* fix(core): strip system-reminder pollution from session title and recap prompts
The `filterToDialog` function in sessionTitle.ts and sessionRecap.ts
was including system-reminder blocks (skills list, CLAUDE.md, MCP
announcements) in the prompt sent to the title/recap model. When the
user's first message was short, `flattenToTail` would reach back into
these injected entries, causing titles like "resolve-cr-comments" instead
of reflecting the actual conversation.
- Skip startup prelude entries via `getStartupContextLength`
- Strip `<system-reminder>` blocks from text parts
- Add shared `stripSystemReminderBlocks` helper in environmentContext.ts
- Add regression tests for both sessionTitle and sessionRecap
Closes#6419
* fix(core): preserve mixed reminder prompt turns
* fix(core): support large text range reads
Allow text reads to stream bounded line ranges for files larger than the previous 10MB guard, while preserving media size limits and forwarding cancellation through read_file/read_many_files/ACP paths.
Refs #6403
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): address large text review feedback
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): propagate abort signals in text reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): validate streamed utf8 reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): handle disabled line truncation for large reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): use kebab-case for text range reader
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): preserve artifact size errors for large sources
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): address large text review follow-up
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): allow default large text reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): honor text read byte caps
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): clarify invalid utf8 range read errors
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): prevent truncated full large reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): preserve unbounded line-zero reads
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): forward artifact read cancellation
Pass artifact execution abort signals into source file reads and preserve cancellation semantics when the read is aborted.
Add regression coverage for unbounded large UTF-8 range reads and offsets beyond EOF.
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): address file read review feedback
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): preserve large text mutation reads
Allow default unbounded readTextFile calls to keep reading full large text files so mutation tools can prepare complete snapshots after a prior ranged read.
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.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>
Split-view panes now render a compact streaming status: the spinner, elapsed time, token count, and cancel hint stay, but the rotating "witty" loading phrase is suppressed. StreamingStatus gains a showPhrase prop (default true, so the main chat is unchanged); when false it also skips the phrase-rotation timer so each pane avoids a needless interval.
Place the Qwen brand mark to the left of the sidebar's New chat button.
The artwork is the same SVG used for the browser-tab favicon (and the
QwenLM GitHub avatar), inlined rather than hot-linked because the Web
Shell CSP is `img-src 'self' data: blob:`, which blocks remote images.
When the sidebar is collapsed there is no room beside the compact
button, so the mark is hidden and only the New chat button remains.
* fix(web-shell): keep split-view session list fresh and preserve panes across view switches
The in-window split view's "add pane" picker read a stale session snapshot —
`useSessions` only fetches on mount — so sessions created after entering the
split never appeared. And switching away from the split and back cleared the
panes, because the live pane set lived in local state that died on unmount while
the seed it re-mounted from was never updated (and the no-arg "Open Split View"
button reset it to empty).
- Reload the picker list when it opens and when the parent's session-list reload
token changes, so it never offers a removed session or misses a new one.
- Mirror the live pane set up to the app via onPanesChange so it survives
SplitView unmounting; restore it (instead of reseeding empty) when the split is
reopened without an explicit selection.
* test(web-shell): cover split-view refresh/restore per review; coalesce token reloads
Addresses review feedback on #6418:
- SplitView: skip a token-driven reload while one is already in flight, so a
burst of session-list changes (bulk create/delete) doesn't fire a redundant
concurrent round-trip per bump (matches the sidebar's poll guard).
- SplitView test: the freshness test now proves the picker re-renders with the
refreshed list — a session appearing only after reload shows up — not just
that reload() was called.
- App test: cover the openSplitView preserve/restore path end-to-end — a reported
pane set survives leaving the split and is restored on reopen.
* fix(web-shell): reload split picker on every token bump (drop in-flight guard)
The in-flight guard added in the previous commit could drop a session-list
reload token that arrives while a reload is still running: the effect has
already run for that token value, and clearing the in-flight flag in `finally`
doesn't re-run it, so the picker could stay stale after burst create/delete/
rename activity — and the split has no polling fallback to recover.
Reload on every distinct token bump instead. `useDaemonResource` serializes
responses via its sequence counter (last write wins), so overlapping reloads are
correct, and the token is bumped only on discrete session-change events — an
occasional redundant fetch is far cheaper than a lost refresh.
* test(web-shell): cover openSplitView explicit-selection branch (dedupe + cap)
Per review: the restore branch of openSplitView was covered but the
explicit-selection branch (dedupe + MAX_SPLIT_PANES cap, replacing any prior
set) was only exercised, not asserted. Add a `?split=` URL test with duplicate
and over-cap ids that asserts the split seeds exactly the deduped, capped
selection.
When a skill is invoked multiple times in a session, each invocation
previously appended the full SKILL.md body content to the conversation
history as a new tool result, wasting context tokens.
Add an isSkillLoaded callback to SkillToolInvocation that checks
loadedSkillNames before building the full content. On re-invocation,
return a short confirmation message instead of the full body. The
check runs after successful skill load (so disabled/not-found paths
are unaffected) but before content construction, hooks registration,
and allowedTools application (which are idempotent and already applied
on first load).
Fixes#6427
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* feat(core): add tools.visible config for selective deferred-tool visibility
* fix(cli): wire tools.visible from settings.json into Config
Add settingsSchema entry for tools.visible and plumb it through loadCliConfig into ConfigParameters.visibleTools. Without this the core-level visibleTools support was unreachable from settings.json.
Adds 3 CLI-level tests: visibleTools passthrough, empty default, safe-mode suppression.
Fixes#6368
* fix(core): exclude visibleTools from tool_search candidates and reveal path
Add visibleTools gate to collectCandidates() and loadAndReturnSchemas in tool-search.ts to prevent KV-cache invalidation when tool_search is invoked for a visible-deferred tool.
Without this, a tool listed in tools.visible would still appear as a keyword-search candidate and select: would still trigger revealDeferredTool + setTools, defeating the purpose of promoting it to first-class visibility.
3 tests: keyword-search exclusion, select: no-reveal/no-setTools, and select: still works for non-visible deferred tools.
* fix(cli): include visibleTools in /context per-tool token breakdown
Add config.getVisibleTools().has(tool.name) gate to the deferred-tool skip condition in collectContextData. Without this, visible-deferred tools appear in the headline total (via getFunctionDeclarations()) but are excluded from the per-tool breakdown, causing the sum to mismatch.
1 test: visibleTools included in breakdown despite deferred+unrevealed.
* fix: address all 7 review suggestions for tools.visible
1. settingsSchema: user-facing description instead of internal jargon
2. config.ts: use normalizeToolNameList (generic name) for both disabled and visible
3. config.test.ts: add bare-mode exclusion test
4. tool-registry.test.ts: disabledTools > visibleTools priority test
5. tool-registry.ts: update JSDoc for getDeferredToolSummary
6. tool-search.test.ts: mixed select: visible+non-visible test
7. tool-registry.test.ts: visible survives clearRevealedDeferredTools
* chore: regenerate settings.schema.json after description update
CI check detected that settings.schema.json was out of sync with settingsSchema.ts after the description was changed in commit 73e879882.
