf("retained tail lost, got %q", r)
+	}
+}
+
+// budgetBashCapture reports the TRUE total (not the retained size) in the marker
+// and raw_bytes, even though only a bounded head+tail was ever held in memory.
+func TestBudgetBashCaptureReportsTrueTotal(t *testing.T) {
+	setTestTempDir(t)
+	// Retained head+tail as boundedBuffer would hand over after overflow: the
+	// frozen head is full at bashCaptureBudgetBytes, then the rolling tail; the
+	// real command produced far more than was kept.
+	retained := "HEAD_START" + strings.Repeat("y", bashCaptureBudgetBytes) + "TAIL_END"
+	total := 10 * bashCaptureBudgetBytes
+
+	meta := map[string]string{}
+	out, _, truncated := budgetBashCapture(retained, total, "", 0, meta)
+
+	if !truncated {
+		t.Fatal("oversized capture must report truncated=true")
+	}
+	if !strings.Contains(out, "HEAD_START") || !strings.Contains(out, "TAIL_END") {
+		t.Fatalf("head/tail lost after budgeting:\n%s", out[:min(120, len(out))])
+	}
+	// Spill hint is the only allowed overage past the budget.
+	if len(out) > bashOutputBudgetBytes+512 {
+		t.Fatalf("emitted %d bytes exceeds budget %d", len(out), bashOutputBudgetBytes)
+	}
+	if meta["raw_bytes"] != strconv.Itoa(total) {
+		t.Fatalf("raw_bytes = %s, want the true total %d", meta["raw_bytes"], total)
+	}
+	// The marker must cite the true omitted count (total-budget), not retained-budget.
+	if !strings.Contains(out, strconv.Itoa(total-bashOutputBudgetBytes)) {
+		t.Fatalf("marker should cite the true omitted byte count:\n%s", out[:min(200, len(out))])
+	}
+
+	// The spill must be recorded, sectioned, and carry a capture-gap marker at
+	// the head/tail junction — the capture dropped the middle, so the spilled
+	// log must never read as contiguous.
+	if meta["spill_path"] == "" {
+		t.Fatal("spill path missing from meta")
+	}
+	content, err := os.ReadFile(meta["spill_path"])
+	if err != nil {
+		t.Fatalf("spill file unreadable: %v", err)
+	}
+	if !strings.Contains(string(content), "### stdout") {
+		t.Fatal("spill must be sectioned by stream")
+	}
+	gapMarker := fmt.Sprintf("capture gap: %d bytes omitted", total-len(retained))
+	gapIndex := strings.Index(string(content), gapMarker)
+	if gapIndex < 0 {
+		t.Fatalf("spill missing capture-gap marker %q", gapMarker)
+	}
+	// The junction sits right after the frozen head (plus the section header),
+	// not appended at the end of the stream.
+	headerOffset := len("### stdout\n")
+	if gapIndex < headerOffset+bashCaptureBudgetBytes-64 || gapIndex > headerOffset+bashCaptureBudgetBytes+64 {
+		t.Fatalf("capture-gap marker at offset %d, want ~%d (the head cap)", gapIndex, headerOffset+bashCaptureBudgetBytes)
+	}
+}

internal/tools/bash_proc_unix.go

@@ -42,3 +42,7 @@ func hardenProcessLifetime(command *exec.Cmd) {
 		return nil
 	}
 }
+
+// applyWindowsShellCommandLine is a no-op outside Windows; there is no
+// cmd.exe command-line-parsing quirk to work around here.
+func applyWindowsShellCommandLine(command *exec.Cmd, commandText string, wrapped bool) {}

internal/tools/bash_proc_windows.go

@@ -5,7 +5,10 @@ package tools
 import (
 	"os/exec"
 	"strconv"
+	"syscall"
 	"time"
+
+	zeroSandbox "github.com/Gitlawb/zero/internal/sandbox"
 )
 
 // bashWaitDelay bounds how long Wait blocks for the I/O pipes to drain after the
@@ -28,3 +31,17 @@ func hardenProcessLifetime(command *exec.Cmd) {
 		return nil
 	}
 }
+
+// applyWindowsShellCommandLine overrides command's raw child command line so
+// commandText reaches cmd.exe unescaped instead of auto-quoted the way
+// exec.Cmd would normally encode a single Args element. Skipped when wrapped
+// is true: the sandbox engine then routes execution through a separate
+// zero-windows-command-runner process, which builds its own child command
+// line from scratch (internal/sandbox/windows_process_windows.go) rather than
+// inheriting whatever this outer exec.Cmd is configured with.
+func applyWindowsShellCommandLine(command *exec.Cmd, commandText string, wrapped bool) {
+	if wrapped {
+		return
+	}
+	command.SysProcAttr = &syscall.SysProcAttr{CmdLine: zeroSandbox.WindowsShellCommandLine(commandText)}
+}

internal/tools/bash_tool_test.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"testing"
 	"time"
@@ -17,6 +18,18 @@ import (
 
 func TestMain(m *testing.M) {
 	if len(os.Args) >= 3 && os.Args[1] == "--zero-bash-helper" {
+		if os.Args[2] == "echo-arg" {
+			// Echoes back exactly the one argument it received, to verify a
+			// quoted, space-and-slash-containing argument (the same shape as
+			// `python -c "print(15 / 3)"`) survives the shell invocation
+			// intact instead of being truncated/corrupted.
+			if len(os.Args) < 4 {
+				fmt.Fprintln(os.Stderr, "echo-arg requires exactly one argument")
+				os.Exit(1)
+			}
+			fmt.Println(os.Args[3])
+			return
+		}
 		runBashToolHelper(os.Args[2])
 		return
 	}
@@ -168,9 +181,13 @@ func TestDetectShellCommandIssueFlagsPipedPosixUtilities(t *testing.T) {
 		`cat file.txt | wc -l`,
 		`some_command | tail -n 20`,
 	} {
-		if issue := detectShellCommandIssue(command, "windows"); issue == nil {
+		issue := detectShellCommandIssue(command, "windows")
+		if issue == nil {
 			t.Fatalf("expected POSIX utility pipeline to be flagged for %q", command)
 		}
+		if issue.Kind != "windows_msys_sandbox" {
+			t.Fatalf("expected windows_msys_sandbox for %q, got %q", command, issue.Kind)
+		}
 	}
 }
 
@@ -182,7 +199,7 @@ func TestDetectShellCommandIssueAllowsUnrelatedCommands(t *testing.T) {
 		`echo header text`,
 		`tail-log.sh`,
 		`grep-cli --version`,
-		`sed.exe --help`,
+		`mytool.exe --help`,
 	} {
 		if issue := detectShellCommandIssue(command, "windows"); issue != nil {
 			t.Fatalf("expected unrelated command to pass for %q, got %#v", command, issue)
@@ -191,7 +208,7 @@ func TestDetectShellCommandIssueAllowsUnrelatedCommands(t *testing.T) {
 }
 
 func TestDetectShellOutputIssueAddsWindowsSyntaxHint(t *testing.T) {
-	issue := detectShellOutputIssue(`cd /d/tmp/zero-pr-158 && ls -la`, "The syntax of the command is incorrect.", "windows")
+	issue := detectShellOutputIssue("The syntax of the command is incorrect.", "windows")
 	if issue == nil {
 		t.Fatal("expected Windows syntax error to get shell guidance")
 	}
@@ -240,6 +257,43 @@ func TestBashToolRunsCommandInWorkspace(t *testing.T) {
 	}
 }
 
+// A command with runaway output must not be buffered whole in memory: the capture
+// is bounded to head+tail, yet raw_bytes still reports the true (much larger) size.
+// If capture were unbounded this would balloon Zero's memory before truncation.
+func TestBashToolBoundsRunawayOutputCapture(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("uses a POSIX shell pipeline (yes | head)")
+	}
+	const produced = 500000 // ~5× the 96 KiB budget
+
+	result := NewBashTool(t.TempDir()).Run(context.Background(), map[string]any{
+		"command": fmt.Sprintf("yes ABCDEFGH | head -c %d", produced),
+	})
+
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok status, got %s: %s", result.Status, result.Output)
+	}
+	if !result.Truncated || result.Meta["truncated"] != "true" {
+		t.Fatalf("runaway output should be flagged truncated, meta=%v", result.Meta)
+	}
+	if !strings.Contains(result.Output, "output truncated") {
+		t.Fatalf("expected a truncation marker in output")
+	}
+	// The emitted (model-visible) text is capped near the budget...
+	if len(result.Output) > bashOutputBudgetBytes+1024 {
+		t.Fatalf("emitted %d bytes far exceeds budget %d", len(result.Output), bashOutputBudgetBytes)
+	}
+	// ...but raw_bytes reflects the full stream, proving the total was counted while
+	// only a bounded slice was ever held (raw >> the ~2×budget retention cap).
+	raw, err := strconv.Atoi(result.Meta["raw_bytes"])
+	if err != nil {
+		t.Fatalf("raw_bytes not an int: %q", result.Meta["raw_bytes"])
+	}
+	if raw < produced {
+		t.Fatalf("raw_bytes = %d, want >= %d (full stream counted)", raw, produced)
+	}
+}
+
 func TestBashToolUsesRequestedCwd(t *testing.T) {
 	root := t.TempDir()
 	nested := filepath.Join(root, "nested")
@@ -384,6 +438,105 @@ func TestRegistryRunsWithDegradedUnavailableNativeSandbox(t *testing.T) {
 	}
 }
 
