* fix(web-shell): isolate worktree session execution
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(mcp): expect session cwd in effective config
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(acp-bridge): cover shell execution after a failed cwd change (#8068)
* fix(acp-bridge): abort direct shell waiting on a hung cwd change (#8068)
* refactor(core): centralize effective MCP server resolution (#8068)
Extract McpClientManager.getEffectiveMcpServers() as the single source of truth for the populateMcpServerCommand recipe, replacing six duplicated call sites that this PR otherwise had to edit in lockstep. Also cover relocateWorkingDirectory's combined memory + MCP refresh failure path so both errors stay surfaced.
* fix(core): align MCP cwd stamping with transport classifier (#8068)
* fix(core): move MCP cwd stamping to populateMcpServerCommand (#8068)
getMcpServers() injected cwd into the effective config, which changed
the hash used for MCP approval binding (hashMcpServerConfig). This
caused approved project/workspace-scoped stdio servers to show as
pending approval, and re-approving silently gated them off on the next
daemon start.
Move the cwd stamping down to populateMcpServerCommand — the
transport-config seam every spawn path already goes through — so
getMcpServers() stays an identity-stable view of configuration and the
approval hash contract holds. Also make populateMcpServerCommand
non-mutating (spread instead of in-place assignment).
Additionally, clean up the abort listener registered during the
direct-shell cwd-queue race so it does not accumulate on the
long-lived per-session signal (ACP path).
---------
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com>
Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
The side-task creation route still constructed a session service directly on
two rollback paths after the surrounding module moved to the workspace runtime
helper, leaving the identifier unbound and breaking the CLI build on main.
Route both rollbacks through the workspace runtime helper so they honour the
per-workspace runtime base dir like every other session-service use in this
module.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(web-shell): add contextual task panels
* fix(web-shell): harden contextual task panels
* fix(web-shell): preserve side task titles
* fix(web-shell): address review feedback on context panels PR (#7929)
- Add POST /session/:id/side-task to telemetry route catalog (51 routes)
- Increase SDK browser bundle size limit to 184KB
- Fix duplicated data-testid="chat-pane" → "chat-pane-container" on container
- Gate sourceType behind session_source_metadata capability check
- Add removeSession cleanup after killSession in !res.writable path
- Add i18n key sideTask.renameFailed for error fallback
- Add unit tests for selectVisibleHistoryRecords invariant
* fix(cli): update telemetry-catalog route drift guard to 51 routes (#7929)
* fix(web-shell): address review feedback round 2 on context panels PR (#7929)
- Fix /fork sider discarding createSideTask() return value: show toast
when side tasks are unavailable
- Fix layout feedback loop: availableWidth no longer depends on
environmentPanelVisible since the CSS overlay does not change the
chat pane DOM width
- Remove dead environmentPanelSuppressed state (never set to true)
- Restore setArtifactPanelOpen(false) in closeArtifactPanelTab when
the last tab is closed
- Extract agentDisplayName(task) to a local variable to avoid triple
invocation per render
* fix(web-shell): dedupe completed background agents in environment panel (#7929)
getEnvironmentAgentTasks correlated a transcript tool card with the live
/tasks snapshot only on toolUseId, the notification taskId, and a
<subagentType>-<callId> derived id. A completed background agent can lose
that linkage (its live task carries no usable toolUseId and its daemon id
is general-purpose-<internalId>), so the trailing loop appended the live
task as a second entry. Add a conservative content fallback (prompt, or
description+subagentType) mirroring the daemon's legacy resolver.
* feat(web-shell): support side tasks during active turns
* fix(web-shell): deduplicate completed subagents and gate sourceType on capability (#7929)
* fix(web-shell): restore background agent reconciliation and fix agent dedupe (#7929)
Restore the one-shot subagent reconciliation for inline background Agent tool
cards. Persisted notification records do not always retain a toolUseId, so the
SSE discrete-notification path alone can leave a card stuck in Running; the
documented fallback resolves pending cards through the subagent endpoint after
catch-up, reconnect, and terminal notifications.
Also stop the loose description content fallback in getEnvironmentAgentTasks
from claiming a live task that another transcript tool call already links
precisely (by toolUseId, message taskId, or derived id). Two agents sharing a
description previously collapsed into one: the fallback stole the linked task,
its owner re-matched the same task, and the orphan was dropped.
* fix(web-shell): address critical review feedback on context panels (#7929)
* fix(web-shell): reconcile side-task state across sessions and listings (#7929)
* fix(web-shell): preserve contextual panel fallbacks
---------
Co-authored-by: 钉萁 <dingqi.jww@alibaba-inc.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com>
Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com>
Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com>
Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com>
* feat(review): statement-level mutation probes in test-efficacy
The revert probe is all-or-nothing: it reverts whole production files, so a
suite that gates six of a diff's behaviours goes red and the probe says
"gated" — even when the seventh behaviour, a one-line safety statement, has no
test at all. Dogfooded on a live PR: deleting a single `reminders.clear()`
inside the not-continued branch left the full 471-test suite green, and that
line carried the PR's headline safety property (an abandoned task's todos must
not bleed into an unrelated new prompt). A human reviewer found it with a
hand-rolled mutation probe; the command could not.
Add the probe kind the human ran, deterministically. Candidates are added
lines from the committed head (never the dirty worktree) whose trimmed
statement calls a safety verb — .clear(), .delete(), .reset(), .abort(),
.removeListener(), .unref() — or reassigns state to empty ([] / new
Map()/Set()), and that are removable as a whole: single complete expression
statements, brace-balanced via a string/comment-aware scanner, previous
significant line ending ;, { or } (which rejects fluent tails, continuations,
and the brace-less-if silent-rebind trap), outside template literals and block
comments. Selection is conservative by design: a false negative costs
nothing, a false positive burns a suite run.
Each mutant (capped at 8, files with collocated new tests first) deletes its
one line in the existing probe worktree and re-runs the affected suites
through the existing vitest-json classifier: red = killed (the line is
guarded), green = SURVIVED — the invariant it enforces ships unprotected —
filed as a finding in the unreachable/inert register so the Agent-7 pipeline
picks it up without any skill change. Compile/load failure = inconclusive,
never a finding. Mutants run only after a cleanly green baseline, inside the
command's existing deadline budget (each run must leave room for the revert
probe); candidates that no longer fit are counted, not silently dropped.