* fix: address wenshao review — extract isDeferredAndHidden, clean up abstractions
1. Remove normalizeToolNameList (empty wrapper, violates AGENTS.md no-abstraction rule)
2. Extract ToolRegistry.isDeferredAndHidden() — 5 call sites reduced to 1 predicate source
3. Add dirty-input test for tools.visible (whitespace, duplicates, empty strings)
4. Add MergeStrategy.UNION test for tools.visible across user + workspace scopes
---------
Co-authored-by: Aleks-0 <aleks-0@users.noreply.github.com>
* feat(core): add Tool(param:value) permission syntax for parameter-level access control
Introduces key:value parameter matching in permission rules, allowing
users to grant or deny tool access based on specific input parameters.
- Parse key:value pairs from specifiers for literal-kind rules
- Support wildcard patterns (*), multiple params, and mixed syntax
- Thread toolParams through PermissionCheckContext and matchesRule
- Add 11 unit tests covering parsing, matching, wildcards, and edge cases
- Maintain backward compatibility with existing specifier kinds
Example rules:
Agent(model:opus) # deny agents using Opus model
Agent(coder,model:*) # deny coder-type agents with any model
Bash(git:*) # still works (legacy :* → git *)
Closes#6100
* fix(permissions): resolve PR #6106 review comments for tool param permission syntax
- buildPermissionRules now propagates toolParamMatchers for 'Always Allow' flow
- MCP tool rules now check param matchers after name matching
- Added Object.hasOwn check to prevent prototype chain lookup vulnerability
- Replaced matchesCommandPattern with matchesParamValuePattern for param value matching
- Added diagnostic logging for param matching failures
- Added warnings for empty valuePattern, invalid keys, and non-literal key:value syntax
- Added type checking for non-primitive param values
* fix(core): address critical review feedback on Tool(param:value) permission syntax
- Add 's' flag to RegExp in matchesParamValuePattern for multiline support
- Filter buildPermissionRules to stable param keys only (model, subagent_type, skill, server_name) to prevent sensitive data leakage
- Reject MCP rules with unsupported specifiers instead of silently ignoring
- Fix :* wildcard conversion to use global replace for backward compatibility
- Extract shared evaluateParamMatchers helper to deduplicate MCP and standard branches
* fix(permissions): address remaining review feedback for Tool(param:value) syntax
- Remove unused @ts-expect-error in gitWorktreeService.ts (CI blocker)
- Fix ReDoS in matchesParamValuePattern: replace regex with linear-time
glob matcher using indexOf, avoiding catastrophic backtracking on
multi-wildcard patterns like *a*a*a*a*b
- Fix MCP backward compatibility: exclude MCP tools from key:value parsing
in parseRule and buildPermissionRules to preserve existing MCP deny
rule semantics
- Add tests for MCP + param matcher, partial wildcards, ReDoS prevention,
number coercion, and buildPermissionRules with toolParams (stable params,
volatile params, sensitive data, round-trip)
* fix(permissions): address PR #6106 review comments and fix useStatusLine test timeout
- Make matchesParamValuePattern case-insensitive to match matchesDomainPattern convention
- Remove duplicate JSDoc block before matchesParamValuePattern
- Remove dead server_name from stableParamKeys in buildPermissionRules
- Add PermissionManager integration tests with toolParams (evaluate, findMatchingDenyRule, hasRelevantRules, hasMatchingAskRule)
- Add type guard tests for evaluateParamMatchers (null, undefined, boolean, object)
- Fix useStatusLine.test.ts timeout by stubbing cron-task exports in core mock
---------
Co-authored-by: 易良 <1204183885@qq.com>
Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com>
* feat(cli): support stacked slash-skill invocations (#6355)
Allow users to chain multiple slash-skill commands in a single prompt
(e.g. `/feat-dev /e2e-testing implement X`). The skill bodies are
concatenated with the remaining user text and submitted as one
`submit_prompt` to the model, so the model receives all loaded skill
contexts at once.
- Add `parseStackedSlashCommands()` with a `MAX_STACKED_SKILLS = 5` cap
- Integrate stacked detection into both interactive (slashCommandProcessor)
and non-interactive dispatch paths
- Record `recordSkillInvocation` telemetry for each stacked skill
- Emit a warning when more than 5 skills are requested
- 29 new test cases covering parsing edge cases and both dispatch paths
Closes#6355
* fix(cli): address review feedback for stacked skill invocations
- Fix whitespace tokenization to match all \s chars, not just spaces
- Align telemetry recording: record success based on actual result type
(both dispatch paths now consistent)
- Surface error messages from non-submit_prompt skill results
- Propagate modelOverride from first submit_prompt skill
- Add 7 new tests: tab whitespace, mixed whitespace, non-submit_prompt
exclusion, telemetry accuracy, modelOverride propagation
* fix(acp-bridge): use static import in logRedaction test to avoid timeout
The dynamic `await import('./spawnChannel.js')` inside the test body
was pulling in a heavy module graph at runtime. Under CI contention
(667 tests running concurrently), this exceeded the 5-second default
timeout. Convert to a static top-level import — the same pattern used
by spawnChannel.test.ts.
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(cli): move stacked skill block inside try/finally and collect onComplete callbacks
- Move stacked skill dispatch into try block so finally cleanup runs
(setIsProcessing, chat recording, telemetry) preventing TUI freeze
- Set invocationSentToModel=true for stacked invocations so chat
history correctly classifies the command as sent to model
- Collect and forward onComplete callbacks from all submit_prompt
skill results in both interactive and non-interactive paths
---------
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* feat(scheduled-tasks): run each task in its own dedicated, named session
Scheduled tasks created through the Web Shell management page were never firing
in the daemon-only case: the durable-cron tick runs inside an active agent
session, and the Web Shell creates a session only lazily on the first prompt, so
a task created on the management page (with no chat open) had nothing ticking it.
This binds every management-page task to a dedicated session, minted at create
time and named "⏰ <task>". The task fires ONLY inside that session — its
transcript is the task's run history — instead of via the shared per-project
durable owner. A daemon-side keepalive heartbeats those sessions so the idle
reaper doesn't stop them, and a boot-time rehydration reloads them after a
restart. Archiving, deleting, or unarchiving the session disables, removes, or
re-enables the bound task (covered on both the REST and ACP surfaces).
Also adds task editing, a live next-run countdown, run history, a one-per-row
card layout, and a "run now" that executes in the task's bound session and
updates the last-run time. All resident-session management is opt-in and enabled
only by the real daemon (runQwenServe), so createServeApp embeds/tests are
unaffected.
* fix(scheduled-tasks): address code review on per-session task feature
Review fixes for #6389:
- Distinguish archive-disabled from user-disabled tasks: disableTasksForSessions
now marks disabledByArchive; enableTasksForSessions only re-enables tasks
carrying that flag, so a task the user deliberately disabled stays disabled
across an archive/unarchive cycle. [Critical]
- Rehydrate task sessions concurrently with a per-session 30s timeout so one
hung loadSession can't stall the boot sweep or leave healthy tasks dormant.
[Critical]
- Await runScheduledTask + reload before executing the prompt in handleRunNow,
so a record failure surfaces and the card's "last run" reflects the trigger.