+// TestBashToolRequireEscalatedMsysGuard covers the boundary of the fix for an
+// escalation-vs-preflight ordering bug: the MSYS sandbox guard exists only
+// because MSYS/Cygwin coreutils fail under the write-restricted sandbox, so
+// once sandbox_permissions: require_escalated is actually approved (running
+// the command unsandboxed), the guard must get out of the way, but an
+// unrelated windows_shell_syntax issue (a real cmd.exe syntax problem, not a
+// sandbox interaction) must still block regardless of escalation.
+func TestBashToolRequireEscalatedMsysGuard(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows-only MSYS sandbox guard")
+	}
+	root := t.TempDir()
+	newEngine := func() *sandbox.Engine {
+		return sandbox.NewEngine(sandbox.EngineOptions{
+			WorkspaceRoot: root,
+			Policy:        sandbox.DefaultPolicy(),
+			Backend:       sandbox.Backend{Name: sandbox.BackendUnavailable, Message: "native sandbox unavailable"},
+		})
+	}
+	registry := NewRegistry()
+	registry.Register(NewBashTool(root))
+
+	t.Run("default sandboxing still blocks an MSYS-prone command", func(t *testing.T) {
+		result := registry.RunWithOptions(context.Background(), "bash", map[string]any{
+			"command": "cat somefile.txt",
+		}, RunOptions{
+			PermissionGranted: true,
+			Sandbox:           newEngine(),
+			PermissionMode:    string(sandbox.PermissionModeAsk),
+		})
+		if result.Meta["shell_issue"] != "windows_msys_sandbox" {
+			t.Fatalf("expected windows_msys_sandbox block without escalation, got %#v", result)
+		}
+	})
+
+	t.Run("approved require_escalated bypasses the MSYS guard", func(t *testing.T) {
+		result := registry.RunWithOptions(context.Background(), "bash", map[string]any{
+			"command":             "cat somefile.txt",
+			"sandbox_permissions": string(SandboxPermissionsRequireEscalated),
+		}, RunOptions{
+			PermissionGranted: true,
+			Sandbox:           newEngine(),
+			PermissionMode:    string(sandbox.PermissionModeAsk),
+		})
+		// Assert on the preflight block sentinel (exit_code "-1", set only by
+		// shellIssueBlockResult) rather than shell_issue: once the guard is
+		// bypassed, "cat somefile.txt" actually runs, and its real,
+		// PATH-dependent output could otherwise trip the unrelated
+		// post-execution detectShellOutputIssue heuristic and make this
+		// assertion flaky for reasons unrelated to the guard under test.
+		if result.Meta["exit_code"] == "-1" {
+			t.Fatalf("expected approved require_escalated to bypass the MSYS guard, got blocked: %#v", result)
+		}
+	})
+
+	t.Run("approved require_escalated still blocks an unrelated syntax issue", func(t *testing.T) {
+		result := registry.RunWithOptions(context.Background(), "bash", map[string]any{
+			"command":             `cd /d/tmp/zero-pr-158 && dir`,
+			"sandbox_permissions": string(SandboxPermissionsRequireEscalated),
+		}, RunOptions{
+			PermissionGranted: true,
+			Sandbox:           newEngine(),
+			PermissionMode:    string(sandbox.PermissionModeAsk),
+		})
+		if result.Meta["shell_issue"] != "windows_shell_syntax" {
+			t.Fatalf("expected windows_shell_syntax block to still apply under require_escalated, got %#v", result)
+		}
+	})
+}
+
+// TestBashToolIgnoresMsysMarkersInCommandArgumentsAfterFailure guards the
+// post-execution MSYS heuristic against treating the command line itself as
+// evidence. The helper command below fails for a reason unrelated to MSYS
+// (a plain non-zero exit), but its own argument text quotes an MSYS crash
+// marker the way a PR-comment or commit-message argument might. Before the
+// fix, detectShellOutputIssue scanned command+output together and would have
+// misdiagnosed this as an MSYS sandbox failure even though the real
+// stdout/stderr never mentions MSYS at all.
+func TestBashToolIgnoresMsysMarkersInCommandArgumentsAfterFailure(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows-only MSYS output guard")
+	}
+	root := t.TempDir()
+	// The helper only reads argv[2] ("fail"); this trailing quoted argument is
+	// otherwise unused by the helper but still part of the command line text.
+	command := helperCommand("fail") + ` "fatal error - CreateFileMapping S-1-5-21, Win32 error 5. Terminating. cygheap_user::init"`
+
+	result := NewBashTool(root).Run(context.Background(), map[string]any{
+		"command": command,
+	})
+
+	if result.Status != StatusError || result.Meta["exit_code"] != "7" {
+		t.Fatalf("expected the helper's real exit 7 failure, got %s: %#v", result.Status, result)
+	}
+	if result.Meta["shell_issue"] == "windows_msys_sandbox" {
+		t.Fatalf("expected the MSYS marker in the command's own argument text to be ignored, got %#v", result)
+	}
+}
+
 func TestBashToolRunsWithDegradedUnavailableNativeSandbox(t *testing.T) {
 	root := t.TempDir()
 	policy := sandbox.DefaultPolicy()
@@ -535,6 +688,68 @@ func TestBashToolAllowsNonInteractiveCommand(t *testing.T) {
 	}
 }
 
+// TestBashToolPreservesEmbeddedQuotesOnWindows pins a real, previously-broken
+// case: passing commandText as a normal exec.Cmd argument makes Go wrap it in
+// an outer pair of quotes (it contains spaces) with its own quotes escaped as
+// \", and cmd.exe's /C remainder parsing strips the first and last literal
+// quote character in that remainder without undoing the backslash-escaping -
+// corrupting a command whose own text contains embedded double quotes,
+// exactly the shape of `python -c "print(15 / 3)"`, `git commit -m
+// "message"`, `node -e "..."`. Before the fix, the helper below received a
+// truncated, mis-quoted argument (starting with a stray literal `"`) instead
+// of the text between the quotes.
+func TestBashToolPreservesEmbeddedQuotesOnWindows(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("cmd.exe /S/C quote-remainder parsing is Windows-specific")
+	}
+	root := t.TempDir()
+	executable := shellQuote(os.Args[0])
+	// Space and a slash: the two characters whose position in the original
+	// bug report's SyntaxError ("print(15, truncated right before the /")
+	// showed exactly where the corruption happened.
+	commandText := executable + ` --zero-bash-helper echo-arg "hello / world"`
+
+	result := NewBashTool(root).Run(context.Background(), map[string]any{
+		"command": commandText,
+	})
+
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok status, got %s: %s", result.Status, result.Output)
+	}
+	if !strings.Contains(result.Output, "hello / world") {
+		t.Fatalf("expected the embedded-quote argument to survive intact, got %q", result.Output)
+	}
+}
+
+// TestBashToolRunsCommandLineForLoopSyntax pins cmd.exe command-line syntax,
+// not batch-file syntax: a `for %i in (...) do ...` loop with a single
+// percent sign is valid typed directly at a cmd.exe prompt (or via
+// `cmd /C "..."`), but requires %%i inside an actual .bat/.cmd FILE, because
+// batch files perform an extra pass of percent-substitution before FOR ever
+// runs, consuming the single percent so the loop variable never binds. An
+// earlier version of this fix ran commandText from a temporary .cmd file
+// instead of passing it straight through to cmd.exe's /C remainder, which
+// silently broke single-percent syntax like this.
+func TestBashToolRunsCommandLineForLoopSyntax(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("cmd.exe command-line vs batch-file parsing is Windows-specific")
+	}
+	root := t.TempDir()
+
+	result := NewBashTool(root).Run(context.Background(), map[string]any{
+		"command": "for %i in (1 2 3) do echo %i",
+	})
+
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok status, got %s: %s", result.Status, result.Output)
+	}
+	for _, want := range []string{"1", "2", "3"} {
+		if !strings.Contains(result.Output, want) {
+			t.Fatalf("expected the for-loop to expand %%i to %q, got %q", want, result.Output)
+		}
+	}
+}
+
 func helperCommand(name string) string {
 	executable := shellQuote(os.Args[0])
 	return executable + " --zero-bash-helper " + name

internal/tools/edit_file.go

@@ -2,6 +2,7 @@ package tools
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"os"
 	"strings"
@@ -44,7 +45,7 @@ func (tool editFileTool) Run(ctx context.Context, args map[string]any) Result {
 	return tool.RunWithOptions(ctx, args, RunOptions{})
 }
 
-func (tool editFileTool) RunWithOptions(_ context.Context, args map[string]any, options RunOptions) Result {
+func (tool editFileTool) RunWithOptions(ctx context.Context, args map[string]any, options RunOptions) Result {
 	requestedPath, err := aliasedStringArg(args, []string{"path", "file", "file_path", "filename"}, "", true, false)
 	if err != nil {
 		return errorResult("Error: Invalid arguments for edit_file: " + err.Error())
@@ -98,6 +99,35 @@ func (tool editFileTool) RunWithOptions(_ context.Context, args map[string]any,
 		}
 	}
 
+	// Fuzzy fallback: when the exact string (and its CRLF translation) is absent,
+	// run a cascade of tolerant matchers (trimmed lines, block anchors, collapsed
+	// whitespace, indentation drift, escape normalization) to locate the span the
+	// model intended. Only a span that occurs literally in the file is accepted,
+	// so the replacement applied below is still exact.
+	if occurrences == 0 {
+		findOld, findNew := oldString, newString
+		if strings.Contains(content, "\r\n") && !strings.Contains(findOld, "\r\n") {
+			findOld = strings.ReplaceAll(findOld, "\n", "\r\n")
+			findNew = strings.ReplaceAll(findNew, "\n", "\r\n")
+		}
+		search, ferr := fuzzyEditMatch(content, findOld, replaceAll)
+		switch {
+		case ferr == nil:
+			oldString = search
+			// The model's new_string was written at old_string's (mismatched)
+			// indentation; re-shape it to the span actually being replaced so a
+			// tolerant match never strips indentation or a trailing CR.
+			newString = adaptReplacementToSpan(search, findOld, findNew)
+			occurrences = strings.Count(content, search)
+		case errors.Is(ferr, errEditFuzzyAmbiguous):
+			return errorResult("Error: old_string matches multiple locations in " + relativePath + " even after fuzzy matching. Provide more surrounding context to make the match unique, or pass replace_all: true.")
+		case errors.Is(ferr, errEditFuzzyNotFound):
+			// Fall through to the exact-match error below.
+		default:
+			return errorResult("Error editing " + relativePath + ": " + ferr.Error())
+		}
+	}
+
 	if occurrences == 0 {
 		return errorResult("Error: Could not find the exact string to replace in " + relativePath + ". The old_string must match the file byte-for-byte.")
 	}
@@ -120,6 +150,10 @@ func (tool editFileTool) RunWithOptions(_ context.Context, args map[string]any,
 	if err := os.WriteFile(absolutePath, []byte(updated), 0o644); err != nil {
 		return errorResult("Error writing " + relativePath + ": " + err.Error())
 	}
+	// Optional format-on-write (ZERO_FORMAT_ON_WRITE). Must run BEFORE the
+	// FileTracker re-baseline: recording pre-format content would make the very
+	// next edit look like an external modification and trip the conflict guard.
+	updated = maybeFormatWrittenFile(ctx, absolutePath, updated)
 	// Re-baseline to the content we just wrote so subsequent edits in this session
 	// compare against the current on-disk state, not the pre-edit version.
 	newInfo, _ := os.Stat(absolutePath)
@@ -130,6 +164,7 @@ func (tool editFileTool) RunWithOptions(_ context.Context, args map[string]any,
 		suffix = "s"
 	}
 	summary := fmt.Sprintf("Successfully edited %s (replaced %d occurrence%s).", relativePath, replacedCount, suffix)
+	summary += inlineDiagnostics(ctx, options, absolutePath, relativePath)
 	result := okResult(summary)
 	result.ChangedFiles = []string{relativePath}
 	// Card-only preview (Display.Preview): the model's Output stays the one-line