* fix(cli): harden test-efficacy mutant selection text checks (#8020)
Mutant selection ran its end-anchored checks on raw trimmed lines, so a
trailing comment hid a statement's real end (dropping genuine candidates
like `reminders.clear(); // why`) and a safety verb inside a string faked
one (a wasted suite run plus a misleading survivor). Route the
`SAFETY_VERB_RE`, `endsWith(';')`, and predecessor `/[;{}]$/` checks
through a shared `codeOnly()` that strips comments and blanks literal
contents first.
Also guard the template-state escape skip against a backslash-continued
line, mirroring the single/double-quote branch: swallowing that newline
dropped a per-line literal flag and shifted every later line's verdict.
Adds unit coverage for all three selection fixes and an integration test
for the baseline-not-green skip branch.
* fix(cli): gitignore fake vitest in test-efficacy integration fixture (#8020)
The fake vitest bin was committed by `git add -A` and checked out into
the probe worktree as the stale passing copy, so installFailingVitest's
overwrite in the main worktree was never seen by npx in the probe tree.
The baseline read green and the baseline-not-green skip test failed.
* fix(review): make mutation-probe reporting precise (#8020)
Address review feedback on the test-efficacy mutation probe:
- Count candidates the MAX_MUTANTS cap drops in `skippedForCap` instead of
silently losing them, so a capped `survived: 0` cannot read as "every
safety statement is covered" (mirrors the existing `skippedForBudget`).
- Gate the mutant phase per probe file: run each mutant against the files
that are green in the unmutated baseline, so one unrelated quarantined
(all-skip) suite — `inconclusive`, not red — no longer disables the whole
probe.
- Scope the `mutant-survived` finding to the diff's own tests ("confirm an
existing test covers it, or add one") rather than asserting the invariant
ships unprotected, which an untouched pre-existing test may still cover.
* test(review): cover the budget-skip path in mutation-probe integration tests (#8020)
* fix(review): harden mutation-probe selection and diff parsing (#8020)
* fix(review): handle multi-line class headers in mutation-probe selection (#8020)
* fix(review): stop class-body walk at braces before matching class keyword (#8020)
* fix(review): whole-file literal scan for mutant selection; pin the untested guard paths
The per-line scanner pair (codeOnly + lineStartsInsideLiteral) shared a blind
spot: a backtick inside a `${…}` interpolation read as the outer template's
closing backtick. That flipped the literal state for every following line and,
in the single-line skip, exposed nested-template content as code — so a safety
verb inside a string could be selected as a mutant (a false-finding vector) and
a class field below a brace-bearing template could slip the class-body
rejection. Replace the pair with one whole-file pass that tracks interpolation
brace depth: per-line code text with comments stripped and literal contents
blanked, plus the same `${}`-aware skip for delimiter scanning. Differential
audit over every core+cli source file: zero selection differences on real code;
the pathological shapes are pinned by three new tests.
Also pin the remaining untested paths from review: the selection-failure catch
(discloses and still runs the revert probe), the runOneMutant line-mismatch
guard (now exported; inconclusive, file untouched), and the budget-skip stdout
disclosure.
* fix(review): clamp probe deadlines to the whole-command budget (#8020)
* fix(review): harden mutant selection guards and disclosure accounting (#8020)
* fix(review): close silent-zero paths in mutant disclosure and harden diff parsing (#8020)
* fix(review): drop interpolation quote-skip that mis-parsed regex literals (#8020)
* fix(review): track template nesting with a stack; disclose derailed files; harden and pin the remaining probe paths
The nested-template fix that landed as a counter cannot represent a nested
template INSIDE a nested interpolation: at two levels the deep template's text
`}` is charged against the wrong frame, the scan desyncs, and the file either
admits template text as a mutant or derail-drops its real candidates. Replace
the counter with a stack — one frame per open template, `}` closes only the
top interpolation, a backtick closes only the innermost template — and derive
the end state from the stack. The two-level trigger is pinned by a test
written red-first against the counter.
Derailed files are now disclosed, not silently dropped: selectMutants returns
them, and the note composer stacks the derail note with the red-baseline note
instead of clobbering. The hostile-git-config path gets its missing test (repo
diff.srcPrefix/dstPrefix, diff.external, core.quotePath with a non-ASCII
path — fails with the pinned flags removed). The budget test drops its
Date.now call-count coupling for an injected clock threaded through
runTestEfficacy/runProbeSuite. The mutation-phase catch gets an end-to-end
test (ENOBUFS mid-phase → all candidates inconclusive, revert probe still
runs, report still written).
---------
Co-authored-by: verify <verify@local>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com>
Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com>
Co-authored-by: Qwen Code <qwen-code@users.noreply.github.com>
Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com>
* fix(serve): prevent repeated workspace skill rescans
Make workspace skill status reads use committed snapshots and move refresh work to explicit mutation paths. Add generation-safe daemon caching, conditional HTTP responses, SDK revalidation, and multi-session extension refresh safeguards.
Refs #8079
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(serve): narrow the workspace-skills read model and close its regressions
Follow-up to the previous commit on this branch, from reviewing it.
Subtractions — these were separable from the fix and carried more surface
than value, so they move out of this change:
- Revert the ETag / If-None-Match layer (CORS allow+expose headers, the SDK
conditional JSON cache, the browser bundle budget bump). Express already
emits an ETag and answers 304 for these routes, so the only new behavior
was the SDK cache. It saves transfer bytes but no daemon work — the ETag is
a hash of the already-serialized body — and it shipped without a paired
`Cache-Control`, which is what actually keeps an intermediary from serving
a stale snapshot of an authenticated, mutable resource. The SDK cache was
also unbounded, with no eviction or clear entry point.
- Revert moving `extensions_final` ahead of skill initialization in
`Config.initialize`. In non-safe, non-bare mode `extensions_initial` is
already the same argument-less `refreshCache()`, and it runs
`applyStoreActivation`, so `getActiveExtensions()` is fully populated
before skills are enumerated either way. The move changed only startup
event order (and pushed permissionManager past the extension refresh) for
every surface including the interactive CLI.
Regression fixes — the read went pure, but two of its inputs lost their only
path back to disk:
- Extension sources have no watcher, unlike skills. With the per-read
`extensionManager.refreshCache()` gone, an extension installed, removed,
enabled, or disabled outside the daemon would never reach the snapshot
until the child restarted — and because extension-level skills are derived
from the extension set, a skill-watcher tick could not recover it either.
