Problem: the release fragment mentioned static HTML support without explicitly
calling out the change from automatic live rendering.
Fix: state in Chinese and English that HTML opens with scripts disabled and
desktop users select Interactive run to execute the page, as requested in review.
Verification: release-note script tests, make check and whitespace validation.
No runtime behavior, dependencies or test expectations changed.
Problem: frontend-next CI failed 25 assertions across action discovery,
orphan detection, HTML preview races and copy feedback consumers.
Root cause: TooltipButton was not declared as an interaction primitive; new
preview controls had no action identities, and the superseded image component
remained orphaned. Neighbouring tests still expected native title tooltips and
automatic HTML execution rather than the explicit Interactive run action.
Fix: certify and name the actual UI inputs, restrict tooltip blur to local
dismissal, remove the unused image renderer, and align consumer tests with the
new contracts while retaining copy failure/retry and browser race coverage.
The unnamed-input ceiling falls from 218 to 217 rather than being relaxed.
Verification: full frontend suite (749 files, 6010 tests); supplemental action,
tooltip and browser guards; typecheck and production build; go vet, make lint
and make check. No workflow skip, timeout increase or alert suppression.
Problem: delivery-card Add to conversation bypassed the workbench's reference
validation, allowing names with token boundaries or MCP ambiguity into drafts.
Root cause: Pane independently formatted @ references from display paths.
The kernel grammar preserves literal backslashes and only escapes space/tab,
so generic backslash doubling would change the selected file.
Fix: share one grammar-specific formatter, use structured relative paths for
workspace cards, disable unrepresentable references while keeping preview and
path copying, and remove the unused intake refToken helper. Keyboard navigation
and initial focus skip disabled menu actions.
Verification: 170 related frontend tests, production build/typecheck, real
browser insertion, kernel backslash round-trip and scoped-resolution tests,
go vet, make lint and make check. Four new rejection cases fail before the fix.
Problem: the presenter was registered for every install, and review requested
default-off model behavior, repository screenshot removal and PDF hardening.
Root cause: present_files used the unconditional builtin path rather than the
existing user-scoped conditional registration and provider-surface mechanism.
Fix: gate registration with tools.present_files, protect user scope across
project merges, preserve the setting on save, and expose the tool only after
opt-in. Restore the unconditional schema budget and remove PR screenshots.
Document PDF.js 6's removal of eval/isEvalSupported, keep scripting disconnected
and explicitly disable XFA rather than passing an unsupported no-op option.
Verification: boot.Build direct/proxy off/on/forbidden/disabled/excluded effect
matrix; user-scope and save/reload tests; document tests and typecheck; real
browser PDF navigation; go vet, make lint, make check and release-note tests.
Record the user-visible preview formats, responsive delivery cards and refresh fixes for PR #14031. Verified with make check and git diff --check; no runtime behavior changed.
Problem: AI-created documents could not be handed off as validated, durable
file cards, and binary files could not be read through the text editor route.
Root cause: delivery identity lived only in ordinary tool output, while file
reading, view selection and auxiliary actions lacked a shared preview path.
Fix: add a bounded root-confined reader, Controller and raw-byte preview APIs,
and host-confirmed present_files results projected through live and stored
history. Add lazy PDF, worker CSV and sandboxed HTML viewers, responsive file
cards, type icons, current-file state and accessible shared tooltips. Preserve
dirty drafts while surfacing deletion/revocation and avoid metadata reads per
keystroke. Keep the current workbench request-consumption contract intact.
Verification: root and desktop Go tests, go vet, make lint/check, frontend
typecheck, 498 related tests and production build. Browser checks cover PDF
navigation, CSV literals, menus, request consumption and responsive themes.
The builtin schema inventory grows by 143 measured tokens; boot effect tests
verify stable schemas within a turn for direct and capability-proxy calls.
The FTS projection indexed raw user message Content while every display path
stripped the host's prepended blocks (<response-language>, <current-date>,
<project-instructions>, <available-skills>, <workspace>): a query for a
block-only phrase matched a row whose snippet could not contain it.
Move the strip into sessionstore.UserSaidText, the one definition of what a
user turn says, and call it from both sides: the indexer
(historycatalog/documents.go) and the snippet builders (history/search.go,
history/indexed.go). The history package's private stripComposePrefixes is
gone, so the two cannot drift again.
Scope is today's snippet behaviour: the leading blocks stripComposePrefixes
took, plus the plan-mode marker. Trailing <execution-policy>/<memory-recall>
stripping is deliberately not added; whether "the user said" also excludes
host-appended blocks is a product decision left for the PR body.
Bump historycatalog.TokenizerVersion to 3 because the indexed vocabulary
changes. ensureTokenizerVersion wipes history_fts/documents/sources/roots when
the stored stamp differs, and the manager re-registers its roots, which
re-indexes them; Rebuild stamps the version it built with. Covered by
TestTokenizerVersionMismatchClearsMixedProjection and
TestRebuildSurvivesTheVersionCheck.
Tests: TestFTSIndexesWhatTheUserSaidNotTheHostBlocks (red before the change),
TestIndexAndSnippetAgreeOnTheUserText, TestIndexKeepsPlainUserTextUnchanged,
TestSnippetBuildersReadTheIndexedUserText.
Problem: selecting a custom preset captures unused skill templates and
extension packages and waits for optional MCP startup.
Root cause: resource snapshot creation also served as the runtime permission
model, making selection cost scale with installed resource payload.
Fix: add live capability-grant bindings while preserving legacy snapshots.
Resolve current skill/MCP/extension/hook/command authority at use, defer
optional startup, retain stable provider contracts, and fence new event
and archive formats from older readers. Use lightweight status summaries
and per-instance extension call leases for safe concurrent replacement.
Verification: root and Desktop suites, full boot/builtin suites, affected
storage/controller/frontend tests, lifecycle race checks, frontend interaction
tests and typecheck, lint/vet/check, actionlint, and Windows cross-compilation.
Twenty-sample real switch measurements include build, persist and publication;
64 MiB unused templates remain outside the binding and stay below the local
500 ms P95 target. Native Windows execution is delegated to hosted CI.
#13954 let /resume delete a saved session with a confirming second
Ctrl+D. The merge of #13755 (d87d0d959) reverted it: its branch's
"Merge studio into a5/schedule-api" (ba48dfe0b) resolved files only
studio had changed to the branch's older copies, taking out the picker
code, its tests, the i18n catalogue and the release note.
Restore #13954's eleven files to their content on the studio side of
that merge; none has changed since.
#13956 gave user ! commands the login shell's PATH through
shellrun.CommandEnv, shared with the bash tool. The merge of #13792
(ac1a737ee) reverted it: its branch's "Merge studio into
feat/13111-origin-ask" (ed1dfebb9) resolved files only studio had changed
to the branch's older copies, deleting shellrun/loginenv.go, the Env
line in RunShellWith, the regression test and the release note, and
moving the probe back into bash.go. A ! command again ran with only the
host's PATH, so tools from nvm or Homebrew were not found when Studio
was started from a launcher.
Restore the nine files to their content on the studio side of that
merge; none has changed since.
Studio now requires every config key to be declared once with who may change it
and when it takes effect; the two keys this branch adds were not in that table,
so loading a config panicked with `setting.undeclared: network.dns`.
Both are `Guarded` / `OnOpen`, beside `network.proxy_url`: which resolver a
request goes through is a network boundary, so the model must not change it by
itself, and a new value applies on the next start rather than mid-session.
Also rebased onto studio, which the same push asked for.
Problem:
The native Windows link gate fails when capturing a junction source with
an internal junction, despite the ordinary symlink cases passing.
Root cause:
filepath.EvalSymlinks can retain a Windows junction alias and fail when
walking its child. It is not the repository's physical-identity resolver.
Fix:
Use fileutil.ResolveExistingPath for source roots, contained entries, and
single-file hook parents. Preserve containment checks, rooted content reads,
strict retained-directory verification, and filesystem error identities.
Verification:
Local source-capture and composition regressions and Windows amd64 test
compilation pass. Native verification remains owned by the Windows
composition-links gate, including the previously failing junction case.
The Windows runner's Git Bash has no shasum, and the grader's
`shasum -c ... || exit 1` turned "command not found" into the same exit 1 as
a changed file, so the reference solution failed there with nothing printed.
The SHA-256 manifest is now checked with hashlib inside the grader's Python,
over raw bytes, and a mismatch names the files that differ.
gradeSeed now carries the grader's output in its error: the CI log said only
"exit status 1", which cannot tell a missing tool from a failed check.
Verified locally: on a PATH without shasum the reference solution grades
clean; a CRLF Makefile and an edited Go file each fail and are named; the
pristine seed still fails. snapshot.sh --check, the corpus tests and
make check pass.
Problem:
The Windows smoke suite did not include agentpreset's snapshot tests and
symlink fixtures could be skipped without proving the reported platform.
Root cause:
The platform package list omitted agentpreset, and no fixture exercised
directory junctions. One portable assertion assumed Unix executable bits.
Fix:
Add native junction capture/restore, hook bundle, escape, cycle, and retained
snapshot rejection cases. Run nine focused source-link tests on Windows and
fail the gate for skipped or missing cases. Compare retained permissions to
the actual source filesystem mode.
Verification:
agentpreset and testenv suites, Windows amd64 test compilation, go vet,
make lint, make check, and actionlint pass locally. Native Windows execution
is explicitly pending the hosted composition-links step.
Problem:
Saved personas can fail when selected with inherited skills, extensions, or
hooks. MCP startup failures lack structured server identity and guidance.
Root cause:
Ordinary discovery accepts linked sources but composition snapshot capture
rejects them. MCP startup errors carry the server only in prose instead of
the module metadata consumed by the localized refusal boundary.
Fix:
Flatten contained source links through rooted reads, retain strict snapshot
verification, and preserve module identities and sanitized failure causes
through HTTP and Chinese/English rendering. Keep required MCP failures
blocking and preserve the original session on a failed selection.
Verification:
Full root Go tests, go vet, make lint, make check, frontend typecheck, and
198 relevant frontend tests pass. Browser fixtures cover failure, retry,
management access, and light/dark wide/phone error presentation. Regressions
cover source deletion and restart, prefix stability, revocation, containment,
snapshot tampering, and credential redaction.
The e2e fixtures finish in 5-15 requests, while a real task in this tree took
deepseek-flash 38. A change aimed at cutting round trips cannot show up on
the short suite, so this adds four long tasks whose workdir is a git archive
of commit 42636ddcb (pricing, ablation, agentgraph and five stdlib-only
leaves) plus a trimmed REASONIX.md, a Makefile, a module README and a small
benchctl consumer:
- rt-cache-write-rate: a CacheWrite rate threaded through cost, fingerprint,
card identity and the catalog, plus an Anthropic row that uses it
- rt-stale-skills: Markdown-only repair of three project skills with no
recognised verifier, the shape behind #13293's lingering stale_verification
- rt-rate-lookup-signatures: five lookups moved to (RateCard, bool),
CatalogKey and ModelRef across about 20 call sites, most in _test.go
- rt-plan-ablation-module: a five-step plan-first task signed off per step
snapshot.sh rebuilds every workdir from the pin and --check detects drift.
One pinned test compares CheckedOn with the wall clock and would fail every
workdir from 2026-11-20; base/ anchors it at the snapshot day instead.
Each task ships a solution/; the corpus is registered with the authoring
guards, which check that every seed fails its grader and every solution
passes. A deepseek-flash baseline (2 trials) solved 8/8 at 13-31 requests
once two grader false negatives in rt-stale-skills were fixed; the README
records it, along with a multi_edit prompt arm and a complete_step
description arm that did not move the request count.
The guard in perf/branch-alignment.mjs was not listed in the Studio workflow, so nothing ran it; the fix itself is user-visible and carries a fragment.
The unavailable refusal carried the module and detail only inside a nested
diagnostics list, so the window filled a generic sentence from a code alone.
Flat module/detail params now cross the boundary and both languages use them.
StartForSession called m.wg.Add(1) without a lock while Close could already
be inside m.wg.Wait() after cancelling the root: Add racing Wait at a zero
count is the misuse sync.WaitGroup documents. Under
`go test -race -run TestStartRacingClose -count=40 -cpu 1,2,8` on studio it
reports a data race and panics with "WaitGroup is reused before previous Wait
has returned".
The launch moves into launchJob, which decides under m.mu, the lock both
Close paths now cancel under, whether the job joins the wait group or is
refused. A refused job never runs: it ends Interrupted, its metadata is
written, and it is recorded and published like any other completion, so a
task recorder that saw it start sees it end. The goroutine body moves
unchanged.
Split out of #13947 at the reviewer's request so the jobs manager is
reviewed on its own.
Dirs resolves the path on the far machine and lists it; when the listing
fails but a stat says the path is a directory, the answer is the directory with
Unlisted set instead of an error. The picker then names why the rows are
missing and 'use this folder' works on the confirmed path. A path that is not
there, or not a directory, stays an error. Listing a directory this account
cannot read is fs.ErrPermission on the script layer, as it is on SFTP.
Probe took SFTP before any other step, so a server that refuses the subsystem
ended the test right after 'connection OK'. With the script layer behind
Client.Files it now completes, and the report carries sftp or exec.
pkg/sftp reports the refusal as an unnamed error string, so the subsystem is
now requested in sftpfs.New and a server reply of 'failure' becomes
ErrSubsystemRefused. Client.Volume prefers SFTP and, only on that sentinel,
serves the connection from execfs over the exec channel and remembers the
answer until the connection is replaced. Any other failure (dead link, refused
channel) is returned unchanged. The exec runner quotes the script and every
argument whole into one /bin/sh command line.
The WSL file layer was a fixed set of /bin/sh scripts taking every path as a
positional argument. It is now a package of its own that takes a runner, so
another transport can reuse it. Scripts name a missing program by exit status,
refuse a file over the transfer limit by type, and check the byte count before
publishing a write. List, Download and UploadAtomic are added, and the shared
scenario now covers names with quotes, spaces, a leading dash, newlines, shell
metacharacters and a path over 3000 bytes.
`flexString` existed byte for byte in the openai, gemini and responses packages.
One type written three times diverges at the next edge case, so it moves to
`internal/base/flexstr`, which is a leaf (it imports nothing under reasonix) and
is declared in `tools/repolint/layers.go` so the layering gate allows every model
package to use it.
No wire behaviour changes: a string code, a number, an absent field and null read
exactly as before.
Problem: The CUA readiness change expands #13720 beyond its reviewed
launcher-ownership scope.
Root cause: The native CUA startup race was repaired alongside the
launcher conflict resolution, despite having a separate runtime owner.
Fix: Revert cad3c58f8a on this branch.
The CUA fix and regressions are preserved on fix/cua-socket-readiness,
based directly on studio. Keep the original launcher fix and conflict
resolution unchanged.
Verification: The resulting tree exactly matches e67fe2ce96. Electron
tests pass (189 passed, one existing platform skip), as do go vet,
make lint, make check, and the builtin/boot package tests.
Document the user-visible startup and restart fix from #13985 in the unreleased Studio notes, in Chinese and English. Verified with make check and the Studio release-notes renderer tests.
Problem: Embedded CUA startup can fail while a valid driver is still
initializing its socket permissions.
Root cause: The driver binds its socket before chmod sets 0600, while the
host treats the intermediate permissions as a terminal startup error.
Fix: Wait within the existing readiness deadline, preserving socket-type
and identity checks and forbidding metadata requests before private
permissions are observed. Cover initial startup and abort replacement
through the shared launch loop.