internal/tools/edit_replacers.go

@@ -0,0 +1,554 @@
+package tools
+
+// Fuzzy fallback matching for edit_file. When the model's old_string fails to
+// match byte-for-byte (drifted indentation, collapsed whitespace, escaped
+// characters, a slightly-misremembered middle line), these replacers propose
+// candidate spans that plausibly correspond to what the model intended. Only a
+// candidate that occurs literally in the file is accepted, so the replacement
+// itself is always exact even when the match was tolerant.
+//
+// Strategy cascade ported from opencode's edit tool, whose replacers were in
+// turn distilled from Cline's diff-apply evals and gemini-cli's edit corrector.
+
+import (
+	"errors"
+	"regexp"
+	"sort"
+	"strings"
+)
+
+// editReplacer proposes candidate spans for find inside content. Candidates
+// are re-validated by fuzzyEditMatch before use.
+type editReplacer func(content, find string) []string
+
+var (
+	errEditFuzzyNotFound  = errors.New("no fuzzy match for old_string")
+	errEditFuzzyAmbiguous = errors.New("fuzzy match for old_string is ambiguous")
+)
+
+// Minimum average middle-line similarity for block-anchor matches.
+const editAnchorSimilarityThreshold = 0.65
+
+// fuzzyEditMatch runs the replacer cascade and returns the exact span of
+// content to replace. When replaceAll is false the span must be unique in
+// content; an ambiguous candidate is skipped in favor of later candidates and
+// only reported if nothing unique is found. A span wildly larger than
+// old_string is refused outright rather than risking a destructive edit.
+func fuzzyEditMatch(content, find string, replaceAll bool) (string, error) {
+	replacers := []editReplacer{
+		lineTrimmedReplacer,
+		blockAnchorReplacer,
+		whitespaceNormalizedReplacer,
+		indentationFlexibleReplacer,
+		escapeNormalizedReplacer,
+		trimmedBoundaryReplacer,
+		contextAwareReplacer,
+	}
+	found := false
+	for _, replacer := range replacers {
+		// Collect the replacer's DISTINCT candidate spans that literally occur in
+		// content. Two or more distinct spans from one strategy (e.g. duplicate
+		// blocks at different indentation, each occurring once) mean the model's
+		// intent is genuinely ambiguous — silently editing the first would be a
+		// wrong-span write, so it is rejected instead.
+		var candidates []string
+		seen := map[string]bool{}
+		for _, search := range replacer(content, find) {
+			if search == "" || seen[search] {
+				continue
+			}
+			seen[search] = true
+			if !strings.Contains(content, search) {
+				continue
+			}
+			candidates = append(candidates, search)
+		}
+		if len(candidates) == 0 {
+			continue
+		}
+		found = true
+		if !replaceAll && len(candidates) > 1 {
+			return "", errEditFuzzyAmbiguous
+		}
+		search := candidates[0]
+		if isDisproportionateEditMatch(search, find) {
+			return "", errors.New("refusing replacement because the matched span is much larger than old_string; re-read the file and provide the full exact old_string for the intended replacement")
+		}
+		if replaceAll {
+			return search, nil
+		}
+		if strings.Index(content, search) == strings.LastIndex(content, search) {
+			return search, nil
+		}
+		// The single candidate occurs at multiple positions; a later, stricter
+		// strategy may still resolve a unique span, so keep cascading.
+	}
+	if !found {
+		return "", errEditFuzzyNotFound
+	}
+	return "", errEditFuzzyAmbiguous
+}
+
+// isDisproportionateEditMatch guards against anchor-style replacers matching a
+// span far larger than the text the model asked to replace (e.g. first/last
+// line anchors bridging hundreds of unrelated lines).
+func isDisproportionateEditMatch(search, find string) bool {
+	findLines := strings.Count(find, "\n") + 1
+	searchLines := strings.Count(search, "\n") + 1
+	limit := findLines + 3
+	if findLines*2 > limit {
+		limit = findLines * 2
+	}
+	if searchLines >= limit {
+		return true
+	}
+	if findLines == 1 {
+		return false
+	}
+	searchTrimmed := len(strings.TrimSpace(search))
+	findTrimmed := len(strings.TrimSpace(find))
+	byteLimit := findTrimmed + 500
+	if findTrimmed*4 > byteLimit {
+		byteLimit = findTrimmed * 4
+	}
+	return searchTrimmed > byteLimit
+}
+
+// splitFindLines splits find into lines, dropping the trailing empty line a
+// trailing newline produces so windows align with real content lines.
+func splitFindLines(find string) []string {
+	lines := strings.Split(find, "\n")
+	if len(lines) > 0 && lines[len(lines)-1] == "" {
+		lines = lines[:len(lines)-1]
+	}
+	return lines
+}
+
+// lineTrimmedReplacer matches when every line equals the corresponding
+// old_string line after trimming surrounding whitespace.
+func lineTrimmedReplacer(content, find string) []string {
+	contentLines := strings.Split(content, "\n")
+	findLines := splitFindLines(find)
+	if len(findLines) == 0 {
+		return nil
+	}
+	var candidates []string
+	for i := 0; i+len(findLines) <= len(contentLines); i++ {
+		matches := true
+		for j := range findLines {
+			if strings.TrimSpace(contentLines[i+j]) != strings.TrimSpace(findLines[j]) {
+				matches = false
+				break
+			}
+		}
+		if matches {
+			candidates = append(candidates, strings.Join(contentLines[i:i+len(findLines)], "\n"))
+		}
+	}
+	return candidates
+}
+
+// blockAnchorReplacer anchors on the first and last lines (trimmed) and
+// accepts the block when the middle lines average >= the similarity threshold
+// (Levenshtein-based), tolerating a slightly misremembered interior.
+func blockAnchorReplacer(content, find string) []string {
+	findLines := splitFindLines(find)
+	if len(findLines) < 3 {
+		return nil
+	}
+	contentLines := strings.Split(content, "\n")
+	firstAnchor := strings.TrimSpace(findLines[0])
+	lastAnchor := strings.TrimSpace(findLines[len(findLines)-1])
+	searchBlockSize := len(findLines)
+	maxLineDelta := searchBlockSize / 4
+	if maxLineDelta < 1 {
+		maxLineDelta = 1
+	}
+
+	type span struct{ start, end int }
+	var candidates []span
+	for i := 0; i < len(contentLines); i++ {
+		if strings.TrimSpace(contentLines[i]) != firstAnchor {
+			continue
+		}
+		// Only the first occurrence of the last anchor after this start counts,
+		// mirroring the reference implementation: a farther-away closing line is
+		// assumed to close a different block.
+		for j := i + 2; j < len(contentLines); j++ {
+			if strings.TrimSpace(contentLines[j]) != lastAnchor {
+				continue
+			}
+			actualBlockSize := j - i + 1
+			delta := actualBlockSize - searchBlockSize
+			if delta < 0 {
+				delta = -delta
+			}
+			if delta <= maxLineDelta {
+				candidates = append(candidates, span{start: i, end: j})
+			}
+			break
+		}
+	}
+	if len(candidates) == 0 {
+		return nil
+	}
+
+	middleSimilarity := func(candidate span) float64 {
+		actualBlockSize := candidate.end - candidate.start + 1
+		linesToCheck := searchBlockSize - 2
+		if actualBlockSize-2 < linesToCheck {
+			linesToCheck = actualBlockSize - 2
+		}
+		if linesToCheck <= 0 {
+			return 1.0
+		}
+		similarity := 0.0
+		for j := 1; j < searchBlockSize-1 && j < actualBlockSize-1; j++ {
+			contentLine := strings.TrimSpace(contentLines[candidate.start+j])
+			findLine := strings.TrimSpace(findLines[j])
+			maxLen := len(contentLine)
+			if len(findLine) > maxLen {
+				maxLen = len(findLine)
+			}
+			if maxLen == 0 {
+				continue
+			}
+			distance := levenshtein(contentLine, findLine)
+			similarity += 1 - float64(distance)/float64(maxLen)
+		}
+		return similarity / float64(linesToCheck)
+	}
+
+	// Return EVERY candidate that clears the similarity threshold, best first.
+	// Picking only the best would hide competing blocks from fuzzyEditMatch's
+	// distinct-candidate ambiguity check and could silently edit the wrong
+	// block; with all qualifying spans surfaced, one clear winner still
+	// resolves (single candidate) while two plausible blocks are rejected as
+	// ambiguous. replaceAll consumers take the first (most similar) span.
+	type scoredSpan struct {
+		span       span
+		similarity float64
+	}
+	var qualifying []scoredSpan
+	for _, candidate := range candidates {
+		if s := middleSimilarity(candidate); s >= editAnchorSimilarityThreshold {
+			qualifying = append(qualifying, scoredSpan{span: candidate, similarity: s})
+		}
+	}
+	if len(qualifying) == 0 {
+		return nil
+	}
+	sort.SliceStable(qualifying, func(i, j int) bool {
+		return qualifying[i].similarity > qualifying[j].similarity
+	})
+	spans := make([]string, 0, len(qualifying))
+	for _, scored := range qualifying {
+		spans = append(spans, strings.Join(contentLines[scored.span.start:scored.span.end+1], "\n"))
+	}
+	return spans
+}
+
+var editWhitespaceRun = regexp.MustCompile(`\s+`)
+
+func normalizeEditWhitespace(text string) string {
+	return strings.TrimSpace(editWhitespaceRun.ReplaceAllString(text, " "))
+}
+
+// whitespaceNormalizedReplacer matches after collapsing all whitespace runs to
+// a single space: full lines, sub-line spans (via a word-boundary regex), and
+// multi-line windows.
+func whitespaceNormalizedReplacer(content, find string) []string {
+	normalizedFind := normalizeEditWhitespace(find)
+	if normalizedFind == "" {
+		return nil
+	}