Adds `ExtensionManager.refreshCacheIfSourcesChanged()`: a stat-based
fingerprint over the extension directory entries, each manifest, the
enablement file, and the store state, which refreshes only when they moved.
A status read pays one readdir plus one stat per entry instead of a
directory scan and a full parse, and stays self-healing.
The baseline is the pre-load fingerprint, so a change landing during a
refresh stays visible to the next check instead of being masked by a
post-load stat. The directory and store halves are captured at different
points because a refresh writes the store itself but never the manifests.
- Revalidation is skipped in safe and bare mode, and the whole of it —
including that mode check — sits inside its error boundary. Those modes never
populate the extension cache by design, while the snapshot derives extension
skills from `getExtensions()`, so revalidating there would have loaded the
extensions the mode exists to exclude. Keeping the mode check outside the
boundary would also have let a config missing those accessors fail a read.
- `initialized: true` with an empty list when the config has no
`SkillManager` is now `initialized: false`. The daemon latches any
initialized answer into `lastWorkspaceSkillsStatus` and then prefers it
over its own local enumeration, so the old value could suppress the
fallback permanently.
Also:
- The retained-snapshot path bumped the freshness timestamp without checking
its generation, so a read that started before an invalidation could push
out the TTL of a snapshot a later read had committed — letting a
post-mutation snapshot go unrevalidated for longer than the window.
- `setWorkspaceSkillEnabled` folded `configsFailed` into `sessionsFailed`,
but it sends `reason: 'settings'`, which never refreshes a skill cache, so
the term was structurally zero. Report `configsFailed` from the `content`
path instead, where it can actually be non-zero.
- Documents the settings-freshness gap this read model accepts: enablement
now comes from the child's in-memory `LoadedSettings`, which `SettingsWatcher`
keeps current for the User and Workspace scopes but not for System /
SystemDefaults (locked-skill policy) or an untrusted workspace.
Tests: adds a real-filesystem guard that drives 50 consecutive cached reads
and asserts zero additional readdir/readFile calls — the mocked suites could
only prove `refreshCache` was not *called*, which is not the invariant that
broke. Adds coverage for the fingerprint gate (steady state, install,
removal, in-place manifest edit, concurrent callers, and the mid-refresh
race), for the null-manager, moved-sources, and safe/bare-mode read paths, and
for the generation guard. The generation-guard and safe/bare-mode tests were
each verified to fail with their fix reverted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(serve): allow bounded reads of large text files
* fix(serve): bound large-text reads by scan cost, not by which knob was set
Follow-up to the bounded large-text read path. Three changes:
Gate on any explicit window argument, not on `limit`. Gating on `limit`
had the cost model backwards in both directions: `{ line: 900_000_000,
limit: 20 }` was admitted despite walking the whole file, while
`{ maxBytes: 4096 }` — satisfiable from the first 4 KiB — was refused. A
read with no window argument at all still fails, since a caller that
believes it holds the whole file may write it back truncated.
Add MAX_TEXT_SCAN_BYTES (8 MiB). MAX_READ_BYTES caps what a read
returns; nothing capped what it cost. Line offsets are resolved by
scanning from byte 0, so a query param could turn into an
uninterruptible multi-second scan of an arbitrarily large file — and on
Windows hold a read handle for that span, blocking renames and deletes.
Past the budget the read is refused with `file_too_large` pointing at
readBytes, which reaches any offset in O(1).
Tolerate appends on streamed windows. Requiring whole-file size/mtime
stability after reading a prefix rejected reads whose returned bytes
were still valid, and the case it rejected — tailing a live log — is the
one this path exists for. Streamed windows now assert inode identity
plus "did not shrink"; truncation and replacement are still rejected.
Also: non-UTF-8 large text now returns `binary_file` rather than
`file_too_large`, so a client retrying on 413 with a smaller window
can't loop forever; and `readFileWithLineAndLimit` throws instead of
silently ignoring a caller-supplied `fileHandle` on the by-path
fallback.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(core): thread the descriptor instead of forking text-read helpers
PR #7947 pinned large-text reads to one inode by threading a caller-owned
FileHandle into readTextRange as an optional field, plus a second field,
forceStreaming, to suppress the buffering fast path. Two optional fields
produced four combinations: one meaningful, one used by a single test, one
unreachable, and — in readFileWithLineAndLimit — one that silently fell
through to a by-path read, defeating the reason the caller opened a handle.
Unify the two encoding detectors. detectFileEncoding now takes a path or a
borrowed handle, so detectFileHandleEncoding is deleted along with the
message discrepancy between them: an encoding iconv-lite cannot load now
raises LargeNonUtf8TextError naming that encoding rather than deferring to
the decoder's generic invalid-utf8 variant. Both still refuse the file, and
the Serve boundary maps both to binary_file.
Split the reader into readTextRange (path) and readTextRangeFromHandle
(always streams, both byte bounds required). The unreachable combination and
its untested readFileHandleBuffer are gone, and with no fileHandle parameter
left for readFileWithLineAndLimit to ignore, the RangeError guarding that
fallthrough is deleted too — the trap can no longer be expressed.
CoreReadTextFileHandleRequest drops its required stats field. Nothing
downstream read it, and because the ACP request type it extends permits
extra properties, TypeScript accepted the dead argument silently.
readFileHandleChunks becomes chunksFromHandle(fh, from) — the one seam
byte-cursor text paging needs.
No observable change at the Serve boundary: its 222 tests pass unmodified.
Two fileSystemService tests were deleted rather than repaired; they asserted
the arguments readFileWithLineAndLimit received, which is nothing once the
handle path stops calling it. Their coverage lives in read-text-range.test.ts
against real files and in workspace-file-system.test.ts at the real boundary.
258 production lines in core, net -71 overall.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(core): make CoreReadTextFileHandleRequest standalone
Self-audit follow-up to f55c867a. Two fields survived the reshape that the
handle path never reads:
- `stats` was documented as required ("must pass the Stats captured from that
handle") and nothing downstream read it. The handle path always streams, so
it never needs a size to choose a strategy, and the encoding probe does its
own fstat.
- `path` became dead once readTextRangeFromHandle replaced the path-plus-handle
call. Errors are labelled with the path by the Serve boundary that owns it.
Neither was caught by the compiler: the ACP ReadTextFileRequest the type
derived from permits extra properties, so the CLI kept passing both silently.
That is the argument for declaring the type standalone rather than Omit-ing
four of six inherited fields and quietly re-admitting the rest.