[Critical]
- Log keepalive/rehydrate read + heartbeat failures at debug instead of
swallowing them silently, so a persistently-failing keepalive is diagnosable.
[Critical]
- Add integration tests: deleteDaemonSessions -> removeTasksForSessions and
unarchiveDaemonSessions -> enableTasksForSessions (guard the coupling). [Critical]
- DELETE route: single atomic updateCronTasks that captures the bound session
and removes the task in one cycle, closing the read-then-remove TOCTOU.
- Stop the keepalive timer during shutdown (matters for embedders that don't
process.exit) so it can't fire against a disposed bridge.
- Deduplicate DEFAULT_BUILDER: export it once from scheduledTasksSchedule and
drop the dialog's copy so the create form and cron-reversal can't drift.
- Reject empty-string sessionId in isValidTask: a bound task with "" would
silently run unbound under the scheduler's truthy guard.
* fix(scheduled-tasks): isolate task sessions, fix catch-up/jitter/revive
Second review round (#6389):
- Force `sessionScope: 'thread'` when minting a task's session. The daemon's
default scope is 'single', which attaches to (reuses) the shared workspace
session — so a second task, or a task alongside an open chat, would bind to
the same session, rename it, land runs in the wrong transcript, and close it
on delete. Thread scope guarantees each task an isolated session. [Critical]
- Re-seat a recurring task's schedule anchor to now when a PATCH changes its
cron (or flips one-shot→recurring), not just on re-enable. A bound task's
catch-up runs on every file-watch reload, so a bare cron edit to an
expression with an already-past slot would fire immediately on save. [Critical]
- Revive a non-resident bound session from the keepalive when its heartbeat
fails (reaper let it go while disabled/archived, now re-enabled). Covers the
unarchive and PATCH false→true paths uniformly and retries each interval, so
a re-enabled task actually resumes instead of showing a live countdown that
never fires. Best-effort, timeout-bounded, non-blocking. [Critical]
- Report `nextRunAt` using the scheduler's jittered fire time
(`nextDurableFireMs`) instead of the bare cron boundary, so the UI countdown
lines up with the real fire (the tick offsets each fire by up to the jitter
window) rather than expiring early and advancing prematurely.
All four are mutation-verified. The cross-daemon double-fire on bound tasks
(same session live in two schedulers) is a separate, architecturally-invasive
fix (claim-then-fire on the durable file) tracked as a follow-up.
* fix(scheduled-tasks): sync bound session name on task rename
Create names a task's session after the task (`⏰ <name>`), but a later PATCH
that renamed the task (or edited the prompt of an unnamed task) left the
session's display name stale. The PATCH route now re-applies
`updateSessionMetadata` with the task's effective label whenever that label
actually changes — a bare cron/enabled edit does not touch the session.
Best-effort: a metadata failure doesn't fail the committed schedule change.
Mutation-verified.
* fix(scheduled-tasks): mirror run sessionId on client type; clarify server wiring
Review follow-up (#6389, qqqys):
- [Medium] `DaemonScheduledTaskRun` now mirrors the daemon's `CronTaskRun`
`sessionId?: string`, so run-attribution the wire already sends isn't silently
dropped by the client type (not surfaced in the UI yet; passthrough cast means
no mapping change needed).
- [Nit] Comment the `app.locals.stopScheduledTaskKeepalive` set site, noting it
follows the same convention as `fsFactory`/`boundWorkspace`/`acpHandle` and is
read by the run-qwen-serve shutdown path (kept the convention rather than
diverge to a one-off return value / declaration merge).
- [Nit] Comment the outer `.catch(() => {})` on rehydrate as intentional
defense-in-depth (the function already handles read + per-session failures).
* fix(scheduled-tasks): couple archive/enable + record manual run only on enqueue
Two [Critical] review items (#6389, gpt-5-codex):
- PATCH re-enable coupling: reject `enabled: true` on a task disabled BY
archiving its session (`disabledByArchive`) with 409 `task_session_archived`.
Re-enabling it here would show an enabled task with a countdown while its
bound session stays archived and can never fire — the caller must unarchive
the session (which clears the marker and reloads it). A user-disabled task
(no marker) and non-enable edits are unaffected.
- Manual "run now" ordering: record the run only AFTER the prompt is enqueued,
not before. `runTaskManually` now returns a promise that resolves on enqueue
and rejects if the bound session can't be opened (archived/deleted), is
superseded, or times out; the dialog awaits it before writing
/scheduled-tasks/:id/run, so a failed session switch no longer leaves a
phantom run in history. Runs are serialized (one pending at a time, button
disabled) so two quick clicks can't drop a prompt on the single bound-run
latch. Added coverage for failed session load and double-click; all new
tests mutation-verified.
* fix(scheduled-tasks): close dormancy/orphan/overflow gaps from review
Five items from GPT-5 /review (#6389):
- [Critical] Bind tasks to sessions only when resident management is on:
createServeApp now passes the bridge to the scheduled-task routes only when
`manageScheduledTaskSessions` is set. Embedders that leave it off get UNBOUND
tasks (shared-owner firing) instead of bound tasks nothing keeps resident or
reloads (which would silently go dormant).
- [Critical] Keep the keepalive/revive loop running whenever task sessions are
managed, not only when a reaper is active — archiving closes a task session,
so a re-enabled one still needs reviving with the reaper disabled. Size the
interval under the reaper window (≤ half of it) so a small idle timeout can't
let a session be reaped before its first heartbeat.
- [Critical] Record a manual run only after the prompt is admitted: the bound
run latch now resolves only if `sendPrompt` admitted the prompt and rejects on
cancellation (e.g. onSubmitBefore) / failure, so a cancelled Run now no longer
advances lastFiredAt or appends history.
- [Critical] Clamp the dialog's reload timer to the 32-bit setTimeout ceiling
(~24.8 days) so a months-away schedule can't overflow and spin a reload loop.
- [Suggestion] Pre-check the task cap before spawning a session, so an over-cap
create never mints an orphan task session it must roll back.
New tests (route unbound-when-no-bridge, cap-no-spawn, computeKeepaliveIntervalMs
bounds, far-future timer clamp) mutation-verified; full server suite green.
* fix(scheduled-tasks): guard catch-up double-fire, run-now hang, /run + cron edits
Four items from GPT-5 /review (#6389):
- [Critical] Bound-task catch-up could double-fire: detection ran on every
file-watch reload and read the stale on-disk lastFiredAt, so a reload racing
the async catch-up persist (a foreign write to the tasks file) re-detected and
re-fired the same overdue slot. Track ids whose catch-up was DELIVERED but not
yet persisted (`deliveredCatchUp`) and skip re-detecting them until the write
lands; a merely-buffered-then-dropped catch-up isn't tracked, so it still
re-detects from disk (recovery preserved).
- [Critical] "Run now" hung the full 30s switch timeout when the bound session
was ALREADY the current, loaded one (no dep change → the consuming effect
never re-ran). Fire the enqueue directly after loadSidebarSession resolves as
well as from the effect; whoever runs first nulls the latch, so it runs once.