+	contentLines := strings.Split(content, "\n")
+	var candidates []string
+	var subLinePattern *regexp.Regexp
+	for _, line := range contentLines {
+		normalizedLine := normalizeEditWhitespace(line)
+		if normalizedLine == normalizedFind {
+			candidates = append(candidates, line)
+			continue
+		}
+		if !strings.Contains(normalizedLine, normalizedFind) {
+			continue
+		}
+		if subLinePattern == nil {
+			words := strings.Fields(find)
+			quoted := make([]string, len(words))
+			for i, word := range words {
+				quoted[i] = regexp.QuoteMeta(word)
+			}
+			pattern, err := regexp.Compile(strings.Join(quoted, `\s+`))
+			if err != nil {
+				continue
+			}
+			subLinePattern = pattern
+		}
+		if match := subLinePattern.FindString(line); match != "" {
+			candidates = append(candidates, match)
+		}
+	}
+
+	findLines := strings.Split(find, "\n")
+	if len(findLines) > 1 {
+		for i := 0; i+len(findLines) <= len(contentLines); i++ {
+			block := strings.Join(contentLines[i:i+len(findLines)], "\n")
+			if normalizeEditWhitespace(block) == normalizedFind {
+				candidates = append(candidates, block)
+			}
+		}
+	}
+	return candidates
+}
+
+// stripCommonIndentation removes the minimum leading-whitespace width shared
+// by all non-empty lines, so blocks match regardless of their nesting depth.
+func stripCommonIndentation(text string) string {
+	lines := strings.Split(text, "\n")
+	minIndent := -1
+	for _, line := range lines {
+		if strings.TrimSpace(line) == "" {
+			continue
+		}
+		indent := len(line) - len(strings.TrimLeft(line, " \t"))
+		if minIndent < 0 || indent < minIndent {
+			minIndent = indent
+		}
+	}
+	if minIndent <= 0 {
+		return text
+	}
+	stripped := make([]string, len(lines))
+	for i, line := range lines {
+		if strings.TrimSpace(line) == "" {
+			stripped[i] = line
+			continue
+		}
+		stripped[i] = line[minIndent:]
+	}
+	return strings.Join(stripped, "\n")
+}
+
+func indentationFlexibleReplacer(content, find string) []string {
+	normalizedFind := stripCommonIndentation(find)
+	contentLines := strings.Split(content, "\n")
+	findLines := strings.Split(find, "\n")
+	var candidates []string
+	for i := 0; i+len(findLines) <= len(contentLines); i++ {
+		block := strings.Join(contentLines[i:i+len(findLines)], "\n")
+		if stripCommonIndentation(block) == normalizedFind {
+			candidates = append(candidates, block)
+		}
+	}
+	return candidates
+}
+
+var editEscapeSequence = regexp.MustCompile("\\\\(n|t|r|'|\"|`|\\\\|\\n|\\$)")
+
+// unescapeEditString undoes one level of string escaping (\n, \t, \", \\, a
+// backslash-newline continuation, \$) — the model sometimes reproduces file
+// content as it appeared inside a quoted string literal.
+func unescapeEditString(text string) string {
+	return editEscapeSequence.ReplaceAllStringFunc(text, func(match string) string {
+		switch match[1:] {
+		case "n":
+			return "\n"
+		case "t":
+			return "\t"
+		case "r":
+			return "\r"
+		case "\n":
+			return "\n"
+		default:
+			// ', ", `, \, $ all unescape to themselves.
+			return match[1:]
+		}
+	})
+}
+
+func escapeNormalizedReplacer(content, find string) []string {
+	unescapedFind := unescapeEditString(find)
+	var candidates []string
+	if strings.Contains(content, unescapedFind) {
+		candidates = append(candidates, unescapedFind)
+	}
+	contentLines := strings.Split(content, "\n")
+	findLines := strings.Split(unescapedFind, "\n")
+	for i := 0; i+len(findLines) <= len(contentLines); i++ {
+		block := strings.Join(contentLines[i:i+len(findLines)], "\n")
+		if unescapeEditString(block) == unescapedFind {
+			candidates = append(candidates, block)
+		}
+	}
+	return candidates
+}
+
+// trimmedBoundaryReplacer tolerates stray leading/trailing whitespace (often
+// blank lines) around an otherwise-exact old_string.
+func trimmedBoundaryReplacer(content, find string) []string {
+	trimmedFind := strings.TrimSpace(find)
+	if trimmedFind == find || trimmedFind == "" {
+		return nil
+	}
+	var candidates []string
+	if strings.Contains(content, trimmedFind) {
+		candidates = append(candidates, trimmedFind)
+	}
+	contentLines := strings.Split(content, "\n")
+	findLines := strings.Split(find, "\n")
+	for i := 0; i+len(findLines) <= len(contentLines); i++ {
+		block := strings.Join(contentLines[i:i+len(findLines)], "\n")
+		if strings.TrimSpace(block) == trimmedFind {
+			candidates = append(candidates, block)
+		}
+	}
+	return candidates
+}
+
+// contextAwareReplacer anchors on the first and last lines and accepts an
+// equal-length block when at least half of its non-empty middle lines match
+// after trimming — a cheaper, stricter cousin of blockAnchorReplacer.
+func contextAwareReplacer(content, find string) []string {
+	findLines := splitFindLines(find)
+	if len(findLines) < 3 {
+		return nil
+	}
+	contentLines := strings.Split(content, "\n")
+	firstAnchor := strings.TrimSpace(findLines[0])
+	lastAnchor := strings.TrimSpace(findLines[len(findLines)-1])
+	var candidates []string
+	for i := 0; i < len(contentLines); i++ {
+		if strings.TrimSpace(contentLines[i]) != firstAnchor {
+			continue
+		}
+		for j := i + 2; j < len(contentLines); j++ {
+			if strings.TrimSpace(contentLines[j]) != lastAnchor {
+				continue
+			}
+			if j-i+1 == len(findLines) {
+				matching, nonEmpty := 0, 0
+				for k := 1; k < len(findLines)-1; k++ {
+					blockLine := strings.TrimSpace(contentLines[i+k])
+					findLine := strings.TrimSpace(findLines[k])
+					if blockLine == "" && findLine == "" {
+						continue
+					}
+					nonEmpty++
+					if blockLine == findLine {
+						matching++
+					}
+				}
+				if nonEmpty == 0 || float64(matching)/float64(nonEmpty) >= 0.5 {
+					candidates = append(candidates, strings.Join(contentLines[i:j+1], "\n"))
+				}
+			}
+			break
+		}
+	}
+	return candidates
+}
+
+// adaptReplacementToSpan re-shapes the model's replacement to the span a
+// tolerant matcher resolved. When old_string only matched after normalization,
+// new_string was written at old_string's (wrong) shape, so applying it raw
+// would strip the file's indentation or drop a trailing CR:
+//
+//  1. Uniform re-indent: when every span line equals delta + the corresponding
+//     find line's indentation (the line-trimmed / indentation-flexible shapes),
+//     the same delta is prepended to every non-blank replacement line. Any
+//     line that breaks the uniform-delta relationship disables the shift —
+//     block-anchor matches with a drifted interior are left untouched.
+//  2. Trailing CR: a span from a CRLF file ends mid-line at "\r" (candidates
+//     are built by joining lines split on "\n"); the replacement gets the same
+//     trailing "\r" so the file's CRLF pairs stay intact.
+func adaptReplacementToSpan(span, find, replacement string) string {
+	if delta, ok := uniformIndentDelta(span, find); ok && delta != "" {
+		lines := strings.Split(replacement, "\n")
+		for i, line := range lines {
+			if strings.TrimSpace(line) == "" {
+				continue
+			}
+			lines[i] = delta + line
+		}
+		replacement = strings.Join(lines, "\n")
+	}
+	if strings.HasSuffix(span, "\r") && !strings.HasSuffix(replacement, "\r") {
+		replacement += "\r"
+	}
+	return replacement
+}
+
+// uniformIndentDelta returns the indentation prefix that, prepended to every
+// non-blank find line, yields the corresponding span line's indentation. ok is
+// false when line counts differ, any line pair disagrees on the delta, or the
+// span is not simply a uniformly deeper-indented copy of find.
+func uniformIndentDelta(span, find string) (string, bool) {
+	spanLines := strings.Split(span, "\n")
+	findLines := splitFindLines(find)
+	if len(spanLines) != len(findLines) {
+		return "", false
+	}
+	leadingWhitespace := func(line string) string {
+		return line[:len(line)-len(strings.TrimLeft(line, " \t"))]
+	}
+	delta := ""
+	haveDelta := false
+	for i := range findLines {
+		spanLine := strings.TrimSuffix(spanLines[i], "\r")
+		findLine := strings.TrimSuffix(findLines[i], "\r")
+		if strings.TrimSpace(spanLine) == "" && strings.TrimSpace(findLine) == "" {
+			continue
+		}
+		spanIndent := leadingWhitespace(spanLine)
+		findIndent := leadingWhitespace(findLine)
+		if !strings.HasSuffix(spanIndent, findIndent) {
+			return "", false
+		}
+		lineDelta := spanIndent[:len(spanIndent)-len(findIndent)]
+		if !haveDelta {
+			delta = lineDelta
+			haveDelta = true
+			continue
+		}
+		if lineDelta != delta {
+			return "", false
+		}
+	}
+	return delta, haveDelta
+}
+
+// levenshtein computes edit distance with a two-row rolling matrix.
+func levenshtein(a, b string) int {
+	if a == "" {
+		return len(b)
+	}
+	if b == "" {
+		return len(a)
+	}
+	previous := make([]int, len(b)+1)
+	current := make([]int, len(b)+1)
+	for j := range previous {
+		previous[j] = j
+	}
+	for i := 1; i <= len(a); i++ {
+		current[0] = i
+		for j := 1; j <= len(b); j++ {
+			cost := 1
+			if a[i-1] == b[j-1] {
+				cost = 0
+			}
+			minimum := previous[j] + 1
+			if current[j-1]+1 < minimum {
+				minimum = current[j-1] + 1
+			}
+			if previous[j-1]+cost < minimum {
+				minimum = previous[j-1] + cost
+			}
+			current[j] = minimum
+		}
+		previous, current = current, previous
+	}
+	return previous[len(b)]
+}