Also record the second behaviour delta of the detector merge in the design
doc: detectFileEncoding catches I/O errors and falls back to 'utf-8', where
detectFileHandleEncoding let them propagate. The failure is not lost — a handle
that fails the 8 KiB probe fails the streaming read immediately after — but a
different call now reports it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* feat(serve): page large text files by byte cursor
Line offsets address a byte stream, so `readText` resolves them by scanning
from byte 0. Paging a large log that way is O(n^2) across pages, and past
MAX_TEXT_SCAN_BYTES (8 MiB) a deep page is refused outright — agents had no
O(1) path short of dropping to GET /file/bytes and splitting lines themselves,
losing encoding handling, multibyte safety, and the binary_file refusal.
A response that leaves content behind now returns `hasMore`, and where a file
byte offset is derivable, an opaque `nextCursor`. Passing it back as `cursor`
resumes in O(1). Page 1 is an ordinary `limit` read, so clients never compute
byte offsets themselves, and a paging loop does not break when a file happens
to be small.
The cursor is unsigned base64url JSON carrying {off, size, dev, ino}, matching
encodeOrganizedCursor rather than the HMAC-signed transcript codec: the path is
re-resolved through the workspace boundary on every request, so a forged cursor
can only move the offset within a file the caller may already read — what
GET /file/bytes?offset= allows today. What the payload is for is staleness:
a replaced or truncated file yields hash_mismatch instead of bytes from the
wrong place, while an append leaves an outstanding cursor valid — the case the
feature exists for.
Every minted cursor points at the start of a line. When a single line exceeds
maxOutputBytes the reader emits a truncated prefix and skips to the next line
rather than resuming mid-line, because a mid-line cursor makes the following
page snap forward and silently drop the rest of that line at the seam. Windows
cut mid-line by a byte cap therefore report hasMore with no cursor, as do
non-UTF-8 snapshot reads whose decoded text is a UTF-8 re-encoding with no
mapping back to file offsets. That is why hasMore is a field rather than a
restatement of nextCursor.
Cursor reads branch before the size check, not by widening the window gate:
a cursor read of a file under MAX_READ_BYTES would otherwise land on the
snapshot path, which knows only line/limit, and silently return line 0.
Adds the workspace_file_read_cursor capability, per the convention that new
behavior gets a new tag, and retargets the scan-budget hint at cursor paging.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(core): advance UTF-8 cursors after truncation
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* docs(serve): clarify cursor bootstrap limits
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(sdk): raise daemon browser bundle budget
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(serve): cover ACP cursor dispatch and cursor binary_file mapping (#8002)
* fix(core): only set sawCrlf for emitted lines in cursor paging (#8002)
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com>
* feat(core): tag UserPromptSubmit hook context and record display provenance
UserPromptSubmit additionalContext was appended to the request as a bare
text part and persisted verbatim, so hook-injected text was
indistinguishable from user-authored text in the transcript, polluted
resumed sessions, telemetry, and auto-memory recall queries.
- Wrap injected context in a reserved
<qwen:user-prompt-submit-context> tag (hook output already escapes
angle brackets, so the tag cannot be forged from inside).
- Record the pre-injection user prompt as systemPayload.displayText plus
the injected string as hookContext on the user record; the model-bound
message stays verbatim for faithful resume replay.
- Use the pre-injection prompt text for telemetry prompt attributes and
managed auto-memory recall.
- Resume projection prefers displayText, strips a trailing whole-part
tagged block when no payload exists, and leaves legacy bare-injected
records unchanged.
- Apply the same tag wrapping on the ACP session injection path, which
already records the pre-injection prompt.
Closes#7940
Co-authored-by: Cursor <cursoragent@cursor.com>
* docs: note UPS promptText TDZ ordering and sole-part resume guard
Document the conflict-resolution constraint that promptText must be
declared before the injection assignment, and the sole-part read-path
guard that keeps a user-authored whole-tag message intact.
Co-authored-by: Cursor <cursoragent@cursor.com>
* test(cli): cover at_command resume with tagged UPS context
Confirm the at_command branch still prefers payload.userText when a
paired user record carries a trailing tagged hook-context part, and
falls back to the tag-stripping projection only when userText is absent.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(core): address PR 7956 review findings and Goal recording spy
Omit the optional UserPromptRecordPayload third arg when no hook
injected, so Goal admission spies expecting two args stay exact and
CI client-goal.test.ts passes.
Project plain UserPromptSubmit-augmented records through
transcript-replay with the same displayText / trailing-tag strip
fallback as the TUI, covering ACP/export surfaces. Strengthen the
displayText preference fixture so it disagrees with the tag-strip
path, and use the named UserPromptRecordPayload type in resume.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(acp-bridge): import UPS tag helper via Node-free package export
transcript-replay is inlined into the browser daemon/transcript SDK
bundle. Importing isUserPromptSubmitContextPartText from the core
package barrel pulled the whole Node-bound core graph into that
bundle and failed CI (esbuild Could not resolve "node:*") across
Test, web-shell E2E, and Real daemon E2E.
Export the pure helper as @qwen-code/qwen-code-core/userPromptSubmitContext
and import that path instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(test): alias userPromptSubmitContext for Vitest source resolution
CLI and acp-bridge Vitest configs already map goalWire/transcriptRecords
to TypeScript sources; without the same alias the new package export
fails import analysis and breaks dozens of CLI suites.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(acp-bridge): keep images when projecting displayText user records
Preferring UserPromptSubmit displayText previously returned early and
skipped projectMessageParts, dropping multimodal inlineData. Rebuild
parts so displayText replaces text while images keep their order.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(core): drop unused hookContext and cover image-only displayText
UserPromptRecordPayload.hookContext had no read sites; keep displayText
only and recover injected text from the tagged message part. Also cover
the image-only !replaced append path and simplify the recording guard.
Co-authored-by: Cursor <cursoragent@cursor.com>
* test: cover remaining UserPromptSubmit provenance Suggestions
Share stripTrailingUserPromptSubmitContextPart between TUI resume and
ACP replay, assert ACP Session tags additionalContext, and lock
telemetry to the pre-injection prompt text.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(autofix): salvage race-lost pushes by merging the moved head and retrying
The review-address push is one-shot: when anything pushes to the PR head
during the agent's ~50-minute window, the final push dies 'fetch first'
and the entire verified agent run is discarded. The per-PR head-write
concurrency group cannot prevent this — it serialises this repo's
workflows, not the PR author or the fork side. Observed twice in one day
(#7983 after a 56-minute run, #7985 after 43 minutes).