Verification: Deterministic permission-ordering and deadline regressions,
file/symlink rejection, 188 passing Electron tests with one platform skip,
native macOS CUA acceptance, go vet, make lint, make check, and builtin/boot
package tests. Split from #13720 at the maintainer's request.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
Problem: Native macOS CUA acceptance can reject a valid driver startup with
"Invalid private CUA socket".
Root cause: The pinned driver binds its socket before chmod sets 0600.
The host treats the intermediate mode as a terminal error. The previous
fixture creates its socket with private permissions immediately, hiding
the production ordering.
Fix: Treat non-private socket permissions as pending readiness within the
existing deadline. Keep non-socket rejection and identity validation, and
never request metadata until private permissions have been observed.
The shared launch loop also covers abort-triggered replacement.
Verification: Deterministic regressions force observation before chmod on
initial launch and restart, and cover deadline expiry, child cleanup, files,
and symlinks. Both permission regressions fail before the fix. Electron
tests pass (193 passed, 1 existing skip), as do native macOS CUA acceptance,
go vet, make lint, make check, and the builtin/boot package tests.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
The kernel-side pieces this branch came with are their own PRs now, so this keeps
only `android/`. Rebasing onto studio also drops the mailbox delivery code and the
mailbox credentials that the review asked to remove.
Problem: PR #13720 conflicts with the current studio branch in the Electron
unit test suite.
Root cause: Both branches inserted independent regression tests at the same
location after the host startup failure tests.
Fix: Preserve the launcher ownership, relaunch handover, and closed-console
tests together with studio's credential-cookie restoration test. Retain the
automatically merged shell changes and updated Electron dependencies.
Verification: Electron unit and Linux packaging tests passed (189 passed,
1 existing skip). Formatting, go vet ./..., make lint, make check, and the
builtin and boot package tests passed.
TestEffectTruncatedSummaryOnPressureKeepsCodeAndRetries and
TestEffectModelSwitchKeepsTheCompactionProjection pin what happens turn
by turn around the compaction trigger, but the turn the trigger falls on
depended on how large the first request happened to be. Measured on
studio: about 550 more characters of prompt or tool schema flip the first
to "7 failed summaries" and the second to "switch spent 1 summarization
call", with no change to compaction. So any prompt or tool-description
edit, or a host whose tool surface is a few hundred characters larger,
failed them, and a change could pass on one runner and fail on another.
Each fixture now builds once, measures its first request the way the
prompt estimator does, and pads its system prompt up to a fixed size, so
the trigger and pressure turns are the fixture's. The sizes sit in the
middle of ranges where the outcome does not move: 30000-40000 characters
for the truncated-summary fixture (6 attempts throughout), 40500-48000 for
the model-switch fixture (3 folds before the switch, none after). With
0-5000 characters of extra first-request weight injected ahead of the
measurement, both now give the same result; before, both flipped near
550. The bounds are unchanged.
waitForDone returns on the turn-done event, which the controller emits before it drops its running flag, so ClearSession could answer "cannot clear while a turn is running". Under 8 parallel race-enabled processes at one CPU the unchanged test failed 3 times in 1200 runs; with waitIdle it failed 0 times in 1200.
Windows has a backend now, so Available() is true, but it enforces only after the
user turns on windows_integrity. Tests and the e2e bench that ask whether a default
configuration is confined use ConfinesByDefault (false on Windows), and a caller's
override to enforce (an observed run asks for one) is honoured on Windows only
when the user opted in, since enforcing edits permissions on their folders.
A relocation moves only the entries StateRootEntries names. The grant record sat in
sandbox/ outside it, so a move would have left the record behind while the
permission entries it lists stayed on the folders, with nothing to take them back.
The restricted token now drops every privilege but the traverse check, so backup
semantics cannot write outside the roots even from an elevated session. A SID is
synthetic only if no account resolves it. The launcher refuses to run under a
restricted token (a command's own launcher inherits it), grants only session temp
folders, hands the child just the standard handles, runs it in a job with UI
limits, and the default DACL and the grant mask are narrower. Links are not
followed at grant time and a link root is refused. REASONIX.md, the guide, the
Settings text and the release note list what is and is not bounded.
The host already confines bash's $TMPDIR to a per-session directory, the
file tools resolve a leading $TMPDIR to it, and the host deletes it when
the session ends. The model was never told the last two, so it verified
with `scratch=$(mktemp -d); ...; rm -rf "$scratch"` or wrote tests to /tmp,
and the host refused them (shell.delete_nonliteral, shell.delete_sequence,
write scope): 26 refused calls over 36 deepseek-flash benchmark runs, one
model round each.
A <session-scratch> turn block now says: put throwaway files at
`$TMPDIR/<name>`, spelled that way in bash and in the file tools, and do
not delete them. The host path is deliberately not projected: under bwrap
bash sees the directory at /tmp, and on Windows a backslashed literal does
not survive a bash command line, so $TMPDIR is the one spelling every tool
resolves. The block follows the dateAnchor/projectionDebt pattern: it rides
the turn and never the prefix, settles only when a turn reaches the runner,
is re-owed after a fold, retracts itself once if the directory can no
longer be created, is projected only for a manager the assembly supplied,
and is inherited across a rebuild. Its bytes carry no per-session value,
so two sessions over the same state project identically.
No shell-safety rule changes; the effect test asserts that a recursive
delete through $TMPDIR or a mktemp variable is still refused.
supersedeStandingState judged only the fold region's kept turns, so a
standing block the region held once survived beside newer copies in the
recent tail. A fold whose region kept a single user turn dropped nothing:
TestEffectFoldDropsSupersededStandingState went from 3 <workspace> copies
to 4 across such a fold once a turn block shifted the fold one turn
earlier. partitionFoldWith now takes the tail, and the tail's user-turn
blocks are marked seen before the kept slice is walked newest-first.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Every CompactionDone forgot what the model had been told -- the date,
the skills listing, the project instructions and rules -- so the next
turn sent them again. A truncated or refused digest also emits
CompactionDone, with Messages == 0: nothing was summarised and the turn
that carried those blocks is still verbatim, so re-sending them only
grew the prompt the next attempt had to fold, by one skills listing per
failed fold.
The producer already says which case it is: a successful fold reports
Messages = len(fold) and emitCompactionAborted reports 0. The sink now
reads that instead of treating every completion as a fold.
TestEffectTruncatedSummaryOnPressureKeepsCodeAndRetries goes from 5 to 6
failed summaries over its 10 turns: the prompt no longer grows by a
re-sent listing after each failure, so the pressure fold lands a turn
later. That is inside the test's bound of 6 and is the growth the retry
gate is meant to read.
Inside run_script, call("read_file", ...) returned the model-facing form:
every line prefixed " N→", CRLF folded to LF, the default 2000-line page,
and — past 32 KiB — a window with a "read on from here" marker. A script
that read a file, edited the text and wrote it back put all of that into
the file. With deepseek-flash on migrate-signature-3packages and
agent.code_mode on, one bulk rewrite corrupted 36 files and cost ~15 rounds
to repair.
Two layers each rendered for the model a result whose reader is a program:
- read_file: a call another tool made (tool.IsNested) gets the decoded text
verbatim — line endings and final newline kept, no default limit, offset
and limit selecting whole lines. The model's own read is unchanged.
- nestedInvoker: the calling tool reads what the tool returned, not the
bound fitted to the model's context. The nested card still shows the
bounded form; toolOutcome.keepFull records both originals in one place.
write_file already preserves the existing file's encoding, so a BOM file
round-trips once the read stops adding to it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
model_rounds and round_outcomes treated every call run_script made as a
top-level dispatch, so each inner result reopened a model gap and each inner
dispatch closed one. One recorded code_mode run with 24 executor requests
reported a median of 87.5 rounds. Nested calls carry no ParentID — only an id
derived from the caller's, inside its dispatch→result window — so the digest
had nothing structural to tell them apart.
The nested id is now declared in internal/contract/tool (NestedCallID /
IsNestedCallOf) and the agent's nestedInvoker builds ids through it, so the
producer and the reader share one definition. e2ebench's callNesting tracks
open top-level calls and treats subagent calls (ParentID) and invoker calls
(derived id) alike: neither cuts a round, joins a batch or adds tool time.
roundEnds and splitAtCorrect use the same tracker. NestedCalls reports how
many there were, and ModelRounds' doc now says what it counts.
With -trials N every trial wrote <task>.trajectory.jsonl, so only the last
survived and every trial's digest read it. Trials now write
<task>.trial-N.trajectory.jsonl (legs: <task>.trial-N.segM...); a single
trial keeps the old name. sft mode resolves the task id through
trajectoryTaskID, which strips trial and leg marks; traj, serve and barrier
modes glob *.trajectory.jsonl and label each file by its stem, so trials
read as distinct entries.
Cause: the digest had no identity for tool-made calls; the trajectory path
had no trial in it.
Blast radius: cmd/e2ebench digest and file naming; nestedInvoker's id string
is unchanged (same "<parent>.<n>" format, now built by the contract helper).
Neighbouring behaviours tested: subagent ParentID exclusion, segment paths,
answer-isolation audit across legs, existing round decomposition tests.
Why this layer: the id convention is owned where nested calls are created;
the digest reads it rather than restating the format.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Problem: Windows CI rejects the retained-hook source-deletion fixture as a
credential-bearing launch before reaching its restoration assertions.
Root cause: credential projection scans the complete quoted shell string as a
generic value, splitting a Windows drive colon inside a quoted file argument.
That creates an unmatched quote and falsely marks the command as sensitive.
Fix: reduce shell launches to static arguments before credential inspection,
preserve exec and legacy argv launches, and continue refusing literal token and
header arguments or shell forms that cannot be reduced safely. Keep source
commands and retention/restoration semantics unchanged.
Verification: regression reproduces both quoted Windows path failures before
the fix, while exec/legacy argv and token/header refusal cases stay covered.
Focused retention/security regressions and their race run pass; full builtin
and boot tests, gofmt, vet, lint and changed-file checks pass locally.
Problem: CodeQL reported new path and allocation alerts, and Windows repeatedly
failed the retained-hook restoration test.
Root cause: Checked selectors were still passed onward as raw request values;
some generated paths and capacity hints did not expose their bounds at the IO
owner. The HTTP DTO also carried an internal inherited binding. The hook fixture
fed Windows backslashes through Bash resource parsing.
Fix: Resolve import/export selections to registered values, forward checked
relative names for resume/migration, and guard the generated session slug before
joining it to state storage. Use rooted declaration/resource reads and copies,
bound resource reads, and avoid unchecked capacity addition. Pass inherited
bindings outside the wire DTO. Normalize and quote the hook fixture path, including
spaces, while retaining source-deletion and exact-binding assertions.
Verification: Hosted alert locations reproduced locally with the path-injection
and allocation queries; before/after query results and the base scan are retained
separately. Boundary regressions fail on the original head and pass with the fix.
Owning HTTP, store, config, hook, sidecar, plugin, specification and encoding suites;
full builtin/boot tests; gofmt, vet, lint and changed-file checks. Hosted Windows
and CodeQL results remain required for exact-head acceptance.
Problem: The composer exposed the new mode picker to users who had not opted in,
contrary to the maintainer's product-visibility decision.
Root cause: Removing the persistent visibility preference also removed its
settings/search action and the RunLine mounting gate.
Fix: Restore the original on/off storage key and default-off behavior. Add the
switch and concise explanation to Persona management, restore settings search
and the action, and gate both mounting and catalog reads. Preserve card defaults,
catalog/editor design, persona copy and current conversation bindings. Failed
storage writes retain the previous choice and report an error.
Verification: Missing/disabled/unknown and toggle regressions fail before the
fix. Focused Chinese/English navigation, persistence, storage-failure and binding
tests pass. Full frontend: 729 files / 5,857 tests; build/typecheck, UI census and
ten identical fixture reports; gofmt, vet, lint, make check, builtin and boot
suites. Browser visual acceptance remains unavailable under the URL policy.
Problem: Built-in preset names and descriptions read like configuration modes
rather than collaborators, and their relationship to Harness prompts was unclear.
Root cause: Technical copy was reused across the catalog, composer and persona
editor without explaining the existing short-identity prompt adaptation.
Fix: Rename the built-ins to Everyday partner and Lightweight partner, update
their bilingual descriptions and guidance, and derive the unbound picker label
from its canonical catalog definition. Keep IDs, persona bytes and selections
unchanged, retain historical translations, refresh revision metadata and document
the pinned official Harness reference and native prompt boundary.
Verification: Frontend 729 files / 5,852 tests and production build/typecheck;
contract, store and HTTP suites; gofmt, vet, lint, changed-file check, builtin and
boot suites. Fix a discovered worktree test race by waiting for the asynchronous
checkout state instead of the creation title shared by initial and checkout
phases. Browser visual acceptance remains unavailable under the URL policy.
Problem: Built-in presets use flat rows, choosing a future task mode requires
repetition, and persona fields are hidden behind an advanced disclosure.
Root cause: Management navigation and new-task launch intent have no shared
explicit preference; the editor groups basic persona fields with identifiers.
Fix: Match the selected two-column card hierarchy with badges, IDs, descriptions
and separate details, usage and copy actions. Persist an explicit built-in card
choice in a new machine-local key and inject it through the local/remote open
owners only for new tasks. Explicit request choices override it; restoring
history keeps its binding. Existing server preferences alone remain inert.
Expose both persona textareas, runtime context, scope and ID directly, rename
settings navigation to Persona, and keep literal persona fields authoritative.
Verification: Full frontend suite plus real MockHub launch effects, blank-pane
reuse, local/remote history, explicit overrides, storage compatibility/refusal,
cross-window refresh and independent footer actions. Production build/typecheck,
UI census/fixtures, format, vet, lint, changed-file and builtin/boot checks.
Browser-rendered visual comparison remains blocked by the URL policy, as
recorded in design-qa.md.
Problem: "After the first-party guidance" exposes prompt-construction jargon
without explaining what users should enter.
Root cause: Both persona editors use a position-based label; the composition
editor has no suffix helper text.
Fix: Label both suffix inputs "Additional instructions" and explain that the
text follows the built-in system instructions and may be left blank. Add the
Chinese/English copy and associate the helper with each textarea. Keep the
existing literal suffix binding and update affected interaction expectations.
Verification: 24 focused persona, preset and localization tests; production
build/typecheck; formatting, vet, lint, changed-file checks and builtin/boot
suites.
Problem: The catalog's create icon and label stack inside a cramped button.
Root cause: Studio icons render as blocks while the button lacks a horizontal
layout; the title and description also share a flex item beside the button.
Fix: Give the button an inline-flex row with no wrapping, a 36px minimum height
and 16px horizontal padding. Align title and action in one grid row and place
the description across the next row, retaining the primary theme colors.
Verification: 27 focused frontend tests, production build/typecheck, formatting,
vet, lint, changed-file checks, and builtin/boot suites. The local preview stays
running; browser observation remains blocked by its URL security policy.
Problem: Preset browsing, creation and the Standard persona were mixed in one
long Settings page, hiding the save flow and exposing configuration too early.
Root cause: Catalog selection and editable drafts shared one inline surface,
without an owner-level guard for Settings navigation.
Fix: Follow the selected full-content editor reference. Group the catalog by
scope; open a separate editor with common fields first, expandable capabilities
and advanced options, and a persistent Cancel/Save footer. Preserve literal
persona bytes, revision checks, import/export/delete and session bindings.
Guard dirty exits across return, Settings close, category and search paths.
Keep pending saves mounted and reject stale import results. Reusable switches
declare button type so toggles cannot accidentally submit the editor form.
Verification: Full frontend suite: 727 files, 5843 tests; focused editor,
navigation, pending/refused save, import ownership and binding regressions;
production build, typecheck, UI census and fixture verification; root format,
vet, lint, changed-file checks and builtin/boot suites. Browser visual acceptance
remains blocked by the local URL security gate and is recorded in design-qa.md.