internal/tools/edit_replacers_test.go

@@ -0,0 +1,369 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func runEdit(t *testing.T, dir, initial string, args map[string]any) (Result, string) {
+	t.Helper()
+	path := filepath.Join(dir, "target.txt")
+	if err := os.WriteFile(path, []byte(initial), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if args["path"] == nil {
+		args["path"] = "target.txt"
+	}
+	result := NewEditFileTool(dir).Run(context.Background(), args)
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return result, string(after)
+}
+
+func TestEditExactMatchStillPreferred(t *testing.T) {
+	// The fast path must keep byte-exact semantics: an exact match wins even
+	// when fuzzy strategies would also find candidates elsewhere.
+	result, after := runEdit(t, t.TempDir(), "alpha\nbeta\ngamma\n", map[string]any{
+		"old_string": "beta",
+		"new_string": "delta",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if after != "alpha\ndelta\ngamma\n" {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyLineTrimmed(t *testing.T) {
+	// Model reproduced the lines without the file's leading indentation; the
+	// line-trimmed strategy must find the real indented span and preserve the
+	// file's own indentation on the replaced region boundary.
+	initial := "func a() {\n\tx := 1\n\treturn x\n}\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "x := 1\nreturn x",
+		"new_string": "y := 2\nreturn y",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	// new_string was written at old_string's outdented shape; the span's
+	// indentation must be re-applied so the file keeps its tabs.
+	if !strings.Contains(after, "\ty := 2\n\treturn y\n}") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+	if strings.Contains(after, "x := 1") {
+		t.Fatalf("old span survived: %q", after)
+	}
+}
+
+func TestEditFuzzyWhitespaceNormalizedSingleLine(t *testing.T) {
+	// Whitespace runs collapsed: "x  :=   1" in the file, single spaces from
+	// the model.
+	initial := "start\nx  :=   1\nend\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "x := 1",
+		"new_string": "x := 2",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if !strings.Contains(after, "x := 2") || strings.Contains(after, ":=   1") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyBlockAnchorToleratesMiddleDrift(t *testing.T) {
+	// First and last lines anchor the block; one interior line differs slightly
+	// (comment text drifted). Levenshtein similarity keeps it above 0.65.
+	initial := strings.Join([]string{
+		"func handler(w http.ResponseWriter, r *http.Request) {",
+		"\t// write the response body to the client",
+		"\tw.WriteHeader(http.StatusOK)",
+		"\tfmt.Fprint(w, \"done\")",
+		"}",
+		"",
+	}, "\n")
+	find := strings.Join([]string{
+		"func handler(w http.ResponseWriter, r *http.Request) {",
+		"\t// write the response body to client",
+		"\tw.WriteHeader(http.StatusOK)",
+		"\tfmt.Fprint(w, \"done\")",
+		"}",
+	}, "\n")
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": find,
+		"new_string": "func handler(w http.ResponseWriter, r *http.Request) {\n\tw.WriteHeader(http.StatusNoContent)\n}",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if !strings.Contains(after, "StatusNoContent") || strings.Contains(after, "StatusOK") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyIndentationFlexible(t *testing.T) {
+	// The whole block sits one nesting level deeper in the file than in the
+	// model's old_string, with interior relative indentation preserved.
+	initial := "if ok {\n\t\tfor i := range xs {\n\t\t\tsum += xs[i]\n\t\t}\n}\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "for i := range xs {\n\tsum += xs[i]\n}",
+		"new_string": "sum += total(xs)",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	// The replacement lands at the span's real depth (two tabs), not at
+	// old_string's outdented depth.
+	if !strings.Contains(after, "\t\tsum += total(xs)") || strings.Contains(after, "range xs") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyCRLFLineTrimmedPreservesCarriageReturns(t *testing.T) {
+	// CRLF file + a multi-line outdented old_string: the resolved span carries
+	// tabs and trailing CRs; the replacement must inherit both so the file's
+	// CRLF pairs and indentation survive the edit.
+	initial := "func a() {\r\n\tx := 1\r\n\treturn x\r\n}\r\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "x := 1\nreturn x",
+		"new_string": "y := 2\nreturn y",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if !strings.Contains(after, "\ty := 2\r\n\treturn y\r\n}") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+	if strings.Contains(after, "\n\treturn y\n") {
+		t.Fatalf("bare LF introduced into CRLF file: %q", after)
+	}
+}
+
+func TestUniformIndentDelta(t *testing.T) {
+	cases := []struct {
+		span, find string
+		want       string
+		ok         bool
+	}{
+		{"\tx := 1\n\treturn x", "x := 1\nreturn x", "\t", true},
+		{"\t\tfor {\n\t\t\tgo()\n\t\t}", "for {\n\tgo()\n}", "\t\t", true},
+		{"x := 1", "x := 1", "", true},                          // no shift needed
+		{"\tx := 1\n  return x", "x := 1\nreturn x", "", false}, // mixed delta
+		{"\ta\n\tb\n\tc", "a\nb", "", false},                    // line-count mismatch
+		{"  a", "\ta", "", false},                               // find indent not a suffix
+	}
+	for _, c := range cases {
+		got, ok := uniformIndentDelta(c.span, c.find)
+		if ok != c.ok || (ok && got != c.want) {
+			t.Errorf("uniformIndentDelta(%q, %q) = %q,%v want %q,%v", c.span, c.find, got, ok, c.want, c.ok)
+		}
+	}
+}
+
+func TestAdaptReplacementLeavesNonUniformSpansAlone(t *testing.T) {
+	// Block-anchor style match with a drifted interior: no uniform delta, so
+	// the replacement must pass through untouched.
+	span := "\tfunc h() {\n\t\t// drifted comment\n\t}"
+	find := "func h() {\n// different comment\n}"
+	if got := adaptReplacementToSpan(span, find, "replacement()"); got != "replacement()" {
+		t.Fatalf("non-uniform span must not re-indent: %q", got)
+	}
+}
+
+func TestEditFuzzyEscapeNormalized(t *testing.T) {
+	// Model reproduced the line as it would appear inside a quoted string
+	// literal, with escaped quotes.
+	initial := "console.log(\"hello world\")\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": `console.log(\"hello world\")`,
+		"new_string": `console.log("goodbye")`,
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if !strings.Contains(after, `console.log("goodbye")`) {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyTrimmedBoundary(t *testing.T) {
+	// Stray whitespace around an otherwise-exact old_string; "  two  " is not a
+	// substring of the file, so only the fuzzy cascade can resolve it.
+	initial := "one\ntwo\nthree\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "  two  ",
+		"new_string": "2",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if after != "one\n2\nthree\n" {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestTrimmedBoundaryReplacerYieldsTrimmedSpan(t *testing.T) {
+	candidates := trimmedBoundaryReplacer("one\ntwo\nthree\n", "\ntwo\n")
+	if len(candidates) == 0 || candidates[0] != "two" {
+		t.Fatalf("expected trimmed candidate \"two\", got %v", candidates)
+	}
+	if trimmedBoundaryReplacer("anything", "already-trimmed") != nil {
+		t.Fatal("already-trimmed find must yield no candidates")
+	}
+}
+
+// Two identical blocks that only match old_string after indentation-tolerant
+// matching: exact search finds nothing, fuzzy finds both.
+const ambiguousBlocks = "\tif x {\n\t\tgo()\n\t}\nmid\n\tif x {\n\t\tgo()\n\t}\n"
+
+func TestEditFuzzyAmbiguousReportsError(t *testing.T) {
+	// No strategy can disambiguate identical blocks, so the tool must refuse
+	// rather than guess, and the file must be untouched.
+	result, after := runEdit(t, t.TempDir(), ambiguousBlocks, map[string]any{
+		"old_string": "if x {\n\tgo()\n}",
+		"new_string": "stop()",
+	})
+	if result.Status != StatusError || !strings.Contains(result.Output, "multiple locations") {
+		t.Fatalf("expected ambiguity error, got %q", result.Output)
+	}
+	if after != ambiguousBlocks {
+		t.Fatalf("file must be unchanged, got %q", after)
+	}
+}
+
+func TestEditFuzzyReplaceAll(t *testing.T) {
+	// replace_all applies the fuzzy-resolved span at every occurrence.
+	result, after := runEdit(t, t.TempDir(), ambiguousBlocks, map[string]any{
+		"old_string":  "if x {\n\tgo()\n}",
+		"new_string":  "stop()",
+		"replace_all": true,
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if strings.Count(after, "stop()") != 2 || strings.Contains(after, "go()") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestEditFuzzyNotFoundKeepsExactError(t *testing.T) {
+	result, after := runEdit(t, t.TempDir(), "alpha\n", map[string]any{
+		"old_string": "omega",
+		"new_string": "x",
+	})
+	if result.Status != StatusError || !strings.Contains(result.Output, "Could not find the exact string") {
+		t.Fatalf("expected not-found error, got %q", result.Output)
+	}
+	if after != "alpha\n" {
+		t.Fatalf("file must be unchanged, got %q", after)
+	}
+}
+
+func TestEditFuzzyCRLFFile(t *testing.T) {
+	// CRLF file + LF old_string with indentation drift: the CRLF translation
+	// feeds the cascade and the replacement preserves CRLF endings.
+	initial := "func a() {\r\n\tx := 1\r\n}\r\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "x := 1",
+		"new_string": "x := 2",
+	})
+	if result.Status != StatusOK {
+		t.Fatalf("expected ok, got %q", result.Output)
+	}
+	if !strings.Contains(after, "\tx := 2\r\n") {
+		t.Fatalf("unexpected content: %q", after)
+	}
+}
+
+func TestIsDisproportionateEditMatch(t *testing.T) {
+	// A candidate span that dwarfs old_string must be refused: anchors bridging
+	// unrelated code would otherwise delete it all.
+	big := strings.Repeat("line\n", 10)
+	if !isDisproportionateEditMatch(big, "a\nb\nc") {
+		t.Fatal("10-line span for 3-line find must be disproportionate")
+	}
+	if isDisproportionateEditMatch("a\nb\nc\nd", "a\nb\nc") {
+		t.Fatal("one extra line within tolerance must be allowed")
+	}
+	if isDisproportionateEditMatch("single line", "single") {
+		t.Fatal("single-line finds are exempt from the byte-length guard")
+	}
+	if !isDisproportionateEditMatch("xx\n"+strings.Repeat("y", 900)+"\nzz", "xx\nab\nzz") {
+		t.Fatal("byte-length blowup must be disproportionate")
+	}
+}
+
+func TestLevenshtein(t *testing.T) {
+	cases := []struct {
+		a, b string
+		want int
+	}{
+		{"", "", 0},
+		{"abc", "", 3},
+		{"", "abc", 3},
+		{"kitten", "sitting", 3},
+		{"same", "same", 0},
+	}
+	for _, c := range cases {
+		if got := levenshtein(c.a, c.b); got != c.want {
+			t.Fatalf("levenshtein(%q,%q) = %d, want %d", c.a, c.b, got, c.want)
+		}
+	}
+}
+
+func TestEditFuzzyDistinctCandidatesAreAmbiguous(t *testing.T) {
+	// Two same-content blocks at DIFFERENT indentation: each resolved span is
+	// distinct and occurs exactly once, so the old literal-uniqueness check
+	// passed and silently edited the first block. Distinct-shaped candidates
+	// from one strategy must instead report ambiguity and leave the file alone.
+	initial := "\tif x {\n\t\tgo()\n\t}\nmid\n\t\tif x {\n\t\t\tgo()\n\t\t}\n"
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": "if x {\n\tgo()\n}",
+		"new_string": "stop()",
+	})
+	if result.Status != StatusError || !strings.Contains(result.Output, "multiple locations") {
+		t.Fatalf("distinct fuzzy candidates must be ambiguous, got %q", result.Output)
+	}
+	if after != initial {
+		t.Fatalf("file must be unchanged, got %q", after)
+	}
+}
+
+func TestEditFuzzyBlockAnchorTwoPlausibleBlocksAmbiguous(t *testing.T) {
+	// Two blocks share the same first/last anchor lines and BOTH interiors sit
+	// above the similarity threshold. Picking the "best" one would silently
+	// edit a block the model may not have meant; both must surface so the
+	// cascade reports ambiguity and the file stays untouched.
+	initial := strings.Join([]string{
+		"func setup(cfg Config) {",
+		"\tvalue := compute(alpha, beta, gamma)",
+		"}",
+		"",
+		"func setup(cfg Config) {",
+		"\tvalue := compute(alpha, delta, gamma)",
+		"}",
+		"",
+	}, "\n")
+	find := strings.Join([]string{
+		"func setup(cfg Config) {",
+		"\tvalue := compute(alpha, omega, gamma)",
+		"}",
+	}, "\n")
+	result, after := runEdit(t, t.TempDir(), initial, map[string]any{
+		"old_string": find,
+		"new_string": "func setup(cfg Config) {}",
+	})
+	if result.Status != StatusError || !strings.Contains(result.Output, "multiple locations") {
+		t.Fatalf("two plausible anchor blocks must be ambiguous, got %q", result.Output)
+	}
+	if after != initial {
+		t.Fatalf("file must be unchanged, got %q", after)
+	}
+}

internal/tools/exec_command.go

@@ -517,7 +517,12 @@ func (tool execCommandTool) run(ctx context.Context, args map[string]any, engine
 	if err != nil {
 		return errorResult("Error: Invalid arguments for exec_command: " + err.Error())
 	}
-	if issue := detectShellCommandIssue(commandText, runtimeGOOS()); issue != nil {
+	// Resolve the command engine before the MSYS preflight check so an
+	// approved require_escalated call (commandEngine == nil, truly
+	// unsandboxed) can actually bypass the MSYS guard instead of being
+	// hard-blocked by the same check it was meant to escalate past.
+	commandEngine := commandEngineForSandboxPermissions(engine, sandboxPermissions)
+	if issue := detectShellCommandIssue(commandText, runtimeGOOS()); issue != nil && !msysGuardBypassed(issue, commandEngine) {
 		return shellIssueBlockResult(*issue)
 	}
 	if interactive := zeroSandbox.DetectInteractiveCommand(commandText, runtimeGOOS()); interactive.Interactive {
@@ -860,6 +865,12 @@ func execToolResult(input execToolResultInput) Result {
 		status = StatusError
 	}
 	body := formatExecCommandOutput(output, input.sessionID, input.exited, input.exitCode, input.interrupted)
+	if status == StatusError && input.exited && !input.interrupted {
+		if issue := detectShellOutputIssue(output, runtimeGOOS()); issue != nil {
+			meta["shell_issue"] = issue.Kind
+			body = appendShellIssueHint(body, *issue)
+		}
+	}
 	if input.outputBufferTruncated {
 		// Appended after truncateExecOutput's own head/tail slicing, not
 		// embedded in the text that goes through it — a marker inside that
@@ -908,6 +919,15 @@ func formatExecCommandOutput(output string, sessionID int, exited bool, exitCode
 }
 