- [Critical] POST /run recorded a run with no enabled/disabledByArchive guard,
unlike PATCH — a direct API caller could write a phantom "ran" record onto a
paused/archived task. Return 409 task_disabled for a disabled task.
- [Suggestion] Anchor re-seat on cron edit compared the raw string, so a
cosmetic change (`0 9 * * *` → `00 9 * * *`) dropped a pending catch-up.
Compare the canonical (parsed) schedule instead.
(The setTimeout-overflow and keepalive-floor reports were already fixed in
2a12cba.) New tests for the first three + the cosmetic-cron case are
mutation-verified; full core scheduler + route suites green.
* fix(scheduled-tasks): block disabled-task run in UI; record manual run at admission
Two [Critical] review follow-ups (#6389):
- A disabled task could still EXECUTE from the Web Shell: the Run button was
only gated on `runningTaskId`, so clicking it enqueued the prompt and the
server's `/run` `task_disabled` guard merely refused the later history write —
a real, unrecorded run. Gate `handleRunNow` and disable the button on
`!task.enabled` too, so a disabled task's prompt is never enqueued.
- Manual run recorded only after the whole turn: the bound-run latch resolved
via sendPrompt, which completes through waitForAcceptedPromptCompletion, so a
long/permission-blocked run or a closed tab could execute without ever being
recorded. Add an `onAdmitted` callback to sendPrompt (fired when the daemon
accepts the prompt, before the turn) and resolve the manual-run latch at
admission instead — cancellation before admission still rejects.
New dialog test (disabled task → no enqueue) mutation-verified; webui/web-shell
typecheck + existing session-action tests green.
* fix(scheduled-tasks): guard tick double-fire, cap rehydration, harden lifecycle writes
Review follow-ups (#6389):
- Extend the fire-persist re-detection guard to ON-TIME tick fires, not just
catch-ups (renamed deliveredCatchUp → firePersistPending): a bound task fired
by the tick advances lastFiredAt asynchronously, so a reload racing that write
(bound detection runs every reload) could re-detect the slot and double-fire.
The tick persist now adds its ids to the guard and clears them when the write
lands, symmetric to the catch-up persist.
- Bound boot-rehydration concurrency (batches of 4): each loadSession forks a
child, so loading up to 50 at once spiked the host and risked spawn failures
that strand tasks. The keepalive revive path was already sequential.
- Archive disable failure is now logged (was fully swallowed) so a broken
archive→pause coupling — where the keepalive would revive the just-archived
session — is diagnosable.
- Unarchive re-enable failure is surfaced in the result `errors` and logged, and
enableTasksForSessions also runs for already-active sessions — so a task left
stranded ({enabled:false, disabledByArchive:true}) by a prior failed enable is
recoverable by re-unarchiving, instead of being permanently stuck.
- Create rollback now removes the persisted session (close + removeSession), so
the loser of a concurrent create at the cap boundary (passes the pre-check,
loses the authoritative write) doesn't leave an orphan named session.
New tests (tick-fire guard, bounded rehydration, already-active recovery)
mutation-verified; full core scheduler + serve suites green.
* fix(scheduled-tasks): one-shot run/edit correctness; tick persist non-regression
Review follow-ups (#6389, ci-bot):
- [Critical] Manual /run on a ONE-SHOT task now removes it from the store. Its
slot is still in the future, so stamping lastFiredAt=now didn't stop the
scheduler firing it again at its original time — a double run. A one-shot's
manual run IS its single fire, so the task is spent.
- [Critical] PATCH recurring:false now re-seats the one-shot's createdAt anchor.
The old (long-past) anchor made the scheduler read it as a MISSED one-shot and
fire + permanently delete it. Re-seating createdAt points its next fire at the
upcoming occurrence. Also covers a cron edit on an existing one-shot.
- [Suggestion] The tick persist no longer regresses lastFiredAt: it skips the
write when the on-disk stamp is already >= the tick slot (a concurrent manual
/run or catch-up may have stamped newer), mirroring the catch-up persist guard.
- [Suggestion] Added the missing create-rollback test: a post-spawn commit
failure closes AND removes the minted session (no orphan).
New tests (one-shot run removal, recurring→one-shot re-seat, rollback teardown)
mutation-verified; core scheduler + route suites green.
* fix(scheduled-tasks): ref-count fire guard, real rehydration cap, authoritative run check
Four [Critical] review follow-ups (#6389):
- Ref-count firePersistPending (was a boolean Set): the same task can have two
lastFiredAt persists in flight (fired again before the first write landed);
clearing on the first settle dropped the guard while the second was still
pending, re-opening the double-fire window. The count holds it until the last
persist settles.
- Rehydration concurrency is now enforced on the REAL loads: loadSession isn't
abortable, so a timed-out load kept forking in the background while the next
batch started. A bounded worker pool holds each slot until the underlying load
actually settles, so in-flight child spawns never exceed the cap.
- Unarchive recovery reports failures for the full resume set: it enables both
unarchived AND already-active sessions but only logged/returned errors for
unarchived, so a failed already-active recovery surfaced errors:[] and left a
task stranded. Deduped one list used for the call, log, and errors.
- Manual "run now" re-checks server-authoritative state before enqueuing: the
dialog snapshot can be stale (another tab/API disabled/deleted the task), so
it would execute the prompt and only the /run record would 409. It now
refreshes, bails if gone/disabled, and enqueues the FRESH prompt/session.
New tests (ref-count, slot-held-past-timeout, stale-disabled re-check)
mutation-verified; core scheduler + serve + dialog suites green.
* fix(scheduled-tasks): catch-up non-regression, disabled-edit re-seat, run/timer/keepalive hardening
Review follow-ups (#6389):
- [Critical] Catch-up persist no longer regresses lastFiredAt: use `>=` like the
tick persist, so a newer stamp (a cross-process manual /run) landing while the
catch-up write is in flight isn't overwritten back to the older minute.
- [Critical] The PATCH anchor re-seat now runs for schedule edits even while the
task is disabled — editing a disabled one-shot's cron then re-enabling it (two
separate requests) no longer leaves a stale anchor that fires + deletes it.
- [High] Manual "run now" of a bound ONE-SHOT consumes it server-side (/run,
which deletes) BEFORE enqueuing, so a record failure leaves a recoverable
"recorded but never ran" instead of a silent double execution at its slot.
- [Medium] The dialog reload timer backs off a stuck past-due nextRunAt (fast
reloads to catch a just-fired advance, then a slow lane) instead of spinning a
1 Hz GET loop.
- [Medium] The manual-run latch bounds the admission phase with a timeout, so a
send that wedges before admission degrades to a visible "run failed" instead
of freezing the run controls.
- [Suggestion] Keepalive: an in-flight guard skips a tick while the previous
pass runs (no duplicate concurrent loadSession spawns), and per-session
exponential backoff stops retrying a permanently-gone session every interval.
New tests mutation-verified. Two deeper items (a task session winning the
durable lock and firing unbound tasks; tearing down a consumed one-shot's
session) are left open as tracked follow-ups — both need new daemon↔child
infrastructure.