Every open tab keeps its pane mounted, and a hidden pane was only
visibility:hidden, so it still took part in layout. The rail width is animated
per frame, so each frame laid out every pane's transcript again. With three
panes the median frame at 4x CPU slowdown went from 94 ms to 18 ms.
Problem: the mode picker required a separate settings toggle before users
could discover or use it.
Root cause: both the picker and its run-status row depended on a persistent
visibility preference unrelated to the conversation's selected mode.
Fix: render the picker directly, remove the retired setting, preference state,
search entry and action, and update the bilingual documentation. Standard
remains the default; displaying the catalog never selects a composition.
Verification: failing-first tests for absent and disabled legacy preferences;
default Standard with a saved Minimal preference; picker, settings, remote-open
and settings-search interaction suites; frontend typecheck, full suite, census,
fixture verification and build; browser preview; go vet, make lint, make check,
builtin tools and boot assembly tests.
Problem: Publishing Studio did not regenerate its website changelog.
Root cause: Product publication and the website deployment had no explicit handoff, and Pages only accepts runs from the website branch.
Fix: Dispatch the website workflow after stable publication with the required version and a unique request ID. Wait for that deployment and propagate failures. Register dispatch on the default branch without enabling deployment from it, and use the workflow commit when recovering old tags.
Verification: All 13 publication tests, shellcheck, signing workflow contracts, go vet, make lint, make check, and the required builtin/boot suites pass.
Problem: the custom-preset form grouped storage scope and ID under one label,
leaving their purpose unclear.
Root cause: two independent controls shared the Preset ID field container
without visible labels or usage guidance for each control.
Fix: give each control its own label and short translated description, and
associate the descriptions with the controls for assistive technology.
Verification: AgentPresets interaction tests, frontend typecheck, production
build, census and fixture verification; browser layout and console checks;
go vet, make lint, make check, builtin tools and boot assembly tests.
Problem: PR #13613 conflicted with the current studio branch.
Root cause: both branches extended session status, ACP configuration, and
project dispatch while the shared stylesheet and session-new function reached
their existing size limits.
Fix: retain composition bindings alongside tariff status, apply the ACP
approval ceiling, and reuse pending dispatch panes without dropping inherited
composition bindings. Move the picker styles to their own sheet and compose
the ACP config projections without increasing the ratchet budgets.
Verification: frontend typecheck, 5,830 frontend tests, production build,
census invariants and fixtures; go vet, make lint, make check; ACP, HTTP serve,
composition contracts/storage, session storage/control, delegation, builtin
tools, and boot assembly tests.
Merged studio 3d9f9ff74c into PR head
ae6a9de150.
The kernel refuses every request without it, and the cookie belongs to the
host rather than the port, so a removal left the window refused until a
relaunch. The shell holds the token and puts it back, logging the cause.
.gitattributes pins *.go to eol=lf but not *.golden, so a Windows checkout turns
the byte-compared goldens into CRLF and TestCompareOutputUnchangedWithoutUnaccountedRuns
fails there. Pin *.golden to eol=lf as well.
The macOS tray loaded the full-colour tray.png, so the menu bar showed the
whole app badge. A Template-suffixed black-on-transparent pair lets macOS tint
it for light and dark bars; Windows and Linux keep their files.
aggregateArm dropped an unaccounted run's wall before adding it to
s.WallMs, so compare's Solved/hour and global wall/task divided by a
smaller wall than the suite report uses for the same -json artifact:
report.go adds every run's wall, and a killed run's wall is measured
unconditionally in main.go (only its tokens never landed). The
per-class accumulator ran before the skip, so one section printed two
populations, and the sign of "which arm is cheaper" could flip.
WallMs is now the all-run wall and AccountedWallMs the accounted
subset: Solved/hour and wall/task divide the all-run total, while
"Wall seconds / solved" keeps dividing accounted wall by accounted
solves because that denominator is solved runs. No new persisted
state, and neither how Unaccounted is decided (segmentrun.go) nor the
suite report changes.
This is the residual of #13446 applied to a field that is NOT
unmeasured. That PR decided which runs are unaccounted and kept their
spend out of the token/cost ratios; it never touched wall, and this
change does not revisit that decision. It fixes the wall population
those adjacent rate lines divide, measured for every run including the
ones #13446 excluded elsewhere. Not a re-report of #13446.
Verified: both probes are red on 53a8db913 (compare Solved/hour 120.0
vs the report's 3.9; global wall/task -30.0s vs the class row's
+870.0s) and green after; a no-unaccounted fixture renders compare
byte-identically (goldens); "Wall seconds / solved" divides accounted
wall by accounted solves under either rate line reverted to the
accounted total, while the matching probe turns red; the suite
report's Solved/hour convention is pinned unchanged. gofmt,
GOOS=windows go vet ./cmd/e2ebench/, go test ./cmd/e2ebench (only the
pre-existing environment TestStateHomeIsOutsideWorkdirParent fails,
identically with this change stashed), make check (repolint: clean),
and golangci-lint run ./cmd/e2ebench/... (0 issues) pass.
Cause: a transcript that could not be read was stamped health=corrupt while its
root was left state=ready with the failing size/mtime signature. Both automatic
repair paths trusted that row: reconcileRoot returned before indexPath ran and a
restart reopened the same ready row. Equal size/mtime says nothing about
readability, so a restored mode or a remounted transcript was never revalidated;
status stayed Indexed 0 / Total 1 until a save or a manual `catalogs reindex
history`.
Fix: chose the scan's skip predicate — reconcileRoot only short-circuits when
state=ready, the signature matches, indexed==total, and no source of that root
is non-ok (health<>'ok' or missing_since<>0). That is the decision that
prevented indexPath from ever being consulted. The same rule is applied at
source granularity: sourceProjectionUnchanged now also requires the stored
health to be 'ok', because after the root is rescanned a legacy source with no
ledger digest (digest=="") would otherwise be certified healthy from its
fingerprints without being read.
Fast path kept: a root whose last scan was complete and healthy still returns
before any transcript is opened, so the five-minute tick remains a directory
walk plus stat, not a re-read of every transcript. A source that stays corrupt
is retried once per scan (one failed read per corrupt source per tick) and
scan_generation advances exactly once per scan, so a permanently unreadable
transcript cannot wedge or loop the worker.
Tests: the audit's TestCorruptSourceIsRetriedOnceTheTranscriptIsReadable (red
before: health stays corrupt after the transcript heals) plus controls for a
healthy unchanged root still skipping without a read (scan_generation, revision
and document ids unchanged; a now-unreadable transcript keeps health=ok) and for
a permanently corrupt source retried on every scan while its healthy sibling is
not re-read. Permission fixtures skip on Windows and root, and every catalog in
these tests points at a temp path.
golangci-lint's intrange rule rejects a C-style loop whose index is never
read. Both loops in the new cases just repeat an emit, so they take the
Go 1.22 form. Behaviour is unchanged.
Store.MigrateV2WithVersion returned on the first fact whose os.ReadFile
failed, so facts later in the same directory and every later directory
were never migrated. MigrationReport carried only { Migrated int } and the
boot caller discarded it, so nothing named the skipped files: the audit
observed {Migrated:1} with zeta and omega never reached.
The contract says otherwise. docs/SESSION_MEMORY_RETRIEVAL.md:259-261: a
read condition is "reported instead of returning partial lists or
calculating a revision from incomplete history". :268-269: "These
conditions ... do not establish data deletion." The pass therefore records
each failed basename in MigrationReport.Unreadable, sorted, and continues
with the rest; boot's warning names them instead of printing only the
first path.
A per-fact read failure returns a nil error on purpose: the field is the
channel for "skipped", and folding it into err would make a skipped fact
indistinguishable from a pass that could not run. Directory and index
failures still return an error, because those leave the pass's own work
incomplete.
Tests cover: the audit fixture (alpha readable, beta chmod 000, zeta after
it, omega in the second directory) now migrating alpha/zeta/omega with
Unreadable=[beta.md], cross-directory sort order, a fully readable control
with no error and an empty list, the repair path for an already-migrated
fact behind an unreadable one, and the boot warning naming beta.md.
Review asked for comments to describe what holds now and leave how the code
got here to the commit history. The behaviour is unchanged.
The history this replaces, kept where it belongs:
* Leaving Source in the zero-probe kept ToolProgress on the passthrough
path, so the first attempt at merging progress merged nothing in
practice. Source is identity — it names the emitting party — so it
belongs in the key, and keeping it there is also what stops a planner's
text merging into an executor's.
* The e2e run over a real kernel is what caught that; the unit tests did
not, because they emitted deltas with no Source at all. Both are now
covered: TestCoalesceMergesDeltasCarryingSource and
TestCoalesceSeparatesBySource pin the key's shape.
Cause: persistCompactionStateLocked reached SaveCompactionState, a blind
atomic overwrite carrying no writer identity, while transcript saves fail
closed with ErrSessionWriteAuthorityMissing. A controller in the
stale-authority window (rebind, controller replace, lease generation change)
therefore installed and published a fold whose transcript save was refused:
it replaced the owner's installed digest and paid for the same span twice.
Enforced in the store rather than in the agent: (*Session).SaveCompactionState
presents exactly the permit save.go presents (requireWriteAuthorityForSave,
including BeginSave's in-flight accounting that lease Release waits for), so
there is one adjudication of the permit and no second check re-reading
WriteAuthority() in the agent. Every write to the owner's sidecar already
funnels through the agent's single persistCompactionStateLocked, which now
calls that method. The raw SaveCompactionState remains for paths that write a
fresh recovery lane and for low-level CAS tests that never bind authority.
The CAS on (generation, projection_version, covered_prefix_hash) is untouched:
this is an additional gate, not a replacement.
On refusal the typed authority error propagates as
errProjectionNotPersisted: commitSummaryProjection restores the previous
projection, so nothing is installed or published, and the next attempt under a
valid permit folds once.
Tests: internal/runtime/agent/compaction_sidecar_authority_test.go proves the
no-authority refusal (sidecar and digest unchanged), the authorized control
(one summary, one digest), and that a refused publish costs one normal retry,
not a second summary.
reanchorPattern stripped the backslash from \!name and \#name before
handing a root pattern to the matcher, which then read the first as a
negation and the second as a comment. Files named !name or #name that git
ignores through the root .gitignore, .git/info/exclude or the global
excludes file were therefore searched and listed. Nested .gitignore files
were not affected, since their patterns are re-anchored under /dir/.
Keep the escape on root patterns when it guards a leading ! or #.
The global ignore file was taken from the first core.excludesFile found:
$XDG_CONFIG_HOME/git/config before ~/.gitconfig, and the first value in a
file. Git reads the files in that order and keeps the last value, so a
user's ~/.gitconfig setting lost to the XDG one. The value was also used
nearly raw, so an inline comment or escaped backslashes produced a path
that does not exist, and the user's global ignore rules were silently
dropped.
Keep the last value across files and within a file, and decode the value
like git: strip unquoted # and ; comments, remove quotes, and unescape
\\ and \". Other backslashes stay as written so raw Windows paths still
work.
useRevealed paces a streamed answer on the frame clock and keeps its cursor
in a ref. The effect returned early when the cursor had caught up and only
handled a shrink inside step(), which the early return made unreachable. A
replacement shorter than what was already revealed — regenerate, rewind,
history restore — therefore left the cursor above the new length, and every
later growth below it was short-circuited: the typing animation silently
stopped for that message while the text itself still arrived in one piece.
Handling the shrink before the early return fixes it. The cursor comes back
with the text, so the next growth is paced exactly like the first one.
The existing three tests did not cover this: they only grew the text. The
new case fails against the old code with "expected 0 to be greater than 0".
A shell emits one ToolProgress per output line, so a build log crossed the
coalescer as one frame per line. Downstream that filled a subscriber's
64-frame queue and the hub then dropped answer text beside it, because text
and progress are both droppable. Measured on a real 500-line command: 285
frames unrepaired, 2 with this.
streamDelta replaces isStreamDelta and returns the key a burst merges under
plus its payload, so progress joins text, reasoning and compaction progress:
* the key carries Source. Every progress chunk an executor emits sets
Source="executor", and leaving it in the zero-probe kept the event on the
passthrough path — the first attempt at this merged nothing in practice.
Keeping it in the key also stops a planner's text merging into an
executor's.
* the key carries the call id, so two tools reporting at once never land
in one another's card.
* progress carrying more than ID+Output — a sub-agent's status channel sets
Name, ParentID and a rate — fails the probe and still crosses untouched.
* progress without an id is not buffered under an empty key.
Verified against a real kernel over SSE: a 500-line command emits 2 frames
that concatenate to the same 500 lines in order, and a 3000-line command
emits 4. Alternating text and progress is unchanged (the key switch flushes
either way, so that burst still cannot merge).
The kill test proves the leftover is removed and the live fact is untouched, but
nothing pinned the branch itself: forcing an active id down the rename-back path
now fails because ListArchived lists the live fact as archived again.
RestoreArchived moved the archive entry away only after the active file
and its index line were published. A process killed in between left one
id both active and archived, so ListArchived reported a live fact and a
retry was refused as "already active".
The entry now moves to a hidden .restoring-<id>-<stamp> name before the
active file is written. The stamp is the archive time, so the name
carries both the id and the original archive name with no journal and no
content comparison. Every restore sweeps leftovers first: an active id
means the dead restore had published, so only the redundant copy goes;
an absent id means it never published, so the entry is renamed back. A
failed sweep reports its error and publishes nothing.
Resolve .reasonix/reasonix.toml only when the root project config is absent. Share selection across loading, local edits, installs and repair; retain project-scope restrictions and reject symlink escapes.
Fixes#6007
chmod 000 does not remove read permission on Windows, so the fixtures that make
MEMORY.md unreadable would fail their own self-check there. The classification and
aside behaviour are unchanged on Unix; the directory-named case stays cross-platform.
After an unreadable index is moved aside and the derived index republished, the
store must list exactly the fact files and the index must not name the aside.
MigrateV2WithVersion is the only entry that recomputes the derived
MEMORY.md, and it returned readIndexIn's error. Every mutation path reads
that same file first, so a restored store whose index has a foreign owner
or a changed ACL froze every memory write for good: boot only warned, and
the docs' promised re-index has no other entry.
Classify the read error by identity: only fs.ErrPermission (permission
denied, a foreign owner, an ACL) and a typed not-regular-file error (a
directory named MEMORY.md) reach the rebuild. Every other error -- an I/O
failure, a resource limit, a momentary lock -- still returns, because the
file may be perfectly good and republishing would overwrite hand-written
text a retry can still read.
Before publishing, atomically rename the unreadable entry to a stamped
name in the same directory (MEMORY.md.unreadable-<unixnano>), so the bytes
survive for a later repair; a rename needs write access to the directory,
not read access to the file. If the rename fails, return the original
error and publish nothing. flushIndexIn keeps the readable path, so
hand-written text is still preserved line for line when it can be read.
No new persisted state beyond the aside copy; the MigrateV2 signature and
boot call site are unchanged, and an unreadable store directory or fact
file still aborts the pass.
Tests: the audit probe is red before and green after; controls cover the
stale readable index, hand-written survival, byte-identical healthy boot,
missing-index rebuild, and a later save succeeding after the repair. New:
a transient read error leaves the index byte-identical (red against the
previous revision), the unreadable bytes survive under the stamped name, a
directory index repairs, and a failed aside rename publishes nothing. Each
new guard was reverted alone to confirm its control goes red.