 func truncateExecOutput(output string, maxOutputTokens int) (string, bool) {
+	return truncateExecOutputSpill(output, maxOutputTokens, "exec_command")
+}
+
+// truncateExecOutputSpill keeps a head/tail window of the output within the
+// token budget and, on truncation, spills the full output to disk so the model
+// can grep/read the elided middle instead of re-running the command with a
+// bigger budget. The spill is best-effort: when it fails the notice simply
+// omits the file hint.
+func truncateExecOutputSpill(output string, maxOutputTokens int, toolName string) (string, bool) {
 	if maxOutputTokens <= 0 {
 		maxOutputTokens = defaultMaxOutputTokens
 	}
@@ -915,9 +935,13 @@ func truncateExecOutput(output string, maxOutputTokens int) (string, bool) {
 	if len(output) <= maxBytes {
 		return output, false
 	}
+	notice := "\n[zero] output truncated\n"
+	if spillPath := spillTruncatedOutput(toolName, output); spillPath != "" {
+		notice = "\n[zero] output truncated — full output saved to " + spillPath + " (grep or read_file it instead of re-running)\n"
+	}
 	head := maxBytes / 2
 	tail := maxBytes - head
-	return utf8Prefix(output, head) + "\n[zero] output truncated\n" + utf8Suffix(output, tail), true
+	return utf8Prefix(output, head) + notice + utf8Suffix(output, tail), true
 }
 
 func utf8Prefix(value string, maxBytes int) string {

internal/tools/exec_command_test.go

@@ -141,6 +141,46 @@ func TestExecCommandRequireEscalatedBypassesNativeSandboxAfterApproval(t *testin
 	}
 }
 
+// TestExecCommandRequireEscalatedBypassesMsysGuardAfterApproval mirrors
+// TestBashToolRequireEscalatedMsysGuard for exec_command: the MSYS sandbox
+// guard exists only because MSYS/Cygwin coreutils fail under the
+// write-restricted sandbox, so once require_escalated is actually approved
+// (unsandboxed execution), the guard must not block the same command it was
+// meant to let escalate past.
+func TestExecCommandRequireEscalatedBypassesMsysGuardAfterApproval(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows-only MSYS sandbox guard")
+	}
+	root := t.TempDir()
+	manager := newExecSessionManager()
+	registry := NewRegistry()
+	registry.Register(NewScopedExecCommandTool(root, nil, manager))
+	engine := sandbox.NewEngine(sandbox.EngineOptions{
+		WorkspaceRoot: root,
+		Policy:        sandbox.DefaultPolicy(),
+		Backend:       sandbox.Backend{Name: sandbox.BackendUnavailable, Message: "native sandbox unavailable"},
+	})
+
+	result := registry.RunWithOptions(context.Background(), ExecCommandToolName, map[string]any{
+		"cmd":                 "cat somefile.txt",
+		"sandbox_permissions": string(SandboxPermissionsRequireEscalated),
+	}, RunOptions{
+		PermissionGranted: true,
+		Sandbox:           engine,
+		PermissionMode:    string(sandbox.PermissionModeAsk),
+	})
+
+	// Assert on the preflight block sentinel (exit_code "-1", set only by
+	// shellIssueBlockResult) rather than shell_issue: once the guard is
+	// bypassed, "cat somefile.txt" actually runs, and its real,
+	// PATH-dependent output could otherwise trip the unrelated
+	// post-execution detectShellOutputIssue heuristic and make this
+	// assertion flaky for reasons unrelated to the guard under test.
+	if result.Meta["exit_code"] == "-1" {
+		t.Fatalf("expected approved require_escalated to bypass the MSYS guard, got blocked: %#v", result)
+	}
+}
+
 func TestExecCommandReturnsExitCodeWhenCommandCompletesDuringInitialYield(t *testing.T) {
 	root := t.TempDir()
 	manager := newExecSessionManager()

internal/tools/format_on_write.go

@@ -0,0 +1,97 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+// Format-on-write for the mutating file tools. When enabled, a successful
+// edit_file/write_file runs the language's standard formatter on the file it
+// just wrote, so the model's output always lands in project-canonical style
+// and never fails a CI format check it cannot see. Off by default (set
+// ZERO_FORMAT_ON_WRITE=1): auto-reformatting changes bytes the model did not
+// write, which strict workflows may not want.
+//
+// Ordering matters: formatting runs BEFORE the FileTracker re-baseline, and
+// the caller records the POST-format content. Formatting after the baseline
+// would make the very next edit look like an external modification and trip
+// the conflict guard.
+
+// formatOnWriteTimeout bounds one formatter run; a wedged formatter must never
+// hang a tool call. On timeout the unformatted write stands.
+const formatOnWriteTimeout = 10 * time.Second
+
+// formatterCommands maps a file extension to the formatter argv; the file path
+// is appended as the final argument. Only in-place, config-respecting,
+// community-standard formatters — a missing binary silently skips formatting.
+var formatterCommands = map[string][]string{
+	".go":    {"gofmt", "-w"},
+	".rs":    {"rustfmt"},
+	".py":    {"ruff", "format", "--quiet"},
+	".ts":    {"prettier", "--log-level", "silent", "--write"},
+	".tsx":   {"prettier", "--log-level", "silent", "--write"},
+	".js":    {"prettier", "--log-level", "silent", "--write"},
+	".jsx":   {"prettier", "--log-level", "silent", "--write"},
+	".json":  {"prettier", "--log-level", "silent", "--write"},
+	".css":   {"prettier", "--log-level", "silent", "--write"},
+	".scss":  {"prettier", "--log-level", "silent", "--write"},
+	".html":  {"prettier", "--log-level", "silent", "--write"},
+	".md":    {"prettier", "--log-level", "silent", "--write"},
+	".yaml":  {"prettier", "--log-level", "silent", "--write"},
+	".yml":   {"prettier", "--log-level", "silent", "--write"},
+	".zig":   {"zig", "fmt"},
+	".dart":  {"dart", "format"},
+	".tf":    {"terraform", "fmt"},
+	".gleam": {"gleam", "format"},
+	".sh":    {"shfmt", "-w"},
+	".bash":  {"shfmt", "-w"},
+	".c":     {"clang-format", "-i"},
+	".h":     {"clang-format", "-i"},
+	".cpp":   {"clang-format", "-i"},
+	".hpp":   {"clang-format", "-i"},
+	".cc":    {"clang-format", "-i"},
+	".kt":    {"ktlint", "-F"},
+	".swift": {"swiftformat"},
+	".lua":   {"stylua"},
+}
+
+// formatOnWriteEnabled reports whether the opt-in env toggle is set.
+func formatOnWriteEnabled() bool {
+	value := strings.TrimSpace(os.Getenv("ZERO_FORMAT_ON_WRITE"))
+	return value != "" && value != "0" && !strings.EqualFold(value, "false")
+}
+
+// maybeFormatWrittenFile runs the configured formatter for absolutePath (when
+// enabled and on PATH) and returns the file's content afterwards. Best-effort
+// throughout: any failure — no formatter, formatter error, timeout, unreadable
+// result — returns writtenContent so the caller's state matches the last write
+// it performed itself.
+func maybeFormatWrittenFile(ctx context.Context, absolutePath string, writtenContent string) string {
+	if !formatOnWriteEnabled() {
+		return writtenContent
+	}
+	command, ok := formatterCommands[strings.ToLower(filepath.Ext(absolutePath))]
+	if !ok {
+		return writtenContent
+	}
+	if _, err := exec.LookPath(command[0]); err != nil {
+		return writtenContent
+	}
+	formatCtx, cancel := context.WithTimeout(ctx, formatOnWriteTimeout)
+	defer cancel()
+	arguments := append(append([]string(nil), command[1:]...), absolutePath)
+	formatter := exec.CommandContext(formatCtx, command[0], arguments...)
+	formatter.Dir = filepath.Dir(absolutePath)
+	if err := formatter.Run(); err != nil {
+		return writtenContent
+	}
+	formatted, err := os.ReadFile(absolutePath)
+	if err != nil {
+		return writtenContent
+	}
+	return string(formatted)
+}

internal/tools/format_on_write_test.go

@@ -0,0 +1,82 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// gofmt ships with the Go toolchain, so it is the one formatter guaranteed to
+// exist wherever these tests run.
+func requireGofmt(t *testing.T) {
+	t.Helper()
+	if _, err := exec.LookPath("gofmt"); err != nil {
+		t.Skip("gofmt not on PATH")
+	}
+}
+
+func TestFormatOnWriteDisabledByDefault(t *testing.T) {
+	requireGofmt(t)
+	t.Setenv("ZERO_FORMAT_ON_WRITE", "")
+	dir := t.TempDir()
+	ugly := "package a\n\nfunc  A( ) {   }\n"
+
+	result := NewWriteFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":    "a.go",
+		"content": ugly,
+	}, RunOptions{})
+	if result.Status != StatusOK {
+		t.Fatalf("write failed: %q", result.Output)
+	}
+	onDisk, err := os.ReadFile(filepath.Join(dir, "a.go"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(onDisk) != ugly {
+		t.Fatalf("formatting must be off by default, got %q", onDisk)
+	}
+}
+
+func TestFormatOnWriteFormatsAndKeepsTrackerConsistent(t *testing.T) {
+	requireGofmt(t)
+	t.Setenv("ZERO_FORMAT_ON_WRITE", "1")
+	dir := t.TempDir()
+	tracker := NewFileTracker()
+
+	write := NewWriteFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":    "a.go",
+		"content": "package a\n\nfunc  A( ) {   }\n",
+	}, RunOptions{FileTracker: tracker})
+	if write.Status != StatusOK {
+		t.Fatalf("write failed: %q", write.Output)
+	}
+	onDisk, err := os.ReadFile(filepath.Join(dir, "a.go"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(onDisk), "func A() {") {
+		t.Fatalf("expected gofmt-formatted content, got %q", onDisk)
+	}
+
+	// The tracker must have been re-baselined to the POST-format content: a
+	// follow-up edit must not trip the external-modification conflict guard.
+	edit := NewEditFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":       "a.go",
+		"old_string": "func A() {",
+		"new_string": "func B() {",
+	}, RunOptions{FileTracker: tracker})
+	if edit.Status != StatusOK {
+		t.Fatalf("follow-up edit must not conflict after formatting: %q", edit.Output)
+	}
+}
+
+func TestFormatOnWriteSkipsUnknownExtensions(t *testing.T) {
+	t.Setenv("ZERO_FORMAT_ON_WRITE", "1")
+	content := maybeFormatWrittenFile(context.Background(), filepath.Join(t.TempDir(), "notes.xyz"), "raw   text")
+	if content != "raw   text" {
+		t.Fatalf("unknown extension must pass through: %q", content)
+	}
+}

internal/tools/glob.go

@@ -45,18 +45,18 @@ func NewScopedGlobTool(workspaceRoot string, scope PathScope) Tool {
 	}
 }
 