* fix(scheduled-tasks): one-shot anchor on unarchive, memoize next-fire, sanitize + log
Review follow-ups (#6389):
- [Critical] enableTasksForSessions now re-seats a ONE-SHOT's createdAt anchor
(not just recurring's lastFiredAt) on unarchive — otherwise unarchiving a task
that was converted to recurring:false while disabled fires it as a missed
one-shot and permanently deletes it.
- [Critical] Log the DELETE-path removeTasksForSessions failure (was fully
swallowed) like the archive/unarchive paths — the session is already gone, so
a silent write failure leaves the still-enabled bound task a permanent ghost.
- [Medium] Memoize nextDurableFireMs (deterministic per id/cron/recurring/anchor)
— a sparse cron costs hundreds of ms per scan and the route recomputed it per
task on every request, stalling the event loop for 50 yearly tasks.
- [Nit] The consumed one-shot /run response now nulls nextRunAt (it was
advertising a future fire on an entity the next GET omits).
- [Suggestion] scheduledTaskSessionName strips terminal control sequences (the
bridge title guard rejects them → silently drops the rename) and truncates on
a code-point boundary (no lone surrogate broadcast as U+FFFD).
- [Critical/doc] Document that firePersistPending is instance-scoped — the
narrow cross-instance restart window is an accepted edge.
- Added the missing test for editing an enabled one-shot's cron.
New tests mutation-adjacent; suites green. Two deeper items (session deleted
outside the daemon orphaning a bound task; surfacing bound tasks in cron_list)
are left open as tracked follow-ups.
* fix(scheduled-tasks): re-seat one-shot anchor on re-enable; guard duplicate revive
Two [Critical] review follow-ups (#6389):
- Re-enabling a one-shot now re-seats its createdAt anchor (added justReEnabled
to the one-shot branch). A one-shot disabled past its slot then re-enabled was
otherwise read as a missed one-shot on the next reload — fired immediately and
permanently deleted. Updated the prior "leaves anchor untouched" test to the
safe behavior (fires at next occurrence).
- Keepalive revive no longer spawns a duplicate child: loadSession isn't
abortable, so a timed-out revive keeps running; a later tick (past its backoff)
would start a SECOND load for the same session. An in-flight `reviving` set
(cleared on the load's TRUE settlement, not the timeout) blocks that — without
holding the sequential tick, so other sessions' heartbeats aren't delayed.
Added a configurable reviveTimeoutMs for the test.
Both mutation-verified. (The one-shot /run session teardown raised again is the
same item as the open deferral — a synchronous close there would break the run,
which executes after /run; it's tracked for the keepalive orphan-sweep.)
* fix(scheduled-tasks): strip bidi override/isolate chars from session name
The bridge's title guard (hasControlCharacter) only rejects C0/DEL, so
Unicode bidi override/embedding/isolate controls (U+202A–202E, U+2066–2069)
slip past it and can visually reorder a scheduled-task session name in the
session list — a Trojan-Source-style attack (CVE-2021-42574). Strip them
alongside the existing terminal-control-sequence pass, matching core's
stripDisplayControlChars canonical set.
Adds a test built from code points so the test file itself carries no
reordering controls.
* fix(scheduled-tasks): close review findings — rehydrate deadlock, manual-run recording, shared helpers, tests
Addresses the review findings on the per-task-session work:
- keepalive rehydrate no longer awaits a non-abortable loadSession after its
timeout. A genuinely hung load would pin its worker and, with enough hangs,
wedge the whole boot sweep (Promise.all never settles) so later task sessions
never rehydrated. The worker now records the timeout as failed and pulls the
next queued session; the background load is left to settle. Rewrote the test
that pinned the old "hold the slot" behavior into a no-wedge regression guard.
- web-shell manual run drops its pre-admission timeout. sendPrompt isn't
abortable, so rejecting on the timer while the send was still in flight let a
LATE admission execute an UNRECORDED run the user could retry into a
duplicate. The run is now tied to admission (accepted prompts are always
recorded); the "session never becomes active" phase stays bounded by the
switch timeout in runTaskManually.
- extract collectBoundSessionIds() shared by the heartbeat + rehydrate passes
(was duplicated) and isBoundTask() in the lifecycle module (was the lone
`sessionId !== undefined` check vs. the strict one used everywhere else).
- spell the nextDurableFireMs cache-key separator as `\x00` rather than a
literal NUL byte, so cronScheduler.ts no longer reads as binary to ripgrep.
- add App.test coverage for the manual-run orchestration (admission-resolve,
cancel/error reject, immediate fire, supersede, switch timeout) and a
keepalive test that a disabled task gets no heartbeat and no revive.
* fix(web-shell): "create via chat" opens a fresh session in scheduled tasks
The scheduled-tasks "Create via chat" button switched to the chat view but
stayed on the CURRENT session, piling the task-creation conversation onto
whatever the user was already doing. It now starts a new session first
(createNewSession) and jumps to it before priming the composer, so task
creation gets its own chat. Covered by a new App.test case asserting
clearSession() is called.
* fix(scheduled-tasks): address follow-up review findings
- keepalive rehydrate: guard the onError callback with try/catch. If it threw
(e.g. stderr EPIPE during log rotation) the rejection escaped loadOne, failed
its worker, and short-circuited Promise.all — stranding every other queued
session.
- cronScheduler catch-up: use the strict `typeof sessionId === 'string' &&
length > 0` bound-check instead of `!== undefined`, matching every other
"is bound?" site.
- server rehydration: log the outer defense-in-depth catch instead of swallowing
it, so an unexpected throw isn't a silent "tasks never fire".
- session-name sanitizer: also strip the standalone Bidi_Control marks U+061C /
U+200E / U+200F, not just the override/isolate ranges.
- scheduled-tasks dialog: when a consumed one-shot then fails to deliver, show a
specific "deleted but never ran — recreate it" error instead of the generic
"run failed" that hid the deletion. Kept the deliberate consume-first ordering.
* fix(web-shell): don't prime the composer when "create via chat" can't start a new session
onCreateViaChat's deferred composer-priming ran unconditionally: if
createNewSession() failed, the task-starter text was dropped into the CURRENT
session (only onSessionIdChange was gated on success). Gate all post-create
side effects on `created`, matching handleMissingSessionNewSession. Adds an
App.test failure-path case (new session fails → composer not primed).
---------
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
Design docs and implementation plans were scattered across .qwen/design,
.qwen/plans, and docs/superpowers. The .qwen/ locations are git-ignored, so
docs written there never got tracked, while docs/design already held the
richer, version-controlled set. Consolidate everything under docs/design and
docs/plans, relocate two stray root docs into docs/design, and repoint the
references left dangling by the move (moved-doc cross-links and a few source
comments).
Also update AGENTS.md and the feat-dev skill so the documented workflow writes
new design docs and plans to the tracked docs/ locations.
Co-authored-by: DragonnZhang <dragonzhang1024@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Reorder reminderParts in getInitialChatHistory() so stable parts (MCP instructions, skills snapshot, startup context) come first and volatile deferred-tools reminder is last — prefix-caching servers retain the KV-cache for the shared prefix, only the tail recomputes.