The turn-start workspace snapshot was awaited with no bound, so in a very large
workspace (a drive root: 7 s on a warm SSD for the first turn, 0.5 s after) the
first tool could not start until the walk finished, and a command's own
timeout_seconds, which only starts when the command does, never covered it.
The wait now ends at snapshotWait, the bound the end-of-turn snapshot already
has: the walk is cancelled, the snapshot records itself incomplete, and the
bundle says the turn's changes were not compared.
The row sized the name, the model chip and the status to their content and gave
the task only what was left, so with the chip present the task was squeezed to
8 px at a 360 px panel and 0 px at 280 px. Below 460 px of the list's own width
the row now takes up to three lines: name and status first, the task second,
the chip third.
profileForSkill set only the name and count, so a skill sub-agent (explore,
locate, run_skill, read_only_skill) showed no task while a task sub-agent
showed its first line. The label helper moves to the event package, next to
the Profile it fills, and both producers use it.
A task text that names a file in the host's own .reasonix/attachments directory
now gives the delegate that image (judged by the resolved path, read through the
same checks an @-reference gets), on the vision role with the note that it rides
the message, including on a follow-up turn that carried no image. The
read_only_skill runner took the turn's candidates without routing the model or
saying so, so a text-only skill sub-agent dropped them silently.
Each delta did s += piece on the row's text, so a long stream cost time and
memory quadratic in its length: 20000 progress events of one command (4.5 MB)
took about 13 s of the UI's own time. The text now grows in a buffer kept beside
the row, trusted only while the row still holds exactly what it last handed out.
The product keeps its 15 second bound on one ssh -G; the source now carries it
as a field with that default, and the test that runs the installed ssh sets it
to testenv.Budget, a bound that only reports a hang.
internal/session/control/git_push.go measured 0 from the control package's own
tests, so declaring it sensitive recorded a floor of 0 that could never fail.
The tests cover the trust refusal for both calls, a repository that is not
there, a branch with no upstream, and the commit lock released after a refused
push; the floor is now 100.
Two project folders under one git repository share a writer domain on purpose
(the lock and the visibility of each other's claims), but an undeclared writer
claimed the domain root, so sibling projects waited on each other for bash and
MCP writes, and the refusal read held X, requested X. It now claims the
selected folder; a session on the repository root still overlaps both, and a
declared path that cannot be resolved still claims the whole repository.
Problem: trusting a publisher whose plugin source has a mistyped commit
reports a misleading local-or-unsafe-source installation failure.
Root cause: plugin source validation collapsed its refusal into the generic
bad-source class before the HTTP and presentation layers could explain it.
Fix: preserve a distinct invalid-plugin-source error through publication and
installation, translate its complete-commit requirement in both languages,
and align the shared client validator with existing registry path checks.
Verification: Go vet, repository lint, owning market/serve tests, the required
builtin/boot tests, 391 frontend market/i18n tests, and TypeScript checking.
The push is the one git action Reasonix performs that leaves the machine, and
the host-only route list decides which requests a paired device may make. Both
now demand review plus security_review and are held to the coverage they have:
gitpush 84.0, serve/git_push.go 77.0, serve/host_routes.go 93.0, and
control/git_push.go 0 (its behaviour is covered by the serve and boot effect
tests, which the per-package gate does not count).
An Auto Guard card and a call that asks before it writes now give the claim
back for the wait; the lease records granted claims so the resume can name
who took the ground, the model gets workspace.claim_contended and the receipt
workspace_claim_yielded. A blocked acquisition ends at a user-set maximum wait
with workspace.claim_timeout.
closeBootTestHistoryCatalog gave the history and usage catalog closes a
fixed 5 s. Closing a SQLite catalog, and waiting out an open still in
flight, can take longer on a loaded Windows runner, and the posture
effect test failed there in cleanup with "close usage catalogs: context
deadline exceeded". The wait only exists to report a hang, so use
testenv.Budget; a close that never returns still fails the same way.
A child that shares the parent's provider starts on the depth in force now, and one
that inherits the parent's model without an effort of its own is built for it, as a
rebuild would have given. The status line reads the depth the session is on.
A provider that exposes the depths it accepts in Request.EffortOverride
(provider.RequestEffortProvider; the OpenAI-shaped adapter does) lets a session change
depth without a rebuild: the controller sets a session depth every main request
carries, updates its running selection, and accepts the change mid-turn. Serve (and so
the terminal and the desktop) and ACP both apply it in place and keep their rebuild for
a depth the provider does not take per request. The level is still saved to the user's
config so a restart opens on it.
Problem: the advancing Studio branch conflicted with the mod-turn settlement
test and changed queue pause and child-dispatch reporting contracts beside the
reviewed composition work.
Root cause: both branches added turn-completion waiting, the combined inbox
owner exceeded the file ceiling, and the effort audit still expected concrete
runtime identity in dispatch profiles after Studio restored selection profiles.
Fix: integrate exact Studio da9fbe7307 normally.
Keep both queued TurnDone and durable idle settlement. Group pause operations
in inbox_pause.go without behavior changes. Assert requested selections on
dispatch events and concrete identities on status events, retaining real
provider effort, rejection and persisted-child checks. Preserve landed queue,
billing, phone and remote protection work and the default-off composition UI.
Verification: format/vet/lint/make check; all 229 root Go packages and required
builtin/boot; affected HTTP/storage/controller and focused boot races; pause
controller regression; Desktop/SDK; 699 frontend files and 5670 tests,
typecheck/build/census; 12 browser/eight Electron renderer scenes; Electron
units (177 pass, one existing skip); 13 official ACP SDK conformance tests.
No threshold or authority expansion; no hosted check result is substituted.
The vendor's published schedule already prices every request; the status bar
now says, live, whether the model is billed at the peak or the off-peak rate
and when that changes, where the model is billed at the vendor's own rate.
Problem: repeated New session reset a blank pane's mode, copied built-in
descriptions ignored the window language, and identifier fields missed theme
styles. The picker was always visible and native archives were parsed in the
renderer. Unrelated transitions also repositioned every open popup.
Root cause: pane reuse resolved an unrequested default; editor styles depended
on a container the identifier row did not share; interface opt-in was missing;
archive transport carried parsed objects; placement had no anchor scope.
Fix: change a reused pane only for an explicit selection and preserve focus on
refusal, translate built-in copies, share field tokens, add a default-off local
picker switch that preserves bindings and the original hidden-state layout,
upload opaque files to the existing host decoder with typed localized errors,
keep the prior JSON envelope, and scope transitions to each popup's anchor or
container. Show catalog unavailability and wrap the editor components.
Verification: failing-first regressions; format/vet/lint/make check; all 228
root Go packages, required builtin/boot, HTTP/storage races and exact-binding
archive transport tests; Desktop/SDK; 691 frontend files and 5608 tests,
typecheck/build/census; 12 browser and eight Electron renderer scenes across
languages/themes/widths; Electron units (176 pass, one existing skip).
No authority, schema, safety policy or verification threshold is widened.
A balance_url on a host that is not a vendor's own now names its format
(balance_format) or is not read, and the key is not sent to it. The connection
editor edits the address and format. Existing relay entries are declared as
deepseek once, and a format the 1.x line drops from the shared file is put back
at the next start.
Problem: ungranted HTTP clients could mutate presets or switch compositions,
and empty hook selections could remove configured safety guards. A paired
remote pane could also borrow the host credential, and a selected native
extension could override the host's denied read.
Root cause: preset routes bypassed request-scoped edit grants, proxying erased
device authority, and frozen automation was treated as the whole safety policy.
Fix: gate preset mutations before store/admission access, refuse paired preset
writes before remote credential forwarding, overlay current authorized guards
outside retained automation, update their shared runner on live edits, resolve
hook inspection/save/approval/reload against one home, and mark bound runtime
packages tighten-only. Preserve ordinary unbound extension behavior, immutable
bindings, optional retained automation and the accepted Standard/Minimal UI.
Verification: failing-first HTTP, proxy and real boot/provider regressions;
format, vet, lint, make check, full root and required builtin/boot suites;
affected package and focused boot races; Desktop/SDK; frontend 689 files/5598
tests, typecheck/build and census; browser eight/Electron renderer four scenes;
Electron units. Correct an inherited mod fixture's premature teardown using
the existing durable-idle notification, with 20 independent process runs.
No test threshold, approval gate or product Close behavior is weakened.
The queued prompt starts a turn of its own, still writing under the test's temp
directory when the test returned, so the directory's cleanup failed with 'directory not
empty'. Seen as 30 failures in 40 runs with eight test processes at once, 0 in 40 after.
The test gave a cold Git Bash two seconds to write the heredoc's file; on a shared
Windows runner the start can take longer, and the failure (4.25 s) was that wait plus
the cancel. The bound is now 15 seconds, only to keep a hang from lasting forever.
The test compared the host's measured thought to the provider's 60 ms sleep minus 10,
which a loaded -race runner breaks (42 ms against 50). The size of the measurement is
the thought clock's own unit test with an injected clock; this test keeps its claim that
a measurement is taken, kept with the turn and framed identically.
A summary request the provider rejects as larger than its context (the documented
context_length_exceeded code, nothing else) is retried at 70%, 50% and 35% of the
refused input, shortened by the same deterministic steps as before; a refused request
costs nothing, so the digest is not given up until three smaller inputs were refused.
A provider that refuses every size ends as summary_input_too_large.
The inbox manifest now records why a queue is paused, as a typed code beside
the old boolean. A message approved on its card is claimed over any pause; one
let in by a standing release or dispatch switch only over a pause the host set
to guard stale input, and stays queued over one the person set. The pause and
older queued items are left as they were.
#13812 made the dispatch-time profile the concrete runtime. The effort-switch audit
test (merged beside it) holds that profile to what the call selected (empty model and
effort when it inherits), so studio went red. The selection is audited as the model
asked it; what the child ran on is on its status events, which the window already
reads, so the dispatch profile goes back to the selection.
Problem: Studio's shared remote-open lifecycle conflicts with the composition
callback, and the HTTP adapter omits the selected composition from its body.
Root cause: composition selection was carried by an App-local path rather than
the shared open-ordering owner and the transport serialization boundary.
Fix: merge exact Studio 5b1c282fed, extend
useOpenRemotePane and SseHub with the selection, preserve newest-open ordering,
duplicate-click coalescing and refusal cleanup, and keep editor styles with
their existing catalog stylesheet without widening the shared CSS budget.
Align the child-effort effect assertion with Studio's concrete dispatch model
and effort reporting while retaining provider, refusal and transcript checks.
Keep Standard/Minimal cards, the mode-only left trigger and right run state,
default-off semantics and all existing source/workspace authorization gates.
Verification: failing-first remote Hook/HTTP body regressions; format, vet,
lint, make check, full root and required builtin/boot tests; affected package
and focused boot races; Desktop/SDK Go; frontend typecheck, 689 files/5598 tests
and build; census fixtures; browser eight and Electron renderer four scenes;
Electron unit suite (176 pass, one existing skip). Local review/security review
does not replace maintainer approval or hosted CI.
[tools].disabled removes a tool from every request and refuses a call to it with
the reason; tools another feature requires declare so and stay on with a note;
a project file can only add names. Settings lists every tool with a switch.
The panel row, the full-record header and the nested card of the spawning call show
provider/model · effort, with the full reference as the title. The card takes it
from the child's status events, so a failed or interrupted run keeps it; a batch's
items each carry their own.
The dispatch profile now names the concrete runtime a call resolves to (an alias
expanded, the parent's model when the call inherits), and each child's status events,
from queued to the terminal one, carry what its run actually started with, per item
for a batch. A failed or interrupted child therefore still says what it ran on.
Bumps the go group in /desktop with 1 update: [golang.org/x/mod](https://github.com/golang/mod).
Updates `golang.org/x/mod` from 0.41.0 to 0.42.0
- [Commits](https://github.com/golang/mod/compare/v0.41.0...v0.42.0)
---
updated-dependencies:
- dependency-name: golang.org/x/mod
dependency-version: 0.42.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go
...
Signed-off-by: dependabot[bot] <[email protected]>
Two controls the audit's contract implies but nothing pinned: a staged copy whose
retired twin came from a different publish is left untouched (no pair, no writes),
and a fact created at the old name after the kill is never overwritten — the retry
refuses and keeps both artifacts for a later attempt. The refusal only named the
conflict once the exist check unwrapped the error, so it now uses
errors.Is(err, fs.ErrExist).
The approval carries the reason code origin_text, so Studio words it in the reader's
language, hides the full-access button (which cannot answer it), and shows the call
whole; the terminal wraps the subject and reason instead of clipping them; ACP
gets a fresh allow-once or reject. The label names up to three origins, then counts
the rest. The guide says headless YOLO refuses such a call.
PermissionLists carries askOnExternalOrigin; a save without it leaves it as it is,
the rules endpoint's rebuild applies it, and it keeps the endpoint's provider-edit
grant. The desktop type and mock mirror it.
A fold that loses the transcript race records FailContextChanged, but no later
turn could release it: retriesOnGrowth did not name the code, the receipt holds
until the generation advances, and the generation only advances on a commit the
held receipt prevents. One racing append then latched automatic compaction off
for the rest of the session, across restarts, and the next overflow failed the
turn instead of folding.
Name the code in retriesOnGrowth so the append that made the fold stale is the
growth that releases the receipt. The receipt already carries InputHash and
InputTokens, and every other code keeps its hold. A plan/provider refusal still
holds, and the retry buys exactly one fold: without further growth the next
prepare neither folds again nor re-records the block.
The recovery may either finish the rename forward or roll it back. Only rollback
keeps an ExpectedRevision from before the rename valid, so the choice is a
contract, not an implementation detail; forcing the forward path makes this test
fail with name=beta revision=2.
publishRename retires the old fact file, flushes the index, then publishes
the staged new file. A kill in that window leaves the fact's only copies as
hidden .retired-* and .staged-* files: Read/List/Revisions report nothing,
the save retry by id reports "memory id not found", and a retry by name
mints a second identity for the same name.
A save by id that finds no active file now looks for the hidden pair a
single publishRename call writes, matched by the nanosecond stamp in the
two file names, with the staged copy's frontmatter carrying that id. The
rule is disk-driven and rollback-first: readable retired bytes are restored
under the old name (same bytes and mode) and reindexed, so the caller's
expected revision still names the fact and the retried save redoes the
rename itself. Only when the retired copy cannot be parsed does recovery
publish the staged copy forward under its new name. With no matching pair
nothing is written and the id is still reported not found, so a retired
copy left by a committed rename is not revived.
Covered by the audit's kill probe, an uninterrupted-rename control, a name
retry after recovery, a failed (not killed) publish, an unpaired-artifact
guard, and a cross-scope move whose pair spans two directories.
With permissions.ask_on_external_origin on (user file only, off by default), a call
that is not a read and whose arguments repeat, byte for byte, text the task's
external results gave it is put to a person as a fresh approval when only the
writer fallback (Auto, YOLO) would have let it run. Explicit allow rules, session
allows, reads and authorized MCP servers are untouched; with nobody to ask the call
is refused naming the origin.
Review follow-up after rebasing onto studio: messages_en/zh/zh_tw.go sit at the
800-line ceiling, so the two new strings move into the embedded QueueMessages and
NotifyMessages catalogues, which have their own files. No call site changes.
session.read and session.submit are separate capabilities, so a mod that only
reads usage cannot put words in front of the model. One prompt a mod queues is
bounded by [mods] submit_bytes (8 KiB, 0 for none) and refused with a typed
identity that does not spend the per-turn count. The docs say the turn a queued
prompt starts is held to this session's own approvals and sandbox.
Problem: Studio browser guards failed on 17f816d11 when the ordering guard
counted SSH rows as conversations in its local project. The remaining quick
action fixture also expected two trailing Archive/Delete buttons.