-func (tool globTool) Run(_ context.Context, args map[string]any) Result {
-	return tool.runWith(args, readExcluder{})
+func (tool globTool) Run(ctx context.Context, args map[string]any) Result {
+	return tool.runWith(ctx, args, readExcluder{})
 }
 
 // RunWithSandbox runs glob while skipping subtrees the sandbox policy denies
 // reads to (DenyRead). With no DenyRead configured the excluder is a no-op and
 // behavior is unchanged.
-func (tool globTool) RunWithSandbox(_ context.Context, args map[string]any, engine *sandbox.Engine) Result {
-	return tool.runWith(args, sandboxReadExcluder(engine))
+func (tool globTool) RunWithSandbox(ctx context.Context, args map[string]any, engine *sandbox.Engine) Result {
+	return tool.runWith(ctx, args, sandboxReadExcluder(engine))
 }
 
-func (tool globTool) runWith(args map[string]any, exclude readExcluder) Result {
+func (tool globTool) runWith(ctx context.Context, args map[string]any, exclude readExcluder) Result {
 	pattern, err := aliasedStringArg(args, []string{"pattern", "glob", "match", "query", "expression"}, "", true, false)
 	if err != nil {
 		return errorResult("Error: Invalid arguments for glob: " + err.Error())
@@ -88,8 +88,11 @@ func (tool globTool) runWith(args map[string]any, exclude readExcluder) Result {
 		return errorResult("Error running glob " + fmt.Sprintf("%q", pattern) + ": " + err.Error())
 	}
 
-	matches, err := scanGlob(root, displayRoot, matcher, includeDirs, exclude)
+	matches, err := scanGlob(ctx, root, displayRoot, matcher, includeDirs, exclude)
 	if err != nil {
+		if res, ok := searchCancelledResult("glob", err); ok {
+			return res
+		}
 		return errorResult("Error running glob " + fmt.Sprintf("%q", pattern) + ": " + err.Error())
 	}
 	if len(matches) == 0 {
@@ -126,9 +129,15 @@ func (tool globTool) runWith(args map[string]any, exclude readExcluder) Result {
 	}
 }
 
-func scanGlob(root string, displayRoot string, matcher *regexp.Regexp, includeDirs bool, exclude readExcluder) ([]string, error) {
+func scanGlob(ctx context.Context, root string, displayRoot string, matcher *regexp.Regexp, includeDirs bool, exclude readExcluder) ([]string, error) {
 	matches := []string{}
 	err := filepath.WalkDir(root, func(path string, entry fs.DirEntry, walkErr error) error {
+		// Checked first, ahead of walkErr: an unscoped scan over a large tree can
+		// run long enough that cancelling the run must stop the walk promptly
+		// rather than visiting every remaining entry to completion.
+		if err := ctx.Err(); err != nil {
+			return err
+		}
 		if walkErr != nil {
 			if path == root {
 				return walkErr

internal/tools/grep.go

@@ -55,19 +55,19 @@ func NewScopedGrepTool(workspaceRoot string, scope PathScope) Tool {
 	}
 }
 
-func (tool grepTool) Run(_ context.Context, args map[string]any) Result {
-	return tool.runWith(args, readExcluder{})
+func (tool grepTool) Run(ctx context.Context, args map[string]any) Result {
+	return tool.runWith(ctx, args, readExcluder{})
 }
 
 // RunWithSandbox runs the search while skipping subtrees the sandbox policy
 // denies reads to (DenyRead), so grep never surfaces content from a read-denied
 // path. With no DenyRead configured the excluder is a no-op and behavior is
 // unchanged.
-func (tool grepTool) RunWithSandbox(_ context.Context, args map[string]any, engine *sandbox.Engine) Result {
-	return tool.runWith(args, sandboxReadExcluder(engine))
+func (tool grepTool) RunWithSandbox(ctx context.Context, args map[string]any, engine *sandbox.Engine) Result {
+	return tool.runWith(ctx, args, sandboxReadExcluder(engine))
 }
 
-func (tool grepTool) runWith(args map[string]any, exclude readExcluder) Result {
+func (tool grepTool) runWith(ctx context.Context, args map[string]any, exclude readExcluder) Result {
 	pattern, err := aliasedStringArg(args, []string{"pattern", "query", "regex", "search", "expression"}, "", true, false)
 	if err != nil {
 		return errorResult("Error: Invalid arguments for grep: " + err.Error())
@@ -138,15 +138,24 @@ func (tool grepTool) runWith(args map[string]any, exclude readExcluder) Result {
 		}
 	}
 
-	files, err := grepFiles(resolvedRoot, target, globMatcher, exclude)
+	files, err := grepFiles(ctx, resolvedRoot, target, globMatcher, exclude)
 	if err != nil {
+		if res, ok := searchCancelledResult("grep", err); ok {
+			return res
+		}
 		return errorResult("Error running grep: " + err.Error())
 	}
 
 	// resolveScopedPath returns an absolute displayRoot when the target resolved
 	// to an extra (non-workspace) granted root; emit absolute match paths there so
 	// they survive a round-trip through read_file/edit_file (mirrors glob).
-	matches := collectGrepMatches(resolvedRoot, filepath.IsAbs(displayRoot), files, compiled)
+	matches, err := collectGrepMatches(ctx, resolvedRoot, filepath.IsAbs(displayRoot), files, compiled)
+	if err != nil {
+		if res, ok := searchCancelledResult("grep", err); ok {
+			return res
+		}
+		return errorResult("Error running grep: " + err.Error())
+	}
 	if len(matches) == 0 {
 		if outputMode == "count" {
 			return okResult("0 matches found")
@@ -261,7 +270,7 @@ func confineGrepFile(resolvedRoot string, path string) (string, string, bool) {
 	return filepath.ToSlash(relative), resolved, true
 }
 
-func grepFiles(resolvedRoot string, target string, globMatcher *regexp.Regexp, exclude readExcluder) ([]string, error) {
+func grepFiles(ctx context.Context, resolvedRoot string, target string, globMatcher *regexp.Regexp, exclude readExcluder) ([]string, error) {
 	info, err := os.Stat(target)
 	if err != nil {
 		return nil, err
@@ -288,6 +297,13 @@ func grepFiles(resolvedRoot string, target string, globMatcher *regexp.Regexp, e
 
 	files := []string{}
 	err = filepath.WalkDir(target, func(path string, entry os.DirEntry, walkErr error) error {
+		// Checked first, ahead of walkErr: an unscoped search over a large tree
+		// (e.g. a broad parent directory, not just the workspace) can run long
+		// enough that cancelling the run must stop the walk promptly rather than
+		// visiting every remaining entry to completion.
+		if err := ctx.Err(); err != nil {
+			return err
+		}
 		if walkErr != nil {
 			if path == target {
 				return walkErr
@@ -342,9 +358,16 @@ func grepFiles(resolvedRoot string, target string, globMatcher *regexp.Regexp, e
 	return files, nil
 }
 
-func collectGrepMatches(resolvedRoot string, absolutePaths bool, files []string, compiled *regexp.Regexp) []grepMatch {
+func collectGrepMatches(ctx context.Context, resolvedRoot string, absolutePaths bool, files []string, compiled *regexp.Regexp) ([]grepMatch, error) {
 	matches := []grepMatch{}
 	for _, file := range files {
+		// Checked per file, matching the walk's per-entry check in grepFiles: a
+		// broad search can already have collected many files before the run is
+		// cancelled, and reading/regex-scanning each one is real work that must
+		// stop promptly rather than running to completion.
+		if err := ctx.Err(); err != nil {
+			return matches, err
+		}
 		// Re-confine at read time (defense-in-depth) AND to compute the clean
 		// workspace-relative path used in output.
 		relative, resolvedPath, ok := confineGrepFile(resolvedRoot, file)
@@ -381,5 +404,5 @@ func collectGrepMatches(resolvedRoot string, absolutePaths bool, files []string,
 			})
 		}
 	}
-	return matches
+	return matches, nil
 }

internal/tools/inline_diagnostics.go

@@ -0,0 +1,18 @@
+package tools
+
+import "context"
+
+// inlineDiagnostics renders the post-write diagnostics block a mutating tool
+// appends to its output. Empty when no Diagnostics callback is wired, the file
+// is clean, or no language server is available — the tool output is then
+// byte-identical to the pre-diagnostics behavior.
+func inlineDiagnostics(ctx context.Context, options RunOptions, absolutePath, relativePath string) string {
+	if options.Diagnostics == nil {
+		return ""
+	}
+	block := options.Diagnostics(ctx, absolutePath)
+	if block == "" {
+		return ""
+	}
+	return "\n\nDiagnostics in " + relativePath + " after this change (fix any errors you introduced):\n" + block
+}

internal/tools/inline_diagnostics_test.go

@@ -0,0 +1,53 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// A wired Diagnostics callback must surface its block in the tool output for
+// both mutating tools, and a clean file (empty block) must leave the output
+// byte-identical to the pre-diagnostics behavior.
+func TestMutatingToolsAppendInlineDiagnostics(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "a.go"), []byte("package a\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	diagnostics := func(_ context.Context, absPath string) string {
+		return absPath + ":1:1: error: undefined: x"
+	}
+
+	editResult := NewEditFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":       "a.go",
+		"old_string": "package a",
+		"new_string": "package b",
+	}, RunOptions{Diagnostics: diagnostics})
+	if editResult.Status != StatusOK {
+		t.Fatalf("edit failed: %q", editResult.Output)
+	}
+	if !strings.Contains(editResult.Output, "Diagnostics in a.go after this change") || !strings.Contains(editResult.Output, "undefined: x") {
+		t.Fatalf("edit output missing diagnostics block: %q", editResult.Output)
+	}
+
+	writeResult := NewWriteFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":    "b.go",
+		"content": "package b\n",
+	}, RunOptions{Diagnostics: diagnostics})
+	if writeResult.Status != StatusOK {
+		t.Fatalf("write failed: %q", writeResult.Output)
+	}
+	if !strings.Contains(writeResult.Output, "Diagnostics in b.go after this change") {
+		t.Fatalf("write output missing diagnostics block: %q", writeResult.Output)
+	}
+
+	clean := NewWriteFileTool(dir).(optionsAwareTool).RunWithOptions(context.Background(), map[string]any{
+		"path":    "c.go",
+		"content": "package c\n",
+	}, RunOptions{Diagnostics: func(context.Context, string) string { return "" }})
+	if strings.Contains(clean.Output, "Diagnostics") {
+		t.Fatalf("clean file must not gain a diagnostics block: %q", clean.Output)
+	}
+}