Co-authored-by: Aleks-0 <aleks-0@users.noreply.github.com>
* feat(cli): add --project and --global flags to /model for per-project model persistence
Add scope control to the /model command so users can persist model
selections to either project-level or user-level settings independently.
- /model --project: persist to workspace .qwen/settings.json
- /model --global: persist to user ~/.qwen/settings.json
- /model (no flag): unchanged behavior (backward compatible)
- Model dialog title shows scope: 'Select Model (this project)' / 'Select Model (global)'
- Completion and argumentHint updated with new flags
- Full i18n support for zh/en
Closes#6052
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): add missing zh-TW translations for /model scope flags
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): address PR review — scope flags, subcommand persistScope, titles, tests
- parseScopeFlags: use (?:^|\s) instead of \b for --flag matching
(\b fails because - is not a word character)
- Completion: strip all flags to isolate model prefix, supports any order
- Subcommand dialogs (fast/voice/vision) now propagate persistScope
- slashCommandProcessor forwards persistScope for all subcommand cases
- ModelDialog title combines subcommand mode + scope label
e.g. 'Select Fast Model (this project)'
- Subcommand confirmations show scope suffix (project/global)
- Extract persistScopeSpread() helper to reduce duplication
- Add 9 tests covering scope flags, dialog returns, confirmations
- Add i18n keys for scope suffix labels in zh/en/zh-TW
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): use Partial<Config> & {[key:string]:unknown} to fix index signature TS error
Replace Record<string,unknown> with Partial<Config> & {[key:string]:unknown}
to satisfy TS4111 index signature access rule in the CI build.
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): add scope suffix to ModelDialog history items
Address review comment: historyManager.addItem for voice/fast/vision/main
model selections now shows scope indicator like ' (this project)' or
' (global)', consistent with CLI direct-set confirmations.
Affected: handleModelSwitchSuccess (main), handleSelect (voice/fast/vision)
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): wrap scopeSuffix in t() and unify wording with ModelDialog
- scopeSuffix in modelCommand.ts now uses t(' (this project)') / t(' (global)')
instead of hardcoded English strings, matching ModelDialog.tsx wording
- Main model confirmation uses shared scopeSuffix instead of separate
i18n keys, eliminating 'Model: {{model}} (project)' duplication
- Remove unused i18n keys from en/zh/zh-TW locales
- Update tests to expect '(this project)' wording
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): address code review feedback — scope validation, i18n, tests
- Reject inline prompt + scope flag combination with clear error (#1)
- Add mutual exclusivity check for --project and --global (#5)
- Verify setValue scope parameter in tests + add --global test (#2)
- Extract scopeSuffix to shared variable, remove duplication (#3)
- Remove dead i18n keys 'Select Model (this project)' / '(global)' (#4)
- Fix scopeSuffix placement on model line not API key line (#8)
- Add fr.js / ja.js translations for scope keys (#10)
- Remove unused export ModelDialogPersistScope (#6)
- Wrap non-interactive help text in t() with new flags (#7)
- Fix argumentHint grouping to show mode vs scope flags (#11)
Signed-off-by: Alex <alex.tech.lab@outlook.com>
* fix(cli): reject --project when workspace is untrusted
Reject --project scope flag before direct persistence or opening ModelDialog
when settings.isTrusted is false. Workspace settings are ignored on merge in
that state, so the save would silently not take effect.
Also mirrors the guard in ModelDialog.tsx resolvePersistScope() to fall back
to user scope when the dialog is opened with --project on an untrusted folder.
Default mock settings now includes isTrusted: true.
Signed-off-by: Alex <alex.tech.lab@outlook.com>
---------
Signed-off-by: Alex <alex.tech.lab@outlook.com>
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
* fix(core): Gate large PDF text extraction
Prevent text-only PDF fallback from injecting full large-document extraction results into the prompt. Large attachment reads now become short references, direct no-pages reads return a short file-too-large error, and page-range extraction is token guarded.
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(core): Stabilize large PDF reference test
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Address PDF budget review feedback
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(core): Stabilize PDF read-file test
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Allow page-range reads for huge PDFs
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Clarify PDF text truncation contract
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Address PDF review follow-ups
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* test(core): Cover multi-page PDF guidance
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Harden paged PDF extraction guards
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Restore authoritative PDF page-count gate
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Keep large PDF references independent of pdftotext
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(core): Narrow dense PDF retry guidance
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
---------
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
* fix(cli): smoother streaming table rendering
Follow-up to the streaming table hold-back, on its own branch so the
cue-removal PR (#6340) can land undisturbed. Makes a live table stream
predictably instead of jittering, flashing, or hanging.
- Atomic rows: hold a frontier row back until it has ALL its columns. A
multi-column row passes through intermediate states that are themselves
valid rows with fewer cells (`| a |`, `| a | b |` toward `| a | b | c |`),
so the old hold-back let it fill in cell by cell. Now the whole row
(border + every cell) appears in one step.
- Widths track the current rows (no freeze): a wider row redraws the whole
table once; a narrower row changes nothing (widths are a max over all
rows, so they only ever grow). Redraw-on-wider only, never per token.
- Bias the streaming preview to the horizontal format: while a table is the
live frontier it only falls back to the vertical `label: value` list when
the terminal is genuinely too narrow, not because an early row wraps tall.
This stops a table from briefly rendering as a vertical list and then
flipping to a horizontal table (a visible jump).
- Hold a forming table back until it is recognizable: a header (and any
partial separator) is trimmed while pending until a separator matching the
header's column count arrives, so the header no longer streams in char by
char as raw `| a | b |` text before snapping into a box. Fenced code-block
content is left untouched.
- Draw the empty header box as soon as the table is recognized, before the
first row completes, so the table area does not sit blank (no box, no cue)
and look like a hang if generation stalls in that window. A zero-row box
omits the header/body divider so it reads as a clean header, not an empty
row.
Only the live frontier table is affected; completed and committed tables use
the normal logic. 211 tests pass (MarkdownDisplay + TableRenderer).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): guard the two remaining zero-row / non-table edge cases
Review follow-up (two [Critical] findings).
- TableRenderer: the maxLineWidth safety check is a second path to the
vertical format, unguarded for zero-row tables. On a very narrow terminal
a zero-row streaming header box would fall through it and render an empty
string — the box vanishes. Skip that fallback when there are no rows so the
header stays visible even if it slightly overflows.
- MarkdownDisplay: the pre-loop header hold-back trimmed ANY trailing run of
pipe-leading lines. When the first line is not a complete `| … |` row,
headerCells was 0 and the run was trimmed anyway — so non-table pipe text
(an un-fenced shell pipeline `| grep foo`, pipe-prefixed log output) would
vanish from the live preview until commit. Only hold back when the first
pipe-line is a plausible table header (a complete row).
Tests cover both. 215 pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): hold a multi-column header mid-type without hiding pipe text
The previous commit (restricting the header hold-back to a complete `| … |`
row, to stop non-table pipe text from vanishing) reintroduced the cell-by-
cell header flash: while a header is typed (`| Alpha`, `| Alpha | Bet`, …)
it is not yet a complete row, so it rendered as raw text.