Root cause: Browser acceptance scripts still encoded the former separate
machine lists, trailing Delete action and individual folding chevrons. The
English wording scan also lacked a newly visible demo conversation title.
Fix: Scope ordering to the named local project and independently assert SSH
rows and keyboard focus are preserved. Exercise default-off and explicit
quick-action choices for local and remote rows, including hit targets,
leading Pin/trailing Archive geometry and menu access. Load the real row
stylesheet in the compact fixture and match its existing 1px spacing while
retaining 34px rows, ellipsis and contained controls. Update manual tree
checks to use project-row folding and five-item previews.
Recognize the existing demo title as user content in the wording fixture list.
Verification: Reproduced the ordering, quick-action and fixture failures
before the repair. Ran all 55 CI browser-guard commands against the built
perf harness and reran repaired guards, the affected manual tree scripts,
gofmt, root vet, make lint,
make check and the required builtin/boot suites. This changes acceptance
scripts only; production UI, workflows and budgets are unchanged.
The reviewer's blocking item: rebindInbox (called from New) started a goroutine
that read c.sink while New was writing it. Capture the sink before starting the
goroutine, for the new unrecovered notice and for the pre-existing recovered one
(it had the same shape and the same race).
Also give the refusal an identity: ErrUnrecovered, wrapped by commitManifestLocked,
with a case in the serve switch so the frontend gets inbox.unrecovered instead of an
internal fault. repolint's refusal-path rule requires that case.
The store already told a caller that the queue's bodies could not be read
(InboxSnapshot.Unrecovered), but nothing user-facing consumed it: the TUI's
/queue line and the notice stream both dropped the field, so an unreadable
queue still read as "inbox empty" and the user could not tell it from a
genuinely empty one.
The TUI mirror and renderQueueList now carry and say the unknown state, and
the controller emits a distinct inbox_unrecovered warn notice (with an
InboxUnrecovered payload naming the queue directory) beside the existing
inbox_recovered one, so its sentence is not confused with "recovered N".
Cause: the frontend contract field had no consumer on the TUI or notice
surfaces.
Blast radius: TUI /queue line, TUI and desktop notice wording, and the two
controller inbox open paths; additive field, code and payload, no on-disk
schema change, recovered behaviour unchanged.
Neighbouring behaviours tested: empty-but-readable queue, recovered salvage
count, TUI mirror JSON decoding of the real GET /inbox body, and both the
serve and control boundaries.
Why this layer: the store owns "unknown vs empty"; the surfaces that render
a queue are where a user can act on it.
Provider-visible prefix: none
salvageOrphanBlobsLocked returned nil for both "no blobs" and "blobs
unreadable", so a corrupt manifest plus an unreadable blobs directory
committed an empty paused manifest (recovered=true, recoveredN=0,
items=null) over the quarantined one. The queued bodies stayed on disk but
were unreachable, the RecoveredN>0 notice never fired, and the queue looked
empty.
The salvage now reports the read failure, and that failure is not turned
into a manifest: the corrupt bytes stay in place as evidence, the store
comes up paused with unrecovered=true on the snapshot, and no commit can
overwrite the evidence while the bodies are unreadable. An absent or empty
blobs directory is still an empty queue, not a failure.
Cause: (items, err) collapsed to (nil, nil) for directory-level and
per-entry read errors.
Blast radius: sessioninbox Open/Snapshot plus the commit path; new
additive InboxSnapshot.Unrecovered field, no on-disk schema change.
Neighbouring behaviours tested: readable-dir salvage, absent/empty blobs
dir, healthy manifest untouched, symlink blob refusal.
Why this layer: the store owns the manifest/blob transaction, so the
"unknown vs empty" distinction belongs where the read fails.
Found while demonstrating the install through Discover: until the switch and
both components read Running there is no reasonix tool, and the model falls
back to running reasonix -p in its own shell, which only looks like a
delegation.
An @token in a peer message or a background-job notice was frozen into the
prompt as a file reference, so a sender could make the receiver read a file
into its prompt, and an unreadable one paused the queue behind it. Items the
host authored are data and are no longer scanned.
Problem: current Studio conflicts in the state-root inventory and adds a move
path that strips the system message containing an immutable composition binding.
That could silently reopen a bound conversation with ordinary capabilities.
Root cause: cross-project adoption intentionally omits project instructions and
permissions, but the new mover did not distinguish immutable composition metadata
from an ordinary system prompt.
Fix: normally merge exact Studio0363720bd. Retain both rules and agent-presets
state roots. Refuse bound conversations in the common MoveSession owner before
copy/delete, preserve the source, and project the typed refusal with bilingual
HTTP wording. Keep ordinary unbound moves, ACP authentication/new-conversation
contracts, directory grants, exact dispatch inheritance and the accepted mode UI.
The preceding source-authority fix remains intact.
Verification: failing-first store and HTTP move tests now refuse blank/started
bound moves and retain the original. All226 root packages pass: mandatory2 plus
the inventoried other224; boot340.765s. Affected race suites pass. The full boot
race invocation previously reached its10minute aggregate limit; final evidence
uses one compiled source snapshot and four disjoint groups covering all753
tests exactly once, without raising the timeout. Static gates, Desktop/SDK,
frontend675files5489tests/typecheck/build/census and browser8/native4 scenes pass.
Local review/security review recorded separately from required maintainer review.
The host keeps the text of a task's external results and counts, per turn, how
many were recorded and how many later non-read-only calls carried a verbatim
string from one. The numbers ride the completion-report audit; nothing is
blocked or shown to the model.
Problem: Review requested WSL boundary evidence and removal of the unrelated
design-qa.md diff after studio introduced non-SSH machines.
Root cause: The new WSL transport supersedes a blanket rejection at shared
remote endpoints. SSH construction must still reject WSL entries, while
generic probe, browse and open must select their configured transport.
The earlier QA cleanup also deleted a document already owned by studio.
Fix: Integrate studio through 0363720bd and cover the real Hub/remotehost/pool
route with disposable WSL and SSH factories. Assert WSL failures survive the
HTTP boundary and never construct an SSH client. Cover Windows, distribution,
empty and fallback default paths. Restore design-qa.md byte-for-byte from
studio so it no longer appears in the PR diff. Retain the upstream SSH form's
refusal to overwrite non-SSH entries.
Verification: Frontend 5529 tests and production build; full serve,
remotehost, remote, attach, wslconn, config, builtin and boot suites; gofmt,
root vet, make lint and make check. No live Windows/WSL qualification claimed.
Problem: PR #13638 became conflicting after studio introduced moving a
conversation to another project.
Root cause: HubPort acquired moveSessionTo beside the archive declaration
that this branch extends with an optional host. Automatic merging also
omitted the errors import needed by the upstream WSL edit guard.
Fix: Keep both capabilities and restore the required import. Retain the
upstream local-only move workflow and the host-qualified pin/archive routes.
Remove an inaccurate rename comment so the combined component stays within
its existing size ceiling without widening the baseline.
Verification: Frontend suite (5529 tests), production build, full serve,
sessionstore, config, builtin and boot tests; gofmt, root vet, make lint and
make check. Transport coverage exercises local project moves beside remote
flags; an interaction regression covers the shared keyboard menu and dialog.
An isolated development browser confirms menu access, project selection and
cancel with hover controls off. Hosted CI remains a separate observation.
The folder in a draft came from the request and went straight to Abs, EvalSymlinks
and Stat. It is now matched, as text, against the remembered projects the person
trusted (scheduler.Folders), and everything after the match uses the list's own
spelling. Anything else is refused as schedule.workspace_unknown, in the kernel,
the CLI and the board's folder list. A run re-checks trust before it starts.
Problem: selecting a repository preset could change an already authorized MCP
server's package arguments or destination and reuse its executable authority or
credentials. The same provenance issue applies to extension args and env.
Root cause: composition parameter validation checked shape and credential
references but did not distinguish repository declarations from personal ones.
Pinning an executable does not authorize the program selected by its arguments.
Fix: enforce source authority in the shared contract. Project MCP/extension
parameters allow only supported timeouts. Validate declarations, imports, old
revisions, and both definition/resolved halves of retained executable bindings.
Keep structural history validation separate so legacy unsafe bindings remain
readable without regaining execution. Align Mock save/import, document the source
table in both languages, and add the release note. Preserve Standard/Minimal,
unbound defaults, persona, additional directories and exact dispatch bindings.
Verification: failing-first boot tests prove changed args start a subprocess and
changed URL contacts another endpoint before the fix; both now refuse before
startup. User positive controls, scope/parameter table, store/revision/old-record
and event-log history tests pass. Format/vet/lint/makecheck, required builtin/boot,
full root, contract/store/sessionstore/serve/ACP race, Desktop/SDK Go and frontend typecheck/build pass;
671 frontend files / 5472 tests pass. Local review/security review recorded;
maintainer approval and new-head hosted CI remain separate requirements.
The full boot race invocation reached its existing aggregate 10-minute limit
as another test started; disjoint test-list groups are used to finish the whole
inventory without raising that timeout or skipping tests.
The index pins the commit that last changed the package; scripts/dsh-market-index.mjs
pins it and validates it with DSH's own validator. docs/DSH.md explains Discover
and the Studio release runbook gains a precondition that keeps the pin current.
A configuration-only package (no code) in integrations/dsh-plugin, installed
from this repository's subdirectory with pnpm's path: selector, plus the
integration guide docs/DSH.md.
How to wire dsh's subagent-acp provider to reasonix acp, what each permission
policy does to Reasonix, what the parent sees, and the limits, verified against
dsh 0.2.1-rc.2.
Problem: newer Studio changes conflict with the ACP session contract and HTTP
runtime opening after the previous assembly integration.
Root cause: Studio adds a new-conversation execution-mode flag where this branch
also owns explicit composition selection and immutable inherited bindings.
Fix: normally merge Studio 4d3c09211. Add NewConversation to the one SessionParams
definition and retain it in HTTP build options. Resolve moved transcript paths
before reading either branch settings or the original composition. Keep exact
dispatch inheritance, workspace refusal, candidate rollback, ACP directories and
the accepted Standard/Minimal composer and management design.
Verification: format, vet, lint, make check, mandatory builtin/boot, full root
Go, Desktop Go, frontend typecheck/build and census pass. Frontend 671 files /
5465 tests and eight browser/four native renderer scenes pass. Affected
boot/ACP/serve/control/remote/delegation race tests pass. Provider request bytes
equal actual Studio with and without persona. Local review/security review are
recorded separately from maintainer review. The unchanged macOS computer
coverage baseline remains 88.2% below 89.8%; no floor is changed.
session/new answered -32603 for a model with no API key or no configured
provider, although initialize advertises a setup auth method. The cause is now
a provider.ErrKeyMissing / ErrNoProvider identity and the frontend maps it to
the ACP schema's -32000.
Problem: Studio advancement made the Agent composition PR conflict in boot
assembly while adding trace export and a workspace-aware mod runtime source.
Root cause: wireMCP now has a composition error result, and extension preflight
filters immutable bindings where Studio introduced its package source.
Fix: normally merge Studio ed1c127d6. Keep OTLP cleanup before returning from
wireMCP and retain workspace-aware discovery before exact composition filtering.
Preserve upstream tighten-only dispatch, replay protection, ACP elicitation and
the withheld mod switch. Keep Standard/Minimal, opt-in binding, ACP directories,
exact dispatch inheritance and the accepted composer layout.
Verification: gofmt, vet, lint, make check, required builtin/boot, full root Go,
affected boot/extension/modhost/ACP/serve/control race suites, Desktop and SDK
tests pass. Frontend typecheck/build, census, 668 files / 5450 tests, Electron
unit tests and eight browser/four native renderer scenes pass. Raw provider
requests equal actual Studio with both persona and empty persona. Local review
and security review found no new issue. Coverage gate still reports only the
unchanged macOS computer baseline, 88.2% versus 89.8%; no floor is lowered.
Rule files under .reasonix/rules, .claude/rules and the user's rules folder are
read from disk each turn. Rules without paths: join the turn as a block owed
like the standing instructions; rules with paths: are owed once the session has
touched a matching file and retracted when they stop applying. Glob parsing and
matching stay with skills' paths: through one exported entry point. Neither layer
reaches the cache-stable prefix; capability diagnostics report what was skipped.
agent.default_mode (user file only, off by default) is applied by the assembly
for builds that open a fresh conversation: Studio panes without a saved
session, the TUI without a resume, and ACP session/new. A resume, a headless
run and sub-agents never take it. Settings gets the switch.
The unverified-change gap names the command the host declined, bounded
at 80 bytes by command[:80]. A check with non-ASCII paths, common in a
Chinese project, could be cut inside a character, and the delivery card
then showed a replacement glyph before the cut mark. Step back to the
start of the character instead, keeping the budget and the mark.
Subprocesses the host starts for the agent carry REASONIX_AGENT_CHILD=1, and the
person-only schedule commands refuse when it is set, so a pty opened by a tool
no longer passes the terminal check. add prints the whole plan and requires its
short id typed back; list shows how each schedule was confirmed. The Windows
preview test compares against the resolved workspace path.
Problem: Agent compositions still carried the persona prerequisite diff and
missed newly added session lifecycle contracts on Studio.
Root cause: the branch predated the landed persona, ACP additional directories
and project-dispatch assembly paths. Combining their independent additions
also exceeded existing file/scalar-state budgets.
Fix: normally merge Studio 896552483 after verifying #13588 actually landed.
Resolve the Persona/composition imports and keep one ACP SessionParams owner.
Clone vetted extra directories through composition switches. Pass the exact
private binding into dispatched runtime assembly; reject cross-workspace
bindings before opening, and reject fresh assembly errors without a history
fallback. Ordinary unbound dispatch remains unchanged. Extract the command
catalog and Mock memory projection, and place the MCP guard with catalog
refresh state. No budget baseline is raised.
Keep Standard/Minimal cards, the left mode-name-only picker, right run status,
default-off/unbound Standard behavior and the landed user-config persona.
Update bilingual behavior/API documentation and release notes.
Verification: format, vet, lint, make check, required builtin/boot tests, full
root count1 atomic-coverage tests, the complete CI race package group, Desktop
and SDK suites; frontend typecheck/build and 667 files / 5443 tests. Electron
unit176 pass with one existing skip. Browser8 and Electron4 controlled rendering
scenes cover locale/theme/width, selection, mode cards, locked/new flows and
left/right layout. Real HTTP requests match landed Studio byte-for-byte with
user persona29139 bytes and empty persona29110 bytes, with fixed probe PATH.
ACP/dispatch regressions fail before propagation and pass afterward. Local
review and security review cover MCP contracts and dispatch/ingress authority.
Known local gate: make coverage-gate reports the untouched macOS computer
package at88.2 below89.8. The exact command on landed Studio reproduces the
same result and uncovered lines. No floor is lowered; hosted Linux gates
remain required. Hosted results for this new head are reported separately.
One draft shape and one set of confirmation facts for every frontend; a
creation is accepted only with the token of the preview the person saw; only
the machine's owner may create or resume, any reader may pause or delete.
The host also answers 'schedule exec' by running the CLI's entry, so the
supervisor can start it as the child. One kernel at a time ticks, under the
machine's ticker lock.
Scheduler.Tick reaps dead runs, claims at most one due slot inside its grace
and hands it to the supervisor; the schedule add, list, pause, resume, renew,
rm and tick commands drive it. Creating, resuming and deleting need a
terminal.
$/cancel_request cancels only the named request and answers -32800;
agent message chunks carry an opaque messageId; an MCP server's form or
page goes through elicitation/create when the client advertised the mode.