On rejection, fetch the moved head, merge it into the local line, and
retry (bounded at 3 attempts). Merge rather than rebase: the agent's own
conflict-resolution rounds create merge commits, and a rebase would
flatten them and can silently re-introduce the conflicts they resolved.
The merge result descends from the remote head, so the retried push is a
fast-forward and rewrites nothing. A genuine content conflict aborts the
merge and falls through to the existing failure path unchanged.
When a salvage merge happened, the round report discloses that the
round's verification predates the merge so mid-run commits get
re-checked by a human.
* fix(autofix): address salvage-loop review findings
- Gate the PUSH_RACE_MERGED disclosure on HEAD actually advancing: a
transient push failure (upload timeout, 503) on an unmoved branch
no-ops the merge ('Already up to date') and must not tell the
reviewer to re-check mid-run commits that never existed.
- Annotate the salvage fetch failure with ::error:: like the two
adjacent failure paths, so a deleted fork branch or network error
does not kill the step with an unannotated exit 128 under bash -e.
- Re-pin the same-repo push URL construction in tests: it lost its old
'origin "${BRANCH}"' pin in this rework, leaving a ${REPO}→${HEAD_REPO}
mutation (malformed remote in the same-repo case) unkillable.
* test(autofix): restore dropped mutation-killing pins and add structural assertions (#8042)
* test(autofix): pin exit 1 in the give-up guard regex to kill the deletion mutation (#8042)
* test(autofix): pin exit 1 in the fetch-failure and merge-conflict salvage paths (#8042)
* test(autofix): strengthen salvage-test pins to kill init-value and capture-order mutations (#8042)
---------
Co-authored-by: verify <verify@local>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com>
* fix(autofix): never defer maintainer feedback in Critical-only mode
Critical-only mode (after 5 change-producing rounds) classifies
feedback lexically: only a literal **[Critical]** tag or a
CHANGES_REQUESTED review survives; everything else is deferred before
the agent reads a word of it. That rule was built to stop the review
bot's suggestion ping-pong, but it catches maintainers too. Observed
four times in two days (#8037, #7944, #7885, #7799): a maintainer's
review with explicit merge-blocking findings — #8037's said 'I'd fix
before merge' on a correctness bug and a security-adjacent one — was
wholesale-deferred as one 'non-Critical item', and the bot then
reported 'No Critical feedback. The Issue-level comments sections are
empty', which was lexically true and substantively false. The bot's
own advertised definition ('correctness bugs, security issues, or
formally requested changes') is exactly what the deferred comments
contained; the agent that could have applied that definition never saw
them.
The lexical test now applies exclusively to the review bot's output:
- All three actionable filters (reviews, inline, issue-level) pass
anything not authored by the review bot straight through in
Critical-only mode — the agent judges maintainer feedback on
content, as everywhere else.
- All three deferred-list builders keep only review-bot items, so a
maintainer comment can never appear as an 'audit record'.
- The deferral note says what is actually deferred (the automated
reviewer's non-Critical suggestions), states that maintainer
feedback is never deferred, and names the exit (@qwen-code /retry
opens a fresh counting window).
- SKILL.md's Critical-only policy now marks everything rendered in
the actionable sections as in scope, so the agent does not re-refuse
what the filter passed through.
Behavioral test updated both ways: maintainer comments/reviews stay
actionable in Critical-only mode across all three sources, bot
suggestions still defer, and structural pins hold the bypass and the
bot-only select in all six filters.
* feat(autofix): per-author feedback budget in Critical-only mode
Follow-up to the author-based split, prompted by the obvious
counterexample: a human account can host an automated reviewer loop
with the exact regeneration property the review bot has — feedback
re-generated after every push at zero marginal cost — so 'not the
bot' cannot mean 'never throttled'. An account is an accountability
unit, not a throttle; the brake has to key on measured regeneration.
Unified model: once Critical-only engages, every source has a bounded
budget of untagged feedback batches per counting window. The review
bot's budget is zero (all deferred, as before). A human's is
CRITICAL_ONLY_HUMAN_BATCHES (2) CONSUMED batches: feedback items are
bucketed into the (prev marker ts, marker ts] span that evaluated
them, only spans from Critical-only rounds count, and an author needs
K distinct consumed spans before their new untagged feedback defers.
Fresh unevaluated feedback never counts against its own author, and
the census is window-scoped, so /retry resets the budget with the
window.
The observed cases (#8037/#7944/#7885/#7799 — one or two late
verification reports each) stay fully served under K=2; a looped
reviewer is throttled after 5+K driven rounds instead of grinding to
the 100-round cap. Past the budget, continuing requires one conscious
act — **[Critical]**, a Request changes review, or /retry — which is
precisely what separates intent from automation. Over-budget authors
are named in the deferral note with those exact escapes.
Tests: the six filter replays gain over-budget cases both ways (the
tagged/CR escapes survive even over budget), and the budget census
itself is replayed over fixture files — two consumed critical-tail
batches list the author; one batch, pre-Critical batches, unconsumed
feedback, untrusted authors, and command comments never count.
* fix(autofix): fix deferred-feedback bash quoting and drop a dead jq binding (#8071)
* test(autofix): exercise census window-isolation guard with a stale-window fixture (#8071)
* test(autofix): make census command-exclusion observable; surface census stderr (#8071)
* fix(autofix): exclude never-deferrable feedback from the budget census (#8071)
The Critical-only per-author budget census counted every trusted review,
inline comment, and issue comment, including feedback the deferred renderer
would never defer: **[Critical]**-tagged comments, Request changes / APPROVED
reviews, inline replies rooted at a Critical comment, and inline comments
attached to a Request changes review. A maintainer who followed the documented
escape hatches (tag Critical, request changes) thereby spent their own budget
and had later untagged feedback silently deferred — the exact bug this PR
fixes, re-created one level down.
Mirror the three deferred-builder predicates in the census item filter so a
batch is counted only when it is actually deferrable. Extend the census replay
test with protected authors (Critical-only, Request changes, APPROVED,
Critical-rooted replies, Request-changes-review inlines, the review bot as a
trusted MEMBER, and a sentinel-ts marker probe) that each carry two
consumed-span batches yet must stay absent, so dropping any one exclusion now
fails the suite. Also fold bash's stderr into the bash -n guard assertion so a
future quoting regression reports the syntax error, not just a non-zero exit.