Discriminate by column count instead of closed-ness: a table header has ≥2
columns; a single-pipe line (shell pipeline `| grep foo`, pipe-prefixed log)
has one cell. Count cells on the first line whether or not it is closed, and
hold the run only when it has ≥2 columns and no matching separator yet. So a
multi-column header held mid-type no longer flashes, while single-pipe non-
table text still renders (the earlier [Critical] fix stands). A header still
typing its very first cell is indistinguishable from a single-pipe line, so
it shows briefly until the second column appears — the narrowest flash
possible without hiding real pipe text.
217 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): make table format decision consistent, not streaming-biased
The horizontal-vs-vertical bias (force a live table horizontal while
streaming) backfired for tables that genuinely belong in the vertical
`label: value` format — a wide table with many columns of long, wrapping
text. It rendered horizontal (tall, clamped, looking stuck) while it was
the streaming frontier, then flipped to vertical the moment it stopped
being the frontier (the next block started) or committed — a visible
format flip, and worse than the vertical-list flash it was meant to avoid.
Drop the streaming bias: the horizontal-vs-vertical decision is now the
same while pending and once committed, so a table never flips format
between the two. Removes the now-unused isStreaming / isStreamingFrontier
plumbing.
Known residual (pre-existing, not from this change): because column widths
track content (redraw-on-wider), a borderline table's wrapped-row height
can still cross the vertical threshold mid-stream. Fully stabilizing that
needs a content-independent format decision — a separate change.
217 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs(cli): note the redraw-on-wider format-oscillation trade-off
Document the accepted limitation next to the horizontal-vs-vertical
decision: because column widths track content (redraw-on-wider), a table
with very long cell text sitting right at MAX_ROW_LINES can still oscillate
format while streaming. Only extreme wide/long-text tables hit it; the
alternatives (content-independent decision, or frozen widths) each cost
more than the residual is worth.
Comment-only; no behaviour change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): count held-back header columns like the table detector
The streaming hold-back counted header columns on the full line with
empty cells filtered out, while the main table detector strips the outer
pipes and splits without filtering. For a header with an empty-named
column like `| A || B |` the two disagreed (2 vs 3), so the hold-back
never found the matching 3-column separator and hid the table for the
whole stream. Count columns the same way in both places.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): release multi-cell non-table pipe content during streaming
The streaming hold-back keeps a run of pipe-lines back until a matching
separator arrives, so a real multi-column header does not flash in cell
by cell. But multi-cell non-table pipe content — a shell pipeline
(`| grep foo | wc -l`), a log excerpt (`| 200 | OK | GET /x`), an
ASCII-art border — also has >=2 cells, so it was held for the entire
stream and only appeared on commit.
A markdown table's separator is the line immediately after the header, so
once a line follows the header and does not even begin like a separator
(optional pipe, optional colon, then a dash), the run is decided: not a
forming table. Release it. A lone header still being typed (no line after
it yet) is still held, so the no-flash behavior is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): anchor the vertical-format decision to the first row (no flip)
The horizontal-vs-vertical choice used maxRowLines measured over EVERY row,
so a table that started horizontal (short first row) flipped to vertical the
moment a later, taller-wrapping row streamed in — a visible mid-stream format
change. Measure only the header + the first data row instead. The first row is
representative for the common case, so the format is decided once and stays
put as rows append. Column widths still track all rows (redraw-on-wider is
unchanged); only the format choice is anchored.
Trade-off: a table whose first row is short but a later row wraps very tall
stays a (taller) horizontal grid rather than flipping to vertical — rare, and
preferable to a visible flip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): release a dash-led data row from the streaming hold-back
The "could this pipe run still become a table?" check treated any line after
the header that merely started with a dash as a possible separator, so an
options table whose first data cell begins with a flag — `| --verbose | … |`
— was held back for the whole stream. Use tableSeparatorRegex instead: it
still matches a partial separator being typed (`|--`) so a real header is
held until its separator lands, but rejects a dash-led data cell (trailing
letters), which now renders live.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): defer a streaming table until its first row (no empty-box flip)
A recognized table with no complete data row yet was drawn immediately as an
empty header box. A zero-row table can only render horizontally (the vertical
fallback needs rows), so once a long first row landed the box flipped to the
vertical label:value format — a visible format change that cannot be avoided
by looking at the header alone (column names are short; width comes from the
values). Defer the table while pending until its first row completes, so it
first appears already in its final format with no flip.
Cost: the table area stays blank while the header + first row stream (the
pre-loop trim already hid the header text, so this only extends that blank).
Committed tables always have rows, so their behavior is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): address review — code-fence tracking, held-back edge cases, committed format
Five review findings:
- [Critical] The pre-loop hold-back's code-fence check used a naive boolean
toggle that ignored fence char/length, so a nested fence (```` with an inner
```) mis-closed and a real code line like `| A | B |` was held back and
vanished while streaming. Track the open fence's delimiter and validate the
close (same char, >= length), mirroring the main parser.
- A COMPLETE separator whose column count already differs from the header can
never match, so release the pipe run instead of holding it for the whole
stream (the main parser treats it as text).
- The end-of-content table flush now uses the same `tableRows.length > 0` guard
as the mid-content handler, so a degenerate zero-row table behaves the same
whichever way it ends — no EOF-vs-mid asymmetry.
- TableRenderer's first-row-only maxRowLines (no-flip) applied to committed
tables too; a committed short-first-row + tall-later-row table wrongly stayed
horizontal. Gate on a new `isPending` prop: measure the first row only while
streaming, every row once committed (most readable, no flip concern).
- Renamed the test block that claimed a nonexistent `isStreaming` prop; added
committed-vs-streaming format tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): don't flip a completed mid-content table's format at commit
A table closed by a following line is complete even while the message keeps
streaming, but it was still rendered with the first-row-only format anchor —
so a short-first-row + tall-later-row mid-content table showed horizontal and
then flipped to vertical the moment the message committed.
Split the two concerns that were both riding on `isPending`:
- the height clamp still tracks whether the MESSAGE is streaming (so a
mid-content table stays bounded and the estimator's clamped cost can't
under-estimate the render);
- the format anchor now tracks whether THIS TABLE is the streaming frontier.
The mid-content flush passes isFrontier={false} → all rows measured → final
format now; only the end-of-content (frontier) table anchors to the first row.
Renamed TableRenderer's format-anchor prop to `isStreaming` (it is not the
message-level pending flag). Added mid-content and tilde-fence tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): don't hold a pipe line inside an open $$ math block
The streaming table hold-back tracked code fences so a `| A | B |` code line
would render, but not display-math (`$$ … $$`) blocks. The main parser pushes
math content verbatim (never as a table), so a `| a | b |` norm/matrix line at
the frontier of an open math block was treated as a forming table and blanked
until the block closed. Track math fences in the trim's fence scan too, mirroring
the main parser's precedence (code block wins, then math), and skip the hold-back
while inside one. Addresses the low-confidence review observation on #6345.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The bundled /review skill is a general command that runs against arbitrary
repositories (and cross-repo PRs), but a previous change baked qwen-code's own
"core infrastructure is maintainer-only" governance into the shipped prompt:
hardcoded packages/core and packages/*/src/{auth,providers,models,config,tools,services}
paths, a 500+ line hard block, and an authorAssociation-based maintainer check.