Problem: Studio's new SearXNG connection probe conflicts with the persona
types in three transport import lines, blocking this PR again.
Root cause: independent settings features append types to the same line.
Fix: normally merge Studio 6f68a14a9, preserving the search probe and persona
contracts. Keep persona-only imports separate so subsequent upstream type
additions do not repeatedly conflict. Remove one upstream trailing space.
Retain the upstream MCP scope upgrades, ACP session directories, Gemini
provider, session ordering and rail-search behavior without modifying their
implementations. Persona prompt, user-only scope and guarded writes remain
unchanged, as do the approved composition UI and default-off behavior.
Verification: three-way simulations of future settings imports merge cleanly
in all three files. Format, vet, lint, make check, required builtin/boot tests,
full root Go tests with atomic coverage, full CI race package group and Desktop
Go pass. Frontend typecheck/build and 662 files / 5425 tests pass. Chromium
checks cover persona navigation in both locales, light/dark, wide/390px;
rail search, session ordering, web access, settings rows and font scaling.
Local review and security review check the integration and fresh MCP consent;
they do not replace maintainer approval. No gate or budget was relaxed.
Problem: the sidebar PR conflicts with studio's new per-session move actions.
Root cause: both branches extend the row key handler and the mock tree's
session projection. Sorting shortcuts also bubbled into tree navigation,
moving focus to a different session after a reorder.
Fix: retain keyboard context menus and session moves together, compose mock
placement with pin/archive decoration, and consume sorting keys at their
owning row. Keep the existing remote host and body-size guards unchanged.
Verification: 5478 frontend tests across 669 files; production frontend
build; full serve, sessionstore, builtin and boot suites; gofmt, root vet,
make lint and make check. A failing-first keyboard regression and isolated
browser checks prove reordered rows retain focus and keyboard menus still
expose pin, archive and move actions.
A 403 insufficient_scope from a signed-in server now stops the model and
tells the user; granting more is a fresh approval asked every time, from
the server row or reasonix mcp auth --upgrade.
Problem: the latest Studio session-reference settings overlap the persona
port imports and development proxy, leaving this PR in conflict.
Root cause: both features add settings types at the same transport boundary.
Fix: normally merge Studio 8c311b02b, retaining both Persona and SessionRef
contracts, Mock state, SSE methods, routes and settings panels. The upstream
dispatch-consent, bounded-fold and subagent-view implementations stay intact.
Persona prompt assembly, user-only scope, guarded settings and save timing
remain unchanged. The preceding test-only commit synchronizes cancellation
observation without changing the mock provider or relaxing error/close checks.
Verification: formatting, vet, lint, make check, required builtin/boot tests,
full root Go tests with atomic coverage, the full CI race package group and
Desktop Go tests. Frontend typecheck/build and 660 files / 5411 tests pass.
Chromium checks cover persona/settings navigation in Chinese and English,
light/dark and wide/390px layouts; upstream session-reference interactions,
settings rows, font scaling and explorer counts are covered as well.
Problem: the hosted race group failed the mock provider cancellation test
when the next received chunk was an already-pending tool call.
Root cause: cancellation does not prioritize the Done case over a send that
is also ready in select. The test canceled the context and immediately read
again without establishing that the producer had observed cancellation.
Fix: acknowledge cancellation through the test context's Err method before
draining the stream. Retain the strict cancellation-error and subsequent
channel-close assertions, plus the unchanged pre-canceled request test.
The hang guard uses the existing test budget. No mock/product behavior,
runtime prompt, UI, thresholds or workflow approvals change.
Verification: reproduced one failure in 10000 race-instrumented iterations
with GOMAXPROCS=2 before the change; 10000 final iterations pass afterward.
A temporary explicit-marker mutation that ignores cancellation is rejected.
The complete CI race package group passes, alongside gofmt, vet, lint,
make check and builtin/boot tests. Prior frontend/coverage evidence applies
to the unchanged product tree.
The prompt result carried a top-level transcriptPath the ACP schema does not
allow, and status_update / mcpStatus_update went to every client, which
standard SDK clients answer with a method-not-found error each. The two
notifications now need clientCapabilities._meta["reasonix.io"].statusUpdates
and the path moves under _meta["reasonix.io"].
Problem
A development Studio shell started with `pnpm start` (`electron .`) had two
ways to outlive the launch that owned it:
- with its launcher killed outright, the shell and its kernel kept running
with nobody holding them;
- with the launcher's console closed underneath it (a terminal gone, a
background job's reader exited), the next kernel stderr line the shell
echoed raised an EPIPE that nothing handled. Electron's modal "A JavaScript
error occurred in the main process" box then held the main thread, so the
shell stopped draining the kernel and ignored SIGTERM (still alive 30 s
after it, 2/2 runs). The kernel logs a request every few seconds, so this
followed within seconds of the console closing.
Root cause
main.js echoed kernel stderr with a bare process.stderr.write and attached
no error listener, and nothing in the shell watched for the loss of the
process that launched it. The kernel already ends with the shell through its
stdin lease; the shell had no equivalent towards its own launcher.
Fix
- src/launcher.js owns both: watchLauncher ends an unpackaged shell once its
launcher is gone (POSIX reparents an orphan; Windows keeps the dead pid, so
there it is probed), and consoleEcho drops output for a console that failed
or backed up instead of throwing or queueing. shell.log and host.log keep
every line.
- Both relaunch paths (the kernel's relaunch act and the X11 relaunch) hand
the next instance over first: the launcher exits with this process, so a
relaunched shell would otherwise read its own start as the launcher dying.
- shell.log records the shell's and the kernel's pids, which is how the live
check finds them.
Verification
- node --test test/unit.mjs test/linux-packaging.test.mjs: 181 pass, 1 skip.
The new real-pipe test fails (child exits 1 on EPIPE) without consoleEcho.
- pnpm launcher-live (new, real shell + kernel, macOS arm64): launcher
SIGKILL after and before the handshake, closed console then SIGTERM, plain
SIGTERM all pass; with the fix removed the first three fail.
- pnpm smoke: 62 ok, 1 fail ("the bridge exposes verbs and nothing else"),
identical on unmodified studio.
Problem: newer Studio settings UI and persona translations both appended to
the shared English catalog, leaving the PR conflicting. Its new setting-open
notices also lacked a destination for the guarded persona fields.
Root cause: translations for an independent settings domain shared an append
point, and the newly introduced frontend routing table did not know persona.
Fix: normally merge Studio 15de9ec33, preserve both sides' translations and
extract persona's unchanged entries into its own catalog module. Route the
three guarded persona setting ids to their editor and cover actual notice
clicks. Preserve the persona's user scope, supplemental prompt semantics,
save timing, settings navigation and existing authorization boundaries.
Verification: format, vet, lint, make check, complete root Go tests with
atomic coverage, Desktop Go, frontend typecheck/build and 657 files / 5400
tests. Verified all 25 extracted translations unchanged. Chromium navigation
passed in Chinese/English, light/dark and wide/390px layouts; settings rows
and font scaling pass. The later dispatch config/catalog combination also
passes an explicit-marker build overlay. No thresholds or approvals changed.
The final-readiness pause answered session/prompt with the non-standard
stopReason "error", so a client that reads only message text took unverified
work for finished. The turn now ends with end_turn, the typed outcome in _meta
and a host-written [paused] line at the end of the streamed text.
Problem: newer Studio introduces the model-facing setting tool across the
same catalog, boot assembly and controller used by the persona feature.
Root cause: validating only the earlier base would leave their combined
authorization and request contracts untested.
Fix: normally merge Studio 4840df982 and retain the existing guarded persona
entries, user-only loading and supplemental prompt assembly. Add a real
boot.Build effect test for all three persona fields: even in YOLO a model
proposal receives a guarded refusal, produces no approval card, preserves
the user file and discloses no value through refusal or settings listing.
Verification: gofmt, vet, lint, make check, full root Go tests with atomic
coverage, Desktop Go and SDK suites, frontend typecheck/build and 654 files
/ 5384 tests; Chromium settings navigation, row layout and font-scaling
checks in light/dark and wide/narrow layouts. Earlier Electron checks apply
unchanged. No baselines, thresholds, workflow approvals or draft state change.
Problem: the approved persona head failed hosted Go tests and coverage, and
the combined Settings/config files exceeded their line budgets.
Root cause: CI tested synthetic merge 6309314327 against newer Studio
5e53be32a. Its attachment ceiling key had no settings declaration, while
independent additions accumulated in the shared settings/config files.
Fix: normally merge Studio 667f2ca16, including the upstream guarded
attachment-key declaration and ProviderEntry method extraction. Keep all
three guarded persona keys and their user-scope/prompt behavior. Extract
the unchanged settings navigation into SettingsNav, retaining keyboard
navigation, status values and the new version dot. Assert that the
attachment storage ceiling stays guarded.
Verification: gofmt, vet, lint, make check, complete root Go tests with
atomic coverage, independent Desktop Go tests, Electron tests, frontend
typecheck/build and 654 files / 5381 tests. Chromium fs-scale and settings
row guards pass, with light/dark 1280px/390px navigation checks. No budgets,
test thresholds, workflow approvals or draft state are changed.
stopReason "error" is not in the ACP prompt-turn enumeration, and the cause
travelled only on a vendor notification, so a text-only client saw an
unexplained failure. The turn now ends with -32603, the redacted cause as the
message and data.code turn_failed.
Problem: remote sidebar mutations could dial an unconfigured SSH target,
accept unbounded JSON, and enable hover actions without a user choice.
Root cause: a missing host resolved to an empty workspace without an error;
remote JSON handlers lacked a shared size boundary, and the preference
treated every value except explicit off as enabled.
Fix: classify unknown hosts before link reuse or acquisition, share a
64 KiB single-body decoder across remote writes, and restore explicit-on
hover preferences with a safe storage-error default. Preserve configured
hosts, far-kernel refusals, menu access and declared workspace resolution.
Align synchronization fixtures and keep translated directory prose in the
interface font.
Verification: failing-first unknown-host and oversized-body regressions;
full serve, builtin and boot tests; 5389 frontend tests; production build;
gofmt, root go vet, make lint and make check; isolated browser interaction
checks for default-off, opt-in persistence, menu access and opt-out.
Problem: The SSH project menu re-derived a default from the project list when
the host response omitted workspace.
Root cause: A client fallback duplicated the kernel's configured-default rule.
Fix: Pass the host response's workspace directly to the menu. Align the fixture
with the server response and cover both an explicit default that differs from
list order and an omitted default that must stay unset.
Verification: The omitted-default regression failed before repair. All 17
remote project/worktree interaction tests and the production build passed.
gofmt, root go vet, make lint, make check and builtin/boot tests passed.
The key landed with the streaming attachment ceiling without a class, so the
settings package panics at registration and every test importing it fails.
config.go is over its file-size budget on studio after two merged changes,
which fails lint for every new pull request. The model-price/override and
API-key methods belong with ProviderEntry; the one function that reads the
process environment stays in config.go so no process-input debt moves.
Problem: the Studio browser guard on the persona PR measured file and
change counts staying at 10px when interface text was enlarged.
Root cause: both explorer headers share a small element without an explicit
font size, leaving the size response to the browser's smaller-text rules.
The stylesheet is inherited from studio; macOS Chromium did not reproduce
the Linux hosted failure.
Fix: give the shared counter rule an explicit 10px base multiplied by the
existing interface scale. Preserve the default appearance and the persona
and composition behavior. Add a bilingual release note.
Verification: gofmt, vet, lint, make check, builtin/boot tests, frontend
typecheck, production/perf builds and 651 files / 5370 tests; Chromium
fs-scale, text-size and explorer guards. In light and dark, both counts
measure 10px, 11.5px and 13px across the three interface size steps.
The metadata list mapped entries to an unkeyed fragment, so React warned about the missing key when Settings -> Extensions rendered a skill's spec. Give the fragment the entry key and add a regression test that fails without it.
Problem: the latest studio settings catalog must know the persona fields, and
opt-in Standard mode must retain access to the prerequisite user-persona UI.
Root cause: the new catalog landed during the composition review, while the
initial composition page had replaced persona editing with a default adapter.
Fix: integrate the corrected persona parent and current studio, reuse its
editor in Standard mode, preserve settings-search navigation, and distinguish
user persona edits from immutable composition bindings. Keep the accepted two
cards, compact left picker and right run information.
Verification: gofmt, vet, lint, make check, root Go suite including boot/builtin,
frontend typecheck/build and 656 files / 5,385 tests pass. Parent persona keys
are Guarded; the actual request-byte and old-reader probes remain unchanged.
Problem: the new studio settings catalog rejects config fields without an
explicit classification, including the three persona fields added here.
Root cause: the catalog foundation landed after the persona implementation.
Fix: integrate studio and classify persona prefix, suffix and optional runtime
context as user-owned Guarded settings. Plain file writes do not rebuild a
runtime. Reject model proposals before any persona value can be changed.
Verification: root Go tests including settings, boot and builtin pass; gofmt,
vet, lint and make check pass; frontend typecheck/build and 5,370 tests pass.
Problem: default selections bound every new session and migrated the user's
persona, changing the ordinary path and making downgrade impossible.
Root cause: entry points manufactured a selection, persona settings were
adapted into the preset store, and blank unbound rewrites were not persisted.
Fix: keep unselected and Standard sessions unbound, preserve the user persona,
require explicit non-Standard selection, and restore durable state on failed
publication from either origin. Reuse the corrected persona parent, preserve
Standard/Minimal cards and the compact unified picker, and remove committed
composition captures. Freeze both skill discovery paths without copying locks.
Verification: root Go suite, vet, lint, make check, Desktop/SDK suites, Electron
unit tests and focused frontend suites pass. Real Build HTTP bytes match the
persona parent; its original reader can continue a new unbound transcript.
Project authorization, rollback and skill discovery regressions pass.
Problem: project configuration could replace the user persona, and complete
mode could suppress built-in guidance. The UI misstated rebuild semantics.
Root cause: persona was absent from the user-scope hold, replacement was a
supported runtime option, and settings metadata described only new sessions.
Fix: hold and clone the user's persona across project decoding, retire complete
throughout config/runtime/API/UI, and describe next-build activation. Integrate
current studio without dropping its settings or transport contracts.
Verification: gofmt, vet, lint, make check, root and Desktop Go suites, boot
scope/prompt regressions, frontend typecheck/build and 5,369 tests pass.
Problem: studio's native-video projection refactor overlaps the local-only
Agent composition metadata filter at the provider boundary.
Fix: keep the shared projection predicate and include composition metadata in
its host-only fields. Preserve stored bindings/video references and remove
answered video payloads only from model requests. Add regressions for a binding
as the sole copy trigger and for binding/video projection together.
Verification: gofmt, go vet, make lint/check, uncached root Go tests,
provider/OpenAI/boot/builtin suites, Desktop Go tests, frontend typecheck,
production build and 5,348 frontend tests. Existing mode layouts are preserved.
Problem: studio's MCP-follow changes and session compositions both own the
publication of a server tool catalog.
Root cause: the new followed-tool metadata must travel with the same catalog
whose fingerprint the immutable session binding validated.
Fix: retain follow metadata and the composition check before catalog publication.
Preserve upstream account localization and MCP-follow behavior. Add refresh,
stale-adapter and reconnect regressions with following enabled and disabled,
plus a real boot/provider test showing old bindings block drift while new
conversations can bind the changed server.
Verification: gofmt, go vet, make lint/check, root Go and boot/builtin suites,
Desktop and SDK Go suites, composition-guard race tests, frontend typecheck,
production build and 5,347 frontend tests. Browser evidence remains in design QA.
Problem: the shipped catalog and composer exposed more mode choices and status
badges than the requested Standard/Minimal workflow needs.