---------
Co-authored-by: verify <verify@local>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
* fix(autofix): answer cap-gate refusals on the PR instead of only in logs
Observed on #7836: the fleet shepherd detected a merge conflict, posted
'dispatched the autofix loop to resolve it', and the dispatch died at
the scan's round-cap gate with only a log line — the PR page showed a
promise, the run showed green, and the conflict sat unhandled for
hours. Three silences stacked: the standard-management cap itself is
silent (the pause notice was takeover-only, so #7836 hit 10/10 with
zero PR-visible notice), the forced-dispatch refusal is silent, and the
shepherd dedups per head SHA — a capped PR gets no pushes, so its head
never changes and conflict handling froze permanently.
Two scan-side changes (the shepherd stays untouched — the windowed
round computation lives in the scan and duplicating it would drift):
- A FORCED dispatch (shepherd conflict lever or a human) refused at the
cap gate now answers on the PR: cap value, what stays unhandled, and
the two recovery commands (/retry for a fresh window, /takeover for
the raised cap). No dedup — the shepherd sends at most one dispatch
per head, and a human asking twice deserves two answers.
- The cap pause notice covers ALL managed PRs: the takeover variant
keeps its wording, standard bot PRs get their own (/retry or
/takeover). Same marker, same once-per-window dedup, same consent
and PAT-identity checks — skip wins everywhere, and only the
takeover variant requires the label to still be present.
After a re-arm the next scheduled scan picks the PR up normally
(conflict targets are label-independent), so the frozen-head loop
resolves without any shepherd change.
* test(autofix): replay the cap-notice consent gate across label/takeover permutations (#8067)
* fix(autofix): gate the loud cap-refusal on workflow_dispatch (#8067)
FORCED_PR is populated for every trusted pull_request_review (route emits
pr_number for those), not just workflow_dispatch, so on a capped PR each
review submission landed in the un-deduped refusal branch — 7 "Dispatch
refused" comments on #7836 where 2 carried the information. Answer only
workflow_dispatch (the shepherd lever or a human); review submissions stay
covered by the once-per-window pause notice. Adds a verbatim behavioral
replay of the guard so a dropped EVENT_NAME condition fails the test.
---------
Co-authored-by: verify <verify@local>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com>
Post-merge measurement of #8016, on the first two live runs after it
landed (#7975 and #8066): the browser installed successfully — "Install
evidence browser: success" in both — and both reports carried **zero
images**. #7975 has 31 tables, #8066 has 407 assertions. Neither has a
witness.
The cause is the third instance of one structural bug, and it is the
same one twice already fixed:
#7917 the /verify recommendation sat inside a section headed
"local invocation ONLY" -> 1/16 reach
#8016 captures were "Optionally ... when text cannot carry the
oracle" -> 0/14
now captures are a required deliverable in the ARTIFACT CONTRACT,
but the plan the agent executes is the Scope-selection budget
list, and that list had zero mentions of evidence, png,
capture or screenshot
An agent that budgets by that list never allocates time for a capture,
however firmly a later section says to produce one. So captures are now
budget item 4 alongside the A/B, the harnesses and the gates, with the
time reserved (~5 min), the gate named (QWEN_VERIFY_CHROMIUM=1), the
count bounded (normally two, at most a handful), and the measured
failure stated so the rule carries its own reason.
The report structure gets the matching half: the A/B table item now says
to reference the capture of those cells by filename. A produced image
with nowhere to go is as dead as an unproduced one.
Mutation-verified 4/4: dropping the budget item, the time allowance, the
count bound, or the report-side reference each turns the test red.
Not changed, and worth stating rather than implying: the seven techniques
from #8010 do not appear in either report. I do not read that as failure
— every one of them is conditional ("when one fix bundles two changes",
"when the oracle is an instrument", "when a PR adds a writer to a shared
store") and neither PR met those conditions. #8010 is not yet falsifiable
from these runs. What both reports DO show is older skill rules working:
a naive-matcher differential in #8066, and #7975's realpath check finding
node_modules/@qwen-code/qwen-code-core resolving into the HEAD tree and
then reasoning about whether that contaminates the control.
107/107 tests; prettier and eslint clean.
Co-authored-by: wenshao <wenshao@example.com>
Pass the auth token to expectSupervisorUnreachable so it distinguishes a
live auth-gated supervisor from a closed socket, remove SIGTERM/SIGINT
listeners on the non-signal shutdown path, exercise prepareSocketPath's
mkdir via a nested socket path in the permission test, and add coverage
for the 1 MB request-line size guard.
Measured on a congested pool: 88 active jobs, 72 hosted and 15
self-hosted. The self-hosted 15 were all running with zero queued; the
hosted 72 were contending, and 20 of them were the SAME job —
`remove-suspicious-attachments`, all queued, none running.
Its real cost is not the work. Recent completed runs:
queue=629s run=5s queue=568s run=2s
queue=518s run=2s queue=340s run=3s
Two to five seconds of API calls behind up to ten minutes of queueing.
And almost none of it needed to happen. The trust check lived INSIDE the
github-script, so a runner was queued, allocated and started before the
job could decide it had nothing to do. Over the 200 most recent comments
on this repo: 184 from trusted associations, 9 from bots, 7 actually
needing a scan. 96.5% of these runs existed to print "Trusted author;
skipping".
Two changes:
- Hoist the association and bot checks into the job `if:`. GitHub
evaluates `if:` BEFORE allocating a runner, so a trusted comment now
costs nothing. The script keeps its own copies: the gate is an
optimisation, not the control, and the two must be able to disagree
without becoming unsafe. Every ambiguity therefore resolves toward
RUNNING the scan — an unrecognised payload yields an empty
association, which is not in the trusted list, so the job runs.
- Add a per-comment concurrency group with cancel-in-progress. The
workflow listens on `edited` as well as `created`, and the bot PATCHes
its own comments constantly, so repeated edits of one comment stacked.
The scan reads the comment's CURRENT body, so a queued earlier scan is
already stale and cancelling it loses nothing. (Contrast the verify
lane, where cancel-in-progress is deliberately false because a
cancelled run destroys evidence.) The key falls back to run_id so an
unexpected payload gets its own group instead of serialising every
scan into one.
Deliberately NOT moved to the self-hosted pool, though it would fit
technically (no checkout, no PR code, API calls only): the 20 stacked
jobs were duplicates, so relocating them just fills the ECS pool
instead — and that pool is what /verify and /triage depend on. It also
holds issues:write while processing untrusted comment bodies, which
belongs on ephemeral hardware rather than reused machines.