Those path names are generic — src/auth, src/config, src/tools, src/services are
common across monorepos — so an external contributor's large PR to an unrelated
repo would be hard-blocked as "must be maintainer-initiated" under a policy that
repo never adopted.
Remove the gate and its escalate-flag plumbing (Steps 1, 6, and 7) from the
bundled skill, along with the matching DESIGN.md rationale and the user-doc
section. qwen-code's maintainer-only policy stays documented in AGENTS.md for
this repo. The Issue Fidelity / root-cause ownership agent (Agent 0) is a
universal review principle and is left unchanged.
Co-authored-by: dragon <dragon@U-2Q53JQG9-0233.local>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* feat(web-shell): add Session Overview panel and in-window split view
Add a large-screen "Session Overview" mission-control panel and an
in-window split view so users can monitor and drive multiple daemon
sessions at once.
- SessionOverviewPanel: ranked live cards (needs-approval -> running ->
idle) merging the workspace session list with the detail=full status
report. Multi-select opens the selected sessions as a split view in
the current tab ("Open in split") or in a new browser tab ("Open in
new tab", via a ?split=a,b URL).
- SplitView + ChatPane: one DaemonWorkspaceProvider hosting N
DaemonSessionProvider panes, each a self-contained interactive chat
(transcript, composer, streaming, tool/ask approvals). Browser focus
scopes the keyboard per pane, so panes never contend over approvals.
- Sidebar entry points gated to large screens; the split view's Back
returns to the Session Overview.
* refactor(web-shell): address review feedback on the session overview / split view
- SessionOverviewPanel: prune the selection Set when a session leaves the list
(so a reappearing session isn't silently reselected) and make select-all use
the intersection rather than prev.size.
- Extract isAskUserPermission into a shared util so App.tsx and ChatPane.tsx no
longer keep verbatim copies that can drift.
- SplitView: dismiss the "add session" picker on Escape or a click outside it.
- Tests: MAX_PANES cap, popup-blocked path, checkbox-selects-without-navigating,
stale-selection pruning, and a direct test for the extracted util.
* fix(web-shell): address /review findings on the split view
- ToolApproval: add a `keyboardActive` prop; split panes pass false so global
Enter/Escape/digit shortcuts can't confirm the wrong session's approval, and
the outer session's approval overlay is no longer rendered behind the split
(where it would keep its global shortcuts while hidden).
- ChatPane: defer the composer commit until sendPrompt resolves, so a rejected
prompt (transcript loading / disconnected / turn active) preserves the draft
instead of silently dropping it.
- SplitView: include a per-mount nonce in each pane's clientId so two tabs
opening the same split don't share a client id — which suppressOwnUserEcho
would treat as a self-echo and drop from the transcript.
- SessionOverviewPanel: cap the split selection to MAX_SPLIT_PANES before
building the ?split= URL or opening the in-window split, with a hint when more
are selected; also dismiss the split picker on Escape / click-outside.
- Tests covering each.
* fix(web-shell): address second /review round on the split view
- SplitView: wrap each pane in its own ErrorBoundary, so a render crash in one
pane (malformed block, unexpected tool shape) shows an inline fallback with a
close action instead of white-screening the whole split.
- splitUrl / overview: carry the daemon token into the new-tab split URL's
fragment. The current tab has already stripped the token from its URL, so a
token-auth (`serve --open`) deployment would otherwise open the split tab
unauthenticated. The token rides the hash (never sent to the server / logs).
- Tests: per-pane error isolation, token-in-fragment (and none without a token),
and the overview polling effects (interval fires, document.hidden skips, and
the in-flight guard prevents overlapping polls).
* fix(web-shell): hide the outer chat under the split and share app-level contexts
- App: hide (display:none) + aria-hide the outer chat subtree whenever
mainView !== 'chat', not only when a panel is open. Previously the outer
chat/composer/toolbar stayed reachable by keyboard/AT behind the full-page
split (it was only covered visually). State is preserved (node stays mounted).
- App: wrap SplitView in the app-level WebShellCustomizationProvider and
CompactModeContext so split panes render markdown / tool-headers / thinking
the same way the single-session chat does. Todo contexts stay chat-only —
they belong to the outer session, not the panes.
* refactor(web-shell): address review suggestions — coverage, dedup, split UX
- ToolApproval: add a dedicated test on the real component that the global
keyboard shortcut is armed by default and NOT armed when keyboardActive=false
(the cross-pane approval safety mechanism).
- SplitView: auto-exit to the Session Overview when the last pane is closed
(guarded so an initial empty seed doesn't bounce straight back out).
- ChatPane: add tests for the cancel action, the empty/whitespace submit guard,
and error routing to the onError prop.
- Extract the shared session-list page size + organization feature flag into
constants/sessions.ts, used by the overview, split view, and sidebar, so the
values can't drift between the three.
* fix(web-shell): surface outer approval + failed refresh in overview/split
- Split view: when the outer (main) session is waiting on an approval
that's hidden behind the split, show a non-blocking notice banner with
a "Go to it" button that returns to the chat where the approval lives.
- Auto-close the split (like the overview panel) when the viewport shrinks
below the large-screen breakpoint, so users aren't stranded.
- Session Overview: surface a failed refresh inline (keeping the last-good
cards) instead of silently swallowing it once cards are on screen.
- Tests: status-report poll cadence, picker dismiss (Escape / outside /
inside click), inline refresh-failure banner.
* fix(web-shell): sever window.opener on split tab; tighten hidden-chat test
- openSelectedInNewTab now clears win.opener (the split tab carries a
daemon token in its URL fragment) to prevent reverse tabnabbing, matching
the existing bug-report window.open path.
- Strengthen the split-view App test so a missing outer-chat subtree fails
instead of passing vacuously through an optional chain.
* fix(web-shell): split-view focus/stability/robustness follow-ups
- Refocus the composer after a shrink-driven split close so keyboard users
aren't dropped onto <body> (skips when an approval or panel takes over).
- Stabilize SplitView onExit via useCallback so its last-pane-close effect
doesn't re-fire on every App re-render.
- ChatPane: surface a per-pane connection-loss banner instead of silently
showing stale messages when a pane's daemon connection drops.
- ChatPane: anchor the streaming timer to the active turn's start (last user
message timestamp) so a pane opened mid-turn shows real elapsed time.
- Tests: split auto-close on shrink, outer-approval split notice + return-to-
chat, connection banner, and streaming-timer anchoring.
* fix(core): reduce multimodal history payload size
* fix(core): use kebab-case image payload filenames
* fix(core): address image payload review blockers
* fix(core): preserve current request image payloads
* ci: disable implicit actionlint pyflakes integration
* fix(core): reattach recent unique image payloads
* fix(core): preserve referenced image payloads
* fix(core): tolerate partial config mocks in MCP discovery
* fix(core): preserve current images during recovery
* fix(core): gate image payload replacement behind threshold
The always-on image payload replacement introduced by PR #6045
replaced ALL historical images with text references on every request,
causing users' old screenshots to