Fix: ship two native compositions, with a full-configuration Standard mode
and a terminal/image Minimal mode using a concise persona. Preserve retained
bindings and report retired defaults explicitly. Separate built-in cards from
custom presets, keep only the mode name in the trigger, and place run status
at the right. Keep source/lock details in the accessible label and menu.
Verification: gofmt, go vet, make lint/check, required boot/builtin tests,
complete root and Desktop Go suites, frontend typecheck/build and 5,343 tests.
Provider-boundary and custom read-only fixtures retain capability-denial
coverage. Browser checks cover light/dark, 1440px/390px and locked controls.
Problem: Host reads could borrow another pane's workspace or reuse an existing
link after its host was disabled. Review also requested explicit defaults,
boundary regressions and removal of development notes from the tree.
Root cause: remoteLink returned the first host pane before reading current
user configuration, so neither workspace identity nor the disabled policy
governed the reuse path.
Fix: Resolve the user-host default before reuse and accept only a matching
host/workspace pane. Share the declared-default-first rule with host views,
show actual and default connection paths separately, and document the rule
in both GUIDE languages. Remove three root QA notes and seven captures;
preserve their useful context in the PR discussion.
Verification: Reproduced both incorrect read behaviors before repair. Added
HTTP regressions for declared defaults, fallback, empty configuration,
reordered extras, explicit opens, project injection and disabled hosts with
existing panes. Focused frontend tests (80), production build, owning Go
serve/config and builtin/boot tests, gofmt, go vet, lint and make check passed.
Problem: the session-composition branch conflicted with the latest studio
settings, composer, peer-thread and runtime-construction changes.
Root cause: both branches extended shared port declarations, provider wiring
and the composer status row; failed-build cleanup also gained a separate
cancellable construction lifetime upstream.
Fix: retain both sets of settings and providers, preserve the top-row unified
Agent menu and responsive anchoring, and combine lifetime cancellation with
composition resource cleanup. Keep upstream accessible status labels and
text-size scaling. Add provider-boundary coverage showing the new image tool
cannot exceed review/research tool selections, plus updated visual evidence.
Verification: gofmt, go vet, make lint/check, required boot/builtin tests,
root and Desktop Go suites, frontend typecheck/build and 5,342 tests, Electron
unit tests, and browser light/dark checks at 1440px and 390px.
Problem: The unified sidebar conflicted with newer Studio worktree grouping
and text scaling changes.
Root cause: Both branches evolved project identity and row presentation from
the same baseline, including helpers extracted by the sidebar change.
Fix: Preserve workspaceContains across grouped local and SSH projects, keep
isolated session directories in open and palette routes, retain isolation and
peer-thread markers, and preserve font scaling. Include reviewed synthetic UI
captures and a release note for the relocated controls.
Verification: The integrated frontend suite passed 5381 tests. Four focused
grouped SSH/palette regressions and the production build passed. gofmt, go vet,
make lint, make check and builtin/boot tests passed without widening baselines.
Problem: Session composition controls competed with task actions in the
bottom toolbar, and the adjacent ellipsis did not explain how to manage
presets.
Root cause: Composition selection and management were rendered as separate
controls below the input. Floating menu placement also missed fold changes
published after window resize and the end of layout transitions.
Fix: Place one Agent presets menu beside the run status, show sources and
capability notes, and group management below a separator. Preserve locked
bindings and explicit new-conversation actions. Reuse the existing Picker
with per-choice action identities, disabled-row keyboard navigation and
current-state semantics; refresh shared floating anchors after reflow.
Update bilingual guides, release notes and before/after screenshot evidence.
Verification: All 633 frontend test files / 5,225 tests pass, including
selection/management, keyboard focus, pending/failure, lock/new-session,
source identity and responsive positioning regressions. Typecheck,
production build, gofmt, go vet, make lint, make check, and required
boot/builtin suites pass. Browser light/dark, wide/390px and resize recovery
were verified with synthetic fixtures.
Problem: Machine-grouped navigation made project switching cumbersome, and
session hover actions had inconsistent selection and keyboard behavior.
Root cause: Local and remote lists owned separate presentation and opening
flows. Remote defaults, paused stream costs, mutation deduplication and focus
restoration did not consistently survive asynchronous navigation.
Fix: Present one project list with source metadata, a common project picker,
fold-all control and bounded recent-session previews. Share neutral selection,
leading pin and trailing archive actions across local and SSH conversations.
Route new sessions and quick switching through the active environment, retain
latest navigation intent, and guard pending mutations and keyboard focus.
Forward remote session flags to the owning kernel and preserve typed errors,
configured workspaces and dedicated-stream connection costs.
Verification: 5263 frontend tests passed, production TypeScript/Vite build
passed, and focused browser and real SSH navigation checks are documented in
the QA reports. Owning Go serve/remotehost/attach suites passed. Pre-commit
gofmt, go vet, make lint, make check and builtin/boot tests passed without
widening the repository lint baseline.
Problem: ordinary binary uploads can consume unbounded workspace disk, and the
review needs direct coverage of file versus directory grants at read consumers.
Root cause: the streaming saver copies until EOF without a byte policy; the
project-config merge must not become an authority for a user disk safeguard.
Fix: enforce a named 1 GiB default while counting real bytes, retain rooted
cleanup on overflow, and return typed ceiling/source facts. Allow user config
to set the limit or explicit zero, hold it across project merges, render it
only at user scope, and strictly read it without degraded config fallback.
Cover selected-file siblings and directory link escape through actual reference
and tool consumers. Incorporate current Studio without rewriting history.
Verification: the former 8-byte/9-byte acceptance fails first and now refuses
cleanly. Full config race and focused control/serve/builtin races, frontend
tests/build, actual Electron rejection/retry, native Windows ARM64 boundary
and cleanup regressions, pinned lint and repolint pass.
Problem: the attachment PR conflicts with Studio's extracted composer helpers
and the newer worktree setup surface.
Root cause: both branches changed the Composer's chip type and helper ownership.
Fix: merge origin/studio, use its shared composerChips owner and worktree setup
flow, and retain cancellable attachment uploads and native path references.
Remove the redundant local chip type instead of duplicating the shared owner.
Verification: TypeScript checking, 155 focused frontend tests, production build,
attachment/CSRF/cloud/control/fileutil Go regressions, 169 Electron unit and
packaging tests (one existing conditional skip), pinned lint and repolint pass.
Problem: an admitted turn could fail before adding its user message and unlock
its composition. Controller-owned continuations could strand ACP config changes,
and MCP disconnect could remove code needed by a subsequent reconnect.
Root cause: the started latch followed message insertion, ACP drained only its
own prompt returns, and connection cleanup owned session package resources.
Fix: durably latch composition admission before executing or acknowledging a
turn and refuse work while the latch is pending. Notify owners at real idle
transitions and drain ACP changes through its existing serialization domain.
Separate transport state from package lifetime, retain script sibling bundles,
exclude VCS metadata, reject linked transcript exports and freeze late tool names.
Add ACP exact-binding/history tests, actual old-reader evidence, bilingual notes
and the complete light/dark wide/390px before-and-after screenshot set.
Verification: deterministic async/sync admission and failed-save regressions,
ten staged background-continuation/config-switch runs, real MCP reconnect cycles,
shared lifecycle race tests, package coverage, vet, lint and repository checks.
Frontend full suite and independent Desktop/SDK/Electron suites remain green.
Refs: #13613
Problem: Studio persona settings cannot select an independent tool and module
combination for each conversation, or preserve that combination across reloads.
Root cause: discovery and runtime construction follow mutable shared configuration
without a versioned session composition or a shared admission boundary.
Fix: introduce versioned scoped definitions, retained resources and authoritative
session bindings. Assemble isolated runtime selections, enforce tool ceilings,
pin MCP contracts, and share the lifecycle gate with first-prompt admission.
Expose independent HTTP, CLI, ACP and Studio management and selection controls,
with migration, native import/export, diagnostics and read-only history recovery.
Verification: composition mechanism tests, provider-wire cache tests, lifecycle
rollback tests, sensitive-boundary regressions, frontend typecheck and all 5,219
frontend tests; gofmt, vet, lint and repository checks. Browser light/dark wide
and 390px screenshots and real Electron backend interactions are included.
Refs: #13588
Problem: CodeQL reports uncontrolled paths in streamed upload cleanup and
attachment creation. A replaced storage directory can redirect failed-upload
cleanup to an external same-name file; a linked .reasonix parent can escape too.
Root cause: filename validation and path-based writes/removals do not preserve
directory identity across filesystem changes during I/O.
Fix: use os.Root for workspace-contained directory/file creation, share an owned
file lifecycle between streamed and legacy byte writes, and retain the directory
handle for failure cleanup. Publish the existing untracked marker atomically
through the same rooted directory rather than reopening its pathname.
Verification: full control/serve/fileutil suites, focused race tests, vet,
pinned linter and repolint passed. Native Windows 11 ARM64 regressions passed
with directory junctions and protected cleanup, plus large bounded streaming.
The independent replay test failure also reproduces with the PR's Go changes
removed through a base-source overlay; no replay behavior or CI gate is changed.
Problem: Review requested a user-visible release fragment and clearer memory
bounds. The live regression also assumed new-session activation must create
a second runtime, which is false when an existing blank pane is reused.
Root cause: A workspace may already own a reusable blank pane, particularly
when native Windows roots have matching canonical spelling. The old setup
waited for a POST and a new pane identity instead of editor readiness.
Fix: Wait for boot completion at a stable viewport, allow a blank-pane reuse,
and preserve assertions for the captured workspace, one shared intake route,
saved upload size and four ready attachments. Allow an explicit test host
binary. Add bilingual release notes and state that memory, not total disk
consumption, is bounded by streaming.
Verification: Windows 11 ARM64 control, stream, HTTP upload/image-limit and
exact-file read-tool regressions passed. Electron 44.2.0 exercised picker,
paste and drop of a 111 MiB file and a 65 MiB streamed File with four ready
attachments. Release-note tests, repository checks and script syntax passed.
Problem: The full frontend suite rejects the upload abort listener as an
unresolved interaction registration.
Root cause: The syntax-based census does not infer a Request.signal member's
AbortSignal type and cannot distinguish it from a possible user-input target.
Fix: Keep an explicitly typed upload signal and teach the shared receiver
certifier to recognize scoped AbortSignal bindings. Preserve unresolved
receivers and refuse borrowing an outer signal type across shadowing.
Do not add a file exception or raise any census or lint ceiling.
Verification: Abort receiver, actions census, cloud upload and attachment
regressions, frontend typecheck, production build and repository checks pass.
Problem: Large ordinary file attachments cause avoidable whole-file
allocations and exceed the legacy JSON attachment transport limits.
Root cause: Picker and paste intake Base64-encode files even when the shell
already knows their native paths. Pathless uploads also pass through JSON.
Fix: Share native path references across picker, paste and drop. Stream raw
ordinary-file bodies to bounded disk writes and clean up failed or cancelled
uploads. Grant external-file access only to the selected file. Preserve
image, legacy JSON and cloud relay limits; bound cloud intake before
materializing an oversized body. Document the transport and live regression.
Verification: Owning Go package suites, focused race tests, go vet, lint,
repository checks, frontend typecheck, production build and attachment
regressions passed on the original tested baseline. Real macOS Electron
exercised native picker, paste and drop with a 111 MiB file and an exact
65 MiB pathless upload. Windows native runtime verification remains open.
2026-10-17 11:08:58 +08:00
2194 changed files with 138845 additions and 5668 deletions
go test -count=1 -timeout=3m -json -run '^Test(CaptureDirectoryMaterializesContainedLinks|CaptureDirectoryRefusesUnsafeLinks|CaptureHookFileFlattensContainedLink|MaterializeStillRejectsLinkedRetainedFilesAndRoot|WindowsCompositionJunctions|WindowsLiveSkillJunctions|CompositionRetainsLinkedSkillAcrossRestart|CompositionLinkedSkillRejectsLinkedResources|CompositionLinkedExtensionResources|CompositionLinkedHookResources)$' ./internal/state/agentpreset ./internal/ext/skill ./internal/assembly/boot | tee "$RUNNER_TEMP/composition-links.jsonl"
platform=$(git diff -z --name-only --no-renames "$BASE_SHA" HEAD | bash scripts/ci-platform-scope.sh)
if git diff --name-only --no-renames "$BASE_SHA" HEAD |
grep -qE '^(desktop/electron/|desktop/packaging/|desktop/cmd/update-helper/|desktop/cmd/studio-manifest/|desktop/go\.(mod|sum)$|cmd/reasonix-studio-host/|internal/platform/(update|appupdate)/|go\.(mod|sum)$|scripts/(linux-install-smoke|verify-linux-package|verify-studio-artifacts)\.sh$|scripts/(windows-signing-lib|verify-windows-package|verify-windows-authenticode|install-smoke)\.ps1$|\.github/workflows/(studio|release-studio)\.yml$)'; then
grep -qE '^(desktop/electron/|desktop/packaging/|desktop/cmd/update-helper/|desktop/cmd/studio-manifest/|desktop/go\.(mod|sum)$|cmd/reasonix-studio-host/|internal/platform/(update|appupdate)/|go\.(mod|sum)$|scripts/(linux-install-smoke|verify-linux-package|verify-linux-appimage|verify-studio-artifacts)\.sh$|scripts/(windows-signing-lib|verify-windows-package|verify-windows-authenticode|install-smoke)\.ps1$|\.github/workflows/(studio|release-studio)\.yml$)'; then
@@ -93,6 +93,16 @@ Prefer a terminal? Download the `reasonix` command-line tool, run `reasonix setu
All four are the same program underneath. What follows is for advanced users and contributors.
## Use it from DeepSeek Harness
DeepSeek Harness (`dsh`) can hand a task to Reasonix as a subagent over ACP. This repository publishes a plugin index for it:
1. Run `reasonix setup` on the machine that runs `dsh`. In the DSH Web UI turn on Custom plugin markets (Plugins, More) and add this index in the Discover tab: `https://raw.githubusercontent.com/esengine/DeepSeek-Reasonix/studio/integrations/dsh-plugin/market-index.json`
2. Install "Reasonix subagent" from it.
3. Open the plugin's page and turn its switch on. While it is off the model has no `reasonix` tool and may run `reasonix -p` in its own shell, which looks like a delegation and is not one.
Setup details, the permission policy and limits are in [Using Reasonix from DeepSeek Harness](./docs/DSH.md).
## Versions
Reasonix ships on two lines. See the [version roadmap announcement](https://github.com/esengine/DeepSeek-Reasonix/discussions/10748) for the reasoning behind the split.
@@ -151,7 +161,7 @@ pre-releases):
| --- | --- | --- |
| macOS | `.dmg` or `.zip` | Apple Silicon (`arm64`) / Intel (`amd64`) |
| Windows | Installer `-installer.exe` or portable `.zip` | x64 |
| Linux | `.deb` or `.rpm` | x64 (`amd64`) / ARM64 (`arm64`, aarch64) |
| Linux | `.deb`, `.rpm`, or `.AppImage` (download only) | x64 (`amd64`) / ARM64 (`arm64`, aarch64) |
Every package ships with a `.minisig` signature, and the release carries
`SHA256SUMS`. Once installed, Studio updates itself in place.
@@ -240,12 +250,14 @@ For advanced CLI usage and configuration, see the **[CLI reference](./docs/CLI.m
`verify-page.sh` is the one to run after a change: it starts the kernel with the
same arguments `Kernel.kt` uses (`--auth token`, `--token-file`, `--port-file`,
`-page`) and checks that it binds a port and answers 200. Without a device, that
is as deep as the chain can be verified.