The `if:` semantics are verified against all payload shapes — 12 cases
covering both `comment.*` and `review.*` associations, bots, and
missing/empty payloads, each asserting which direction it resolves.
CONTRIBUTOR is deliberately NOT trusted: a merged PR does not make
someone's links safe.
Mutation-verified 6/6: dropping the review payload path, dropping the
bot check, adding CONTRIBUTOR to the trusted list, turning off
cancel-in-progress, collapsing the group to a global key, and inverting
the gate so untrusted comments are the ones skipped — each turns a test
red. The last is the one that matters; it is the only mutation here that
would be a security regression rather than a cost regression.
148/148 tests across both suites; actionlint exit 0; prettier and eslint
clean.
Co-authored-by: wenshao <wenshao@example.com>
* feat(triage): make /verify evidence screenshots actually possible
Measured across the 14 real verify reports on this repo: tables appear
in 14/14 (4 to 31 rows), images in 0/14. The hosting machinery has been
complete the whole time — pr-assets branch, PNG magic check, 8 images,
2 MB cap, sanitised names — and it has never once been fed.
Two independent causes, both fixed:
1. The agent physically could not produce a PNG. It runs as `node`
under `env -i` with HOME pointed at a per-run scratch dir, so
`playwright install --with-deps chromium` needs an apt it has no
rights to and would re-download ~170 MB into a directory deleted at
the end of every run. Chromium is now installed by ROOT in the tools
step into a shared, world-readable PLAYWRIGHT_BROWSERS_PATH, and the
agent env carries that path plus QWEN_VERIFY_CHROMIUM=1 — but ONLY
when the install actually succeeded, so the variable's absence is a
real signal rather than a stale promise. The install is best-effort:
a failure warns and the run continues to a text-only report, which
is what every run has produced anyway.
2. The skill discouraged it. Captures were "Optionally … use them when
text cannot carry the oracle", with TUI rendering as the only
example — so an agent verifying a daemon or a CLI correctly
concluded text carried the oracle and skipped. It now says produce
them whenever a harness ran, and says WHY: a table is the agent's
claim about what happened; a capture of the run is a witness that
the numbers came from a real execution, which is the part a reviewer
cannot get any other way. It names the three highest-value shots
(A/B cells side by side, the mutation matrix as printed, raw harness
output behind a headline number) and forbids `playwright install`
explicitly, so the failure mode of cause 1 cannot be rediscovered by
an agent burning budget on it.
Affordable now because #8014 raised the agent budget to 120m; at 20m a
browser download alone was most of the run.
Mutation-verified 4/4: handing the agent the variable unconditionally,
failing the job on a chromium install failure, reverting the skill to
"Optionally", and dropping the do-not-install warning each turn the
test red.
90/90 tests; prettier, eslint, actionlint and shellcheck clean.
* fix(triage): match browser to lockfile and fix pr-assets branch conflict (#8016)
Split the chromium install into system deps (tools step, as root) and
browser binary (new post-checkout step, as node using the checkout's
own Playwright). This eliminates the version pin that produced an
unusable browser tree.
Change evidence hosting from a bare pr-assets leaf branch — which
cannot coexist with the existing pr-assets/* namespace — to per-PR
branches (pr-assets/pr<N>-verify) with orphan-init fallback.
Update terminal-capture skill to warn CI agents away from the
forbidden playwright install command, and close test blind spots
for conditionality (M1b), version matching (M5), and marker
ordering (M6).
* fix(triage): unpin Playwright install-deps so apt list tracks lockfile browser (#8016)
* fix(triage): assert the browser path reaches the agent, not just the flag
Review finding on #8016, verified by mutation before accepting it:
deleting `"PLAYWRIGHT_BROWSERS_PATH=$CHROMIUM_PATH"` from the agent env
left the test green. That is the nastiest arm of this feature — the
agent is TOLD chromium is available via QWEN_VERIFY_CHROMIUM=1, then
Playwright looks in the default ~/.cache/ms-playwright instead of the
shared install and every capture fails, so a successful install still
degrades to a text-only report.
The test now asserts both variables and that BOTH sit behind the
success guard. Mutation-verified 2/2: deleting the path line fails on
the containment assertion, and hoisting it outside the guard fails on
the ordering one (`expected 21928 to be less than 21834`).
Not changed, because the other two findings do not describe this
codebase — replies on the threads carry the evidence.
90/90 tests; prettier and eslint clean.
* fix(triage): cover orphan-init delivery and delete verify asset branches
Two review findings on #8016, both about the per-PR branch scheme the
autofix bot introduced on this branch (029beba3ef), and both correct.
**Orphan-init's success path had no test.** The existing scenarios both
take the clone-failed branch, but one seeded the branch (so clone
succeeds) and the other points at a non-repo (so the push fails too) —
neither proves orphan-init can DELIVER. A third scenario now points at
a valid bare remote with the branch deleted first: clone fails,
orphan-init runs, and the images land. It also asserts the branch has
exactly one commit, so a graft onto unrelated history would fail rather
than pass as "orphan". Without it, a dropped `git init` or
`remote add origin` would silently discard every image on every PR's
first verification run.
Mutation-verified: dropping `remote add origin` and dropping `git init`
each kill it. A third mutation — pointing `checkout --orphan` at the
wrong branch name — SURVIVES, and correctly: the push is
`HEAD:$assets_branch`, which names its destination explicitly, so the
local branch name cannot affect delivery. Recorded as a non-defect
rather than papered over with a test that pins an irrelevant detail.
**Per-PR asset branches were never deleted.** `pr-assets/*` had exactly
one producer and one deleter; the verify lane became a second producer
without being added, so every verified PR would leave a single-commit
branch in the base repo permanently. The cleanup workflow now deletes
both refs, is renamed to match what it does, and carries a note that
every new `pr-assets/*` producer must be added — a branch nothing
deletes is permanent.
The loop deliberately drops `set -e`: most PRs produce neither branch,
so a 404 on the first must not skip the second. A real delete failure
still warns and sets a non-zero exit. The new test executes the script
against a stubbed `gh` where the visuals branch 404s, and asserts the
verify branch is still deleted — which a `set -e` version would fail.
92/92 tests; prettier and eslint clean.