internal/tools/output_ceiling.go

@@ -0,0 +1,81 @@
+package tools
+
+import (
+	"os"
+	"strconv"
+	"strings"
+)
+
+// Universal output ceiling. Every tool result leaving the registry boundary is
+// capped at a single token-denominated budget unless the tool manages its own
+// deliberate budget (selfBudgeting below). This is the safety net for tools
+// with no cap of their own — web_fetch bodies, skill files, MCP server tools,
+// browser snapshots, anything added later — so no single call can flood the
+// context window with output that is then re-billed on every following turn.
+// Truncation keeps head+tail and spills the full text to disk (re-readable via
+// grep/read_file), so nothing is lost, only deferred.
+
+// defaultOutputCeilingTokens is the per-call ceiling in estimated tokens
+// (bytes/4). 16k tokens = 64 KiB — deliberately equal to the search-tool
+// budget: generous enough for a large fetched document, small enough that one
+// call cannot eat a third of a small context window.
+const defaultOutputCeilingTokens = 16_000
+
+// outputCeilingEnv overrides the ceiling (in tokens). Zero or negative
+// disables the ceiling entirely; unset or unparsable keeps the default.
+const outputCeilingEnv = "ZERO_TOOL_OUTPUT_CEILING_TOKENS"
+
+// selfBudgeting marks a tool that enforces its own deliberate output budget —
+// possibly model-raisable (exec_command) — which the registry ceiling must not
+// second-guess. The method is unexported on purpose: only tools in this
+// package can opt out, so an MCP-served tool can never exempt itself.
+type selfBudgeting interface{ managesOutputBudget() }
+
+// The exemption list, kept in one place. Each of these applies its own budget
+// before returning: bash (bashOutputBudgetBytes per stream + spill),
+// exec_command (model-raisable token budget + spill), read tools (128 KiB),
+// search tools (64 KiB).
+func (bashTool) managesOutputBudget()             {}
+func (execCommandTool) managesOutputBudget()      {}
+func (readFileTool) managesOutputBudget()         {}
+func (readMinifiedFileTool) managesOutputBudget() {}
+func (grepTool) managesOutputBudget()             {}
+func (globTool) managesOutputBudget()             {}
+func (listDirectoryTool) managesOutputBudget()    {}
+
+func resolveOutputCeilingTokens() int {
+	raw := strings.TrimSpace(os.Getenv(outputCeilingEnv))
+	if raw == "" {
+		return defaultOutputCeilingTokens
+	}
+	parsed, err := strconv.Atoi(raw)
+	if err != nil {
+		return defaultOutputCeilingTokens
+	}
+	return parsed
+}
+
+// enforceOutputCeiling truncates an over-ceiling result to head+tail within
+// the token budget, spilling the full (already redaction-scrubbed) output to
+// disk with a recovery hint. Runs after scrubResultSecrets so the transcript
+// and the spill file agree on what was hidden.
+func enforceOutputCeiling(toolName string, result Result) Result {
+	ceiling := resolveOutputCeilingTokens()
+	if ceiling <= 0 {
+		return result
+	}
+	if len(result.Output) <= ceiling*4 {
+		return result
+	}
+	rawBytes := len(result.Output)
+	truncated, _ := truncateExecOutputSpill(result.Output, ceiling, toolName)
+	result.Output = truncated
+	result.Truncated = true
+	if result.Meta == nil {
+		result.Meta = map[string]string{}
+	}
+	result.Meta["raw_bytes"] = strconv.Itoa(rawBytes)
+	result.Meta["emitted_bytes"] = strconv.Itoa(len(result.Output))
+	result.Meta["estimated_tokens"] = strconv.Itoa(estimatedTokensFromBytes(len(result.Output)))
+	return result
+}

internal/tools/output_ceiling_test.go

@@ -0,0 +1,140 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// ceilingFakeTool is a minimal tool with no output budget of its own — the
+// stand-in for an MCP-served tool or any future builtin that forgets to cap
+// its output.
+type ceilingFakeTool struct {
+	baseTool
+	output string
+}
+
+func (tool ceilingFakeTool) Run(context.Context, map[string]any) Result {
+	return okResult(tool.output)
+}
+
+// ceilingExemptTool is the same fake but marked self-budgeting.
+type ceilingExemptTool struct{ ceilingFakeTool }
+
+func (ceilingExemptTool) managesOutputBudget() {}
+
+func newCeilingFakeTool(name, output string) ceilingFakeTool {
+	return ceilingFakeTool{
+		baseTool: baseTool{name: name, safety: readOnlySafety("test tool")},
+		output:   output,
+	}
+}
+
+// An unbudgeted tool's oversized output is capped at the universal ceiling,
+// head and tail survive, and the full output is spilled to a re-readable file.
+func TestOutputCeilingCapsUnbudgetedTool(t *testing.T) {
+	setTestTempDir(t)
+	big := "HEAD_MARK\n" + strings.Repeat("z", defaultOutputCeilingTokens*4*3) + "\nTAIL_MARK"
+
+	registry := NewRegistry()
+	registry.Register(newCeilingFakeTool("fake_big", big))
+	result := registry.Run(context.Background(), "fake_big", map[string]any{})
+
+	if !result.Truncated {
+		t.Fatal("over-ceiling output must report truncated=true")
+	}
+	maxBytes := defaultOutputCeilingTokens * 4
+	if len(result.Output) > maxBytes+512 {
+		t.Fatalf("emitted %d bytes, ceiling is %d", len(result.Output), maxBytes)
+	}
+	if !strings.Contains(result.Output, "HEAD_MARK") || !strings.Contains(result.Output, "TAIL_MARK") {
+		t.Fatal("head/tail lost at the ceiling")
+	}
+	if result.Meta["raw_bytes"] != strconv.Itoa(len(big)) {
+		t.Fatalf("raw_bytes = %s, want %d", result.Meta["raw_bytes"], len(big))
+	}
+	if !strings.Contains(result.Output, "full output saved to ") {
+		t.Fatalf("ceiling truncation must include a spill hint: %q", result.Output[:200])
+	}
+	start := strings.Index(result.Output, "full output saved to ") + len("full output saved to ")
+	end := strings.Index(result.Output[start:], " (grep")
+	content, err := os.ReadFile(result.Output[start : start+end])
+	if err != nil {
+		t.Fatalf("spill file unreadable: %v", err)
+	}
+	if string(content) != big {
+		t.Fatalf("spill must hold the full output: got %d bytes, want %d", len(content), len(big))
+	}
+}
+
+// A self-budgeting tool's output passes the boundary untouched even when it
+// exceeds the universal ceiling — its own (possibly model-raised) budget wins.
+func TestOutputCeilingSkipsSelfBudgetingTool(t *testing.T) {
+	big := strings.Repeat("z", defaultOutputCeilingTokens*4*2)
+	registry := NewRegistry()
+	registry.Register(ceilingExemptTool{newCeilingFakeTool("fake_exempt", big)})
+
+	result := registry.Run(context.Background(), "fake_exempt", map[string]any{})
+	if result.Truncated || result.Output != big {
+		t.Fatalf("self-budgeting tool must bypass the ceiling (truncated=%v, %d bytes)", result.Truncated, len(result.Output))
+	}
+}
+
+// Under-ceiling output is untouched and gains no budget metadata.
+func TestOutputCeilingSmallOutputUntouched(t *testing.T) {
+	registry := NewRegistry()
+	registry.Register(newCeilingFakeTool("fake_small", "hello"))
+
+	result := registry.Run(context.Background(), "fake_small", map[string]any{})
+	if result.Truncated || result.Output != "hello" {
+		t.Fatalf("small output altered: %+v", result)
+	}
+	if _, ok := result.Meta["raw_bytes"]; ok {
+		t.Fatal("small output must not gain budget metadata")
+	}
+}
+
+// ZERO_TOOL_OUTPUT_CEILING_TOKENS tightens, loosens, or disables the ceiling.
+func TestOutputCeilingEnvOverride(t *testing.T) {
+	setTestTempDir(t)
+	big := strings.Repeat("z", 100*1024)
+
+	t.Setenv(outputCeilingEnv, "1000")
+	registry := NewRegistry()
+	registry.Register(newCeilingFakeTool("fake_env", big))
+	result := registry.Run(context.Background(), "fake_env", map[string]any{})
+	if !result.Truncated || len(result.Output) > 1000*4+512 {
+		t.Fatalf("ceiling override ignored: truncated=%v, %d bytes", result.Truncated, len(result.Output))
+	}
+
+	t.Setenv(outputCeilingEnv, "0")
+	result = registry.Run(context.Background(), "fake_env", map[string]any{})
+	if result.Truncated || len(result.Output) != len(big) {
+		t.Fatalf("ceiling=0 must disable the net: truncated=%v, %d bytes", result.Truncated, len(result.Output))
+	}
+}
+
+// The exemption list matches the tools that really do budget themselves, and
+// the unbudgeted ones (web_fetch, skill, browser) stay under the net.
+func TestSelfBudgetingExemptionList(t *testing.T) {
+	dir := t.TempDir()
+	exempt := []Tool{
+		NewBashTool(dir),
+		NewExecCommandTool(dir, newExecSessionManager()),
+		NewReadFileTool(dir),
+		NewReadMinifiedFileTool(dir),
+		NewGrepTool(dir),
+		NewGlobTool(dir),
+		NewListDirectoryTool(dir),
+	}
+	for _, tool := range exempt {
+		if _, ok := tool.(selfBudgeting); !ok {
+			t.Errorf("%s must be exempt from the output ceiling", tool.Name())
+		}
+	}
+	if _, ok := NewWebFetchTool().(selfBudgeting); ok {
+		t.Error("web_fetch must NOT be exempt — the ceiling is its only budget")
+	}
+}

internal/tools/registry.go

@@ -40,6 +40,12 @@ type RunOptions struct {
 	// tool-call progress in the specialist card. nil is a no-op (the default
 	// for every non-Task tool).
 	Progress func(streamjson.Event)
+	// Diagnostics, when set, returns a formatted language-diagnostics block for
+	// a file a mutating tool just wrote ("" when clean or no server available).
+	// edit_file/write_file append it to their output so the model sees an error
+	// it introduced in the same turn instead of waiting for a later verification
+	// pass. nil disables inline diagnostics.
+	Diagnostics func(ctx context.Context, absPath string) string
 }
 
 type sandboxAwareTool interface {
@@ -117,13 +123,24 @@ func (registry *Registry) Run(ctx context.Context, name string, args map[string]
 func (registry *Registry) RunWithOptions(ctx context.Context, name string, args map[string]any, options RunOptions) (result Result) {
 	// Every return path passes through scrubResultSecrets exactly once, so denial,
 	// permission, and unknown-tool error messages (which can echo secret-bearing
-	// args/paths) are redacted at the boundary just like tool output.
-	defer func() { result = scrubResultSecrets(result) }()
+	// args/paths) are redacted at the boundary just like tool output. The output
+	// ceiling runs after the scrub so the transcript and the spill file agree on
+	// what was hidden.
+	ceilingExempt := false
+	defer func() {
+		result = scrubResultSecrets(result)
+		if !ceilingExempt {
+			result = enforceOutputCeiling(name, result)
+		}
+	}()
 
 	tool, ok := registry.Get(name)
 	if !ok {
 		return errorResult(`Error: Unknown tool "` + name + `".`)
 	}
+	if _, ok := tool.(selfBudgeting); ok {
+		ceilingExempt = true
+	}
 	if rejecter, ok := tool.(PrePermissionRejecter); ok {
 		if res, rejected := rejecter.RejectBeforePermission(args); rejected {
 			return res

internal/tools/search_cancel.go

@@ -0,0 +1,13 @@
+package tools
+
+import (
+	"context"
+	"errors"
+)
+
+func searchCancelledResult(tool string, err error) (Result, bool) {
+	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+		return Result{Status: StatusError, Output: "Error: " + tool + " cancelled."}, true
+	}
+	return Result{}, false
+}