## Where the code is
| File | Owns |
| --- | --- |
| `MainActivity.kt` | The WebView, the back gesture, downloads, external links, renderer restart |
| `Kernel.kt` | Unpacking the page, starting and stopping the kernel, waiting for the port, health probes, logs, the workspace directory |
| `KernelService.kt` | The foreground service that keeps a long turn alive |
| `NativeBridge.kt` | `@JavascriptInterface`: what the page may ask the phone for |
| `BrowserShellServer.kt` | Loopback HTTP server the kernel's browser tools drive the overlay WebView through |
| `Downloader.kt` | Writing a kernel-produced file through the system document picker |
The page reaches the shell through `window.reasonixHost` and `ReasonixNative`,
defined in `desktop/frontend-next/src/port/host.ts` (`AndroidHost`).
**Phone and desktop share one page build.** What differs only on a phone is
decided at runtime, not by a second set of components. `phoneShell()` in
`src/port/host.ts` and the `[data-shell="android"]` selector in `mobile.css` both
read `<html data-shell="android">`, which `index.html` sets before the bundle
loads.
One attribute feeding both is what keeps style and components from disagreeing
about what a phone is.
## What the page may ask the phone for
`NativeBridge` is reachable from whatever document the WebView holds, so every
method that touches the device asks `trusted()` first.
The activity answers by comparing the **main frame's** origin against the
kernel's and against the one paired computer the person scanned a code for. A
page that navigated elsewhere is not trusted, even though it may still hold the
object.
Five methods answer without the check — `version`, `systemTheme`, `safeArea`,
`setLayerOpen`, `setRunning` — because they report back what the page already
knows about its own window and grant nothing.
`openInApp(url)` is the pairing decision, not a navigation one: it loads a URL in
this window only for the origin the scanner just returned, and only once.
Without that, any page holding the kernel origin could hand itself the bridge.
External links leave through the browser. A non-web scheme is routed to the
system only for `mailto:` and `tel:`; everything else is dropped, because an
`intent://` URL can start an arbitrary exported activity with caller-supplied
extras.
## Permissions, and what they cost
`MANAGE_EXTERNAL_STORAGE` — opening a folder on the phone as a workspace. The
kernel is a Go binary: it opens real paths with `open(2)`, and SAF hands back a
`content://` URI it cannot open.
Requested from the folder-picker action, never at launch. Android shows no dialog
for it and opens system settings instead.
**There is no OS-level sandbox around this kernel.** The PR sets the sandbox to
off on Android, so confining the agent to the chosen folder is Reasonix's own
write-root check and nothing else. A shell tool with all-files access can reach
all shared storage.
`REQUEST_IGNORE_BATTERY_OPTIMIZATIONS` — Doze freezes the app group, which stops
a turn that is still running. Asked once when a turn actually runs, never at
launch, and the person can refuse. Refusing costs battery life during long turns
and nothing else.
**Google Play restricts this permission**: fine for a sideloaded build, a
store-policy risk if this is ever listed.
`FOREGROUND_SERVICE` / `FOREGROUND_SERVICE_DATA_SYNC` — keeping the kernel alive
while the screen is off. `dataSync` is capped per 24 hours from Android 15. It is
the closest honest description of "an agent is working".
`CAMERA` — reading the pairing code a computer shows. Requested when the person
asks to scan, never at launch. `ScanActivity` is not exported.
`READ_EXTERNAL_STORAGE` / `WRITE_EXTERNAL_STORAGE` — the same folder walk below
API 33. `maxSdkVersion`-gated; above 33 they are superseded by the permission
above.
`allowBackup=false` with explicit extraction rules, so a `keystore` or a token in
app storage never leaves the device in a cloud backup.
## Cleartext HTTP
The kernel serves its page on `127.0.0.1:<port>`, and a paired computer serves
the same page over its LAN address. Both are plain HTTP.
Neither address can be written into a `network-security-config`: the port is
chosen at start, the LAN address at scan time, and the format matches literals
and domains, not subnets.
So the config permits cleartext and the boundary is in the code instead.
`shouldOverrideUrlLoading` sends every other http(s) origin to the browser, and
the bridge answers only the kernel and the just-paired computer. A plaintext page
on the LAN cannot become the document holding the bridge.
## What goes into the APK
| Content | Size | Why it is shaped this way |
| --- | --- | --- |
| `lib/arm64-v8a/libreasonix.so` | 45 MB | Go static binary, stripped. It is a `.so` rather than a file in `assets/` because only `nativeLibraryDir` is guaranteed executable — a binary copied to app storage fails on a device whose data partition is mounted `noexec`. |
| `assets/page/` | 14 MB | The frontend build, including the Inter and Noto Sans SC fonts, `noCompress` |
| `assets/LICENSE` | 1 KB | The repository's MIT licence, verbatim. MIT requires the notice and the licence text to accompany a binary distribution. |
`assets/LICENSE` is generated by `sync-page.sh` from the repository's root
`LICENSE`, not committed a second time: a checked-in copy drifts when the root
file is updated.
It sits in `assets/` rather than `assets/page/` because that directory is removed
and rebuilt on every build.
A release APK is about 31 MB.
## Building
**The Gradle wrapper is committed**, jar and all. `build.sh` refuses to run
without `gradlew` rather than generating one from whatever `gradle` is on `PATH`:
the dependency versions are part of the build.
`gradle-wrapper.properties` pins `distributionSha256Sum`, so a mirror that serves
a different distribution zip fails verification instead of being unpacked and
run. Dependency repositories are the official ones only — a mirror belongs in a
user-level init script.
**Signing.** A release with no key configured comes out **unsigned**
(`app-release-unsigned.apk`) with a warning. It used to fall back to the debug
key, which produced an artifact that looked shippable and was signed by a
per-machine key nobody chose: it could not install over the previous build.
Point the build at a real key by writing `keystore/keystore.properties`
(`storeFile`, `storePassword`, `keyAlias`, and optionally `keyPassword`) or by
setting the matching `REASONIX_KEYSTORE*` environment variables. Key material
stays out of the repository and out of build logs.
Debug builds always use the debug key. Signing a debug build with the release key
would let anyone who can read the keystore produce a build that installs over the
shipped one under the same signature.
**Non-ASCII checkout paths.** AGP refuses to build under a path with non-ASCII
characters on Windows unless `android.overridePathCheck=true` is set. That is a
property of one machine's disk layout, not of this project, so it is not in the
repository's `gradle.properties`.
Set it in a user-level `~/.gradle/gradle.properties` or pass
// The extraction script: prefer article/main for content pages,
// fall back to body. innerText preserves visual line breaks. It
// also stamps a ref on every element a person could act on, so the
// kernel can name one in an act step the way the desktop backend
// names an accessibility node. A ref already on an element is kept,
// so it stays the same across snapshots of one document; a fresh
// document starts the numbering over, and the kernel only accepts a
// ref from the snapshot it last read.
valscript="""
(function(){
var seq = window.__rxRefSeq || 0;
var used = {};
var marked = document.querySelectorAll('[data-rx-ref]');
for (var i = 0; i < marked.length; i++) { used[marked[i].getAttribute('data-rx-ref')] = true; }
function roleOf(el) {
var tag = el.tagName.toLowerCase();
var role = (el.getAttribute('role') || '').toLowerCase();
if (role) return role;
if (tag === 'a') return 'link';
if (tag === 'button') return 'button';
if (tag === 'select') return 'combobox';
if (tag === 'textarea') return 'textbox';
if (tag === 'input') {
var t = (el.getAttribute('type') || 'text').toLowerCase();
if (t === 'checkbox' || t === 'radio' || t === 'submit' || t === 'button' || t === 'reset') return t;
return 'textbox';
}
return 'generic';
}
function nameOf(el) {
var n = el.getAttribute('aria-label') || el.getAttribute('placeholder') || el.getAttribute('name') || '';
if (!n) { n = (el.innerText || el.textContent || '').trim(); }
return n.replace(/\s+/g, ' ').slice(0, 160);
}
var sel = 'a[href],button,input,select,textarea,[role=button],[role=link],[role=checkbox],[role=radio],[role=tab],[role=menuitem],[contenteditable=true]';
var els = document.querySelectorAll(sel);
var elements = [];
for (var j = 0; j < els.length; j++) {
var el = els[j];
var ref = el.getAttribute('data-rx-ref');
if (!ref) {
do { seq++; ref = 'e' + seq; } while (used[ref]);
el.setAttribute('data-rx-ref', ref);
used[ref] = true;
}
var raw = (el.value != null && String(el.value) !== '') ? String(el.value).slice(0, 160) : '';
go run ./cmd/e2ebench -suite benchmarks/realrepo -bin "$PWD/bin/reasonix" -task rt-stale-skills
```
The agent reads its provider config from `REASONIX_HOME` (default
`~/.reasonix`); point it at a disposable one to keep benchmark sessions out of
your own.
The default `-budget` (800k tokens) can stop the suite partway through, hence
`-budget 0`. Under `-trials`, each trial keeps its own
`<task>.trial-N.trajectory.jsonl`.
## The snapshot
Every `workdir/` is the same Go module (`module reasonix`, stdlib only, no
network needed to build or test):
- **Pinned code.** `git archive` of commit `42636ddcb425` for
`internal/contract/{pricing,ablation,agentgraph}` and
`internal/base/{retrieval,rulelist,tokencount,neterr,nilutil}` — about 5k
lines of real kernel code with its real tests and doc comments.
- **Fixture-authored files** in `base/`: `go.mod`, a trimmed `REASONIX.md` (the
project's own conventions, so the agent loads standing instructions the way
it does here), a `Makefile` with `check`, a module `README.md`, the
`cmd/benchctl` command that consumes `pricing` and `ablation` from outside
their package, and one project skill.
- **One pinned test anchored to the snapshot day.**
`TestCatalogPricesHaveBeenCheckedRecently` compares each row's `CheckedOn`
with the wall clock, which would fail every workdir 90 days after the oldest
pinned date (2026-11-20). `base/` replaces that file with one that reads
staleness at 2026-10-19 instead; nothing else in the pin is edited.
- **Per-task overlay** in `overlay/<task-id>/`, applied last. Only
`rt-stale-skills` has one.
`snapshot.sh` rebuilds every workdir from those three sources and
`snapshot.sh --check` fails if a committed workdir has drifted from them. Edit
`base/` or `overlay/`, never a workdir, then rerun the script. Moving the pin is
a new suite version: results from two pins do not compare.
## Tasks
| Task | Kind | What it exercises |
| --- | --- | --- |
| `rt-cache-write-rate` | multi-file Go change, test, README | Adds a `CacheWrite` rate to `RateCard` and threads it through cost, fingerprint, card identity and the catalog, then prices Anthropic's `claude-sonnet-4-5` with it: a new `officialRates` row, documentation and checked-on constants, an endpoint host, and `benchctl rate`/`cost` output. The fingerprint has to stay byte-identical for every existing card, so the change to `PricingFingerprint` cannot be the obvious one. |
| `rt-stale-skills` | documentation only, no recognised verifier | Three project skills name paths, functions, `make` targets and `benchctl` subcommands and flags that do not exist, and one contradicts the append-only rule in `REASONIX.md`. Fixing them takes reading `rates.go`, `ablation.go`, both `benchctl` files and the `Makefile`, then editing only Markdown — the shape behind #13293, where `stale_verification` stays owed after a skill edit because no check covers it. |
| `rt-rate-lookup-signatures` | API migration with a likely compile-error loop | Changes `CurrentRate` to `(RateCard, bool)`, `LookupCatalog` to a `CatalogKey` struct, and `IsOfficialCard`, `MatchesCatalog` and `LiveTariff` to a `ModelRef`: about 20 call sites across 11 files. Most are in `_test.go` files and dereference the old pointer, so `go build ./...` can pass while `go vet` and `go test` fail. |
| `rt-plan-ablation-module` | plan-mode with `complete_step` | Five numbered steps adding a `memory` ablation module, summaries, `ModuleHelp` and `benchctl arm -list`. The prompt opens with the host's own plan-first marker (`control.PlannerRouteMarker`), which routes the turn through the planner, and asks for each step to be signed off before the next. An existing test pins the `all` arm name and has to be updated, which adds a red-test step. |
Every grader stages hidden tests and checks after the run: `gofmt -l`, `go vet
./...` and `go test ./...` for the Go tasks; for `rt-stale-skills`, every
backticked path, identifier, `make` target and `benchctl` command line in each
skill must resolve against the tree, and a SHA-256 manifest proves no Go source, `go.mod`,
`Makefile` or `REASONIX.md` changed. Each task fails its pristine seed and
passes its `solution/`, both checked by `cmd/e2ebench`'s corpus tests.
## Reading the numbers
```text
Metric: requests per task
Anchor: the trajectory digest's requests_by_source, summed over sources
Population: every graded trial of a task that ran to its own stop
Excludes: tasks the token budget skipped, and runs that hit timeout_sec
```
Requests is the headline because round trips are what these tasks exist to
move. Read it beside the solve rate and the report's round outcomes: a change
that drops requests by giving up on verification has not cut a round trip, it
has cut a check. `rt-stale-skills` is read for whether the run finishes at all
and how many rounds go to re-verifying after its last write.
## Baseline
`deepseek-flash`, 2 trials per task, at studio `0b8de9e63`. Requests are the
trajectory's executor requests; all 8 runs ran to their own stop.
prompt="""Some providers bill tokens written into their prompt cache at a rate of their own. Add that to the pricing contract in this repository:
- `RateCard` gains a `CacheWrite` rate, per 1M tokens like the others. Zero means the card has no separate cache-write price, and cache-written tokens keep being charged at `Input` exactly as they are today.
- When a card sets `CacheWrite`, `OriginalCostAmount` charges the billed cache-write units (`CacheWriteBilledTokens` when set, otherwise `CacheWriteTokens`) at `CacheWrite` instead of `Input`.
- `PricingFingerprint` must tell apart two cards that differ only in `CacheWrite`, but a card without one must keep exactly the fingerprint it has today: stored ledgers are keyed by it.
- Two cards that differ only in `CacheWrite` are different cards everywhere cards are compared, including official-card recognition and catalog matching.
- `CatalogEntry` carries the rate too, and `RateCardFromCatalog` and `CatalogEntry.RatesAt` hand it back.
- Add Anthropic to the official price book: `claude-sonnet-4-5`, pay-as-you-go in USD, cache hit 0.30, input 3, output 15 and cache write 3.75 per 1M tokens, documented at https://www.anthropic.com/pricing and checked 2026-10-19. `api.anthropic.com` is Anthropic's own endpoint.
- `benchctl cost` gains a `-write` flag for cache-written tokens, and `benchctl rate` shows a card's cache-write rate when it has one.
- Add tests for the new behaviour and update the README's pricing section."""
{ID:"titled",Fields:map[string]string{"title":"Package manager","body":"We migrated to pnpm."}},
{ID:"mention",Fields:map[string]string{"title":"Sprint notes","body":"Someone asked about the package manager once; also weather, lunch, and a package of stickers."}},
}
hits:=RankV2("package manager",docs)
iflen(hits)!=2||hits[0].ID!="titled"{
t.Fatalf("field-weighted doc must outrank body mentions: %+v",hits)
t.Fatal("a free-form fraction was accepted; the axis is four steps on purpose")
}
}
funcTestZeroValueIsTheControlArm(t*testing.T){
varsSet
if!s.Empty()||s.Arm()!="full"{
t.Fatalf("zero Set = %q, want the full control arm",s.Arm())
}
for_,m:=rangeModules(){
ifs.Off(m){
t.Errorf("zero Set disabled %s",m)
}
}
}
Loaded 100 of 2194 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
YHH