* fix(triage): match browser install to the harness Playwright and re-arm the strip guard (#8016)
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(triage): gate chromium marker on deps and resolve Playwright from the harness (#8016)
Address review feedback on the verify evidence-browser path:
- Resolve the Playwright CLI from the capture harness's own directory
(require.resolve of the exported package.json, then join cli.js) instead
of assuming npm hoists playwright to the root, so the installed binary
tracks the harness's import even if the hoist layout changes. cli.js is
not in the package's exports map, which is why a direct path was used
before; resolving the exported package.json keeps that bypass while
removing the hoist assumption.
- Gate the verify-chromium-path marker on a new verify-chromium-deps-ok
marker written by the tools step. apt and the Playwright CDN are
independent servers with no shared success signal, so a binary download
alone must not promise chromium to the agent when system deps failed.
- Add a delete-failure case to the PR asset cleanup test so the status=1
and exit path is executed, not just statically asserted.
* fix(triage): strip cache creds from browser install and guard cli.js resolution (#8016)
The evidence-browser install runs the PR-resolved Playwright CLI but did
not drop ACTIONS_RUNTIME_TOKEN/URL/CACHE_URL the way the adjacent prepare
and agent steps do, leaving runner-injected cache credentials readable by
PR-controlled code. Strip them in the same env -u chain. Also replace the
string-only guard on the require.resolve + cli.js join with an execution
guard that resolves the expression against the installed tree, so a
Playwright bump that relocates cli.js fails the test instead of silently
degrading evidence capture at runtime.
---------
Co-authored-by: wenshao <wenshao@example.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com>
Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com>
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
* feat(autofix): back off scan inspection of idle candidates
The scheduled scan inspects every candidate every tick. The takeover
pool doubled in two days (28 open takeover PRs, 8 of them idle in
'nothing new' state for 10+ hours), and idle candidates crowd the two
SHARED budgets: MAX_CANDIDATE_INSPECTIONS (60) and the 10-target cap.
Observed on #8002: freshly engaged, admitted by the 09:03 scan, then
deferred by the target budget while long-idle PRs re-confirmed their
idleness yet again.
Candidates whose list-provided updatedAt (no extra API call) is older
than 24h are now inspected on roughly every 4th scan, on a
deterministic slot keyed by PR number and UTC hour so no PR waits
forever. The skip is free — it sits with the busy skip before the
inspection-budget increment.
Safe by construction: every real wake-up bumps updatedAt (reviews,
comments, labels, pushes) or routes in real time anyway, so the only
thing deferred is the scheduled re-confirmation of idleness plus
worst-case a few hours of base-conflict-detection latency for a PR
nobody touched in a day. The forced-dispatch path never builds the
list files, so a forced PR is always inspected.
* feat(autofix): idle-backoff review follow-ups
- Corrected the comment's cost model: idle candidates hit 'continue'
before the TARGETS append, so they never contend for the 10-target
cap — the real win is the shared inspection budget plus the serial
scan-walk latency (the walk is what delayed #8002's pickup by ~6
minutes), and the comment now says exactly that.
- Slot quantum changed from the hour to the scan tick (600s, the same
quantum as ROT_OFF): an hourly slot against the */10 cron meant 6
back-to-back inspections then a ~3h blind window per PR — same 25%
average, terrible shape. The gap is now bounded at ~30 minutes, which
is what the operator-facing strings promise ('gap ≤30m').
- The two scan-only signals updatedAt cannot see (a base conflict
appearing when main moves; still-red checks awaiting the redcheck
marker) are named in the comment instead of papered over.
- The per-candidate jq fork became a single precomputed set + a bash
substring test, matching the busy skip's idiom and the 'free' claim.
- Tests: the skip predicate and set builder got a behavioral replay
(idle+out-of-slot defers, idle+in-slot inspects, fresh inspects,
missing-from-lookup inspects); the two byte-distance assertions
became a loop-head slice (comment growth cannot red-light CI, and
budget-consuming code between the skips and the increment fails);
the --json field pin is order-independent; the 3600 quantum is
pinned OUT.
* test(autofix): pin the null-updatedAt defensive guard in idle-backoff replay (#8049)
* fix(autofix): extract idle-backoff predicate from workflow, fix gap bound 30→40m (#8049)
* test(autofix): pin takeover-prs.json in idle-backoff replay (#8049)
* fix(autofix): state idle-backoff gap probabilistically, unpin quantum (#8049)
Round-2 verification showed the scheduled scan lands every ~40-70 min on
this repo, not every 10 min, so the (epoch/600)%4 slot is an independent
~25% draw per scan, not a deterministic 1-in-4 rotation. The gap is
geometric (measured median ~2h, p90 ~6h), not bounded at ~40m.
Reword the operator-facing strings (skip echo, fleet row) and the workflow
comment to state the behaviour probabilistically and drop the false
"bounded" / "no PR is unlucky forever" claims; correct the misattributed
#8002 latency (queue/startup, not the serial walk). Relax the tests so a
truthful number is not a CI failure: pin the mod-4 time-quantum shape
instead of the exact 600s constant, and drop the /3600 exclusion that
forbade the better-tailed quantum. Mechanism logic is unchanged.
---------
Co-authored-by: verify <verify@local>
Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com>
* fix(integration): make interactive read-then-write test deterministic (#8060)
Drive the multi-turn conversation with the fake OpenAI server instead of a
live LLM. The real model made this test flaky on main: it could choose
different tools, phrase the read result without the literal version, or
settle the first turn on its own schedule. Scripting the exact read-then-write
turns keeps the interactive mechanics (typed input, tool execution, file
mutation) under test while removing the nondeterminism, matching the
fake-server pattern already used by the other multi-turn interactive tests.
* fix(integration): add debug output to read-step assertions (#8060)
* test(integration): clarify scripted-echo assertion and verify request count (#8060)
* test(integration): ground version assertion in the read result (#8060)
Address review feedback: assert the read_file tool result the CLI sent
back (requests[1]) contains '1.0.0', so the version check observes the
real read result rather than only the fake model's scripted echo. Also
correct the flake rationale comment to cite the verified mid-stream
stall on turn 2 instead of an unsupported late-settle hypothesis.
The main agent runs in yolo mode with read_file registered, so the model sometimes read the file directly instead of delegating, failing the foundTaskTool assertion. Make the prompt require delegation and forbid direct reads, and make the subagent imperative about calling read_file. Tool restriction (coreTools/excludeTools) is not viable: the subagent shares the session permissionManager, so it would lose read_file too.
Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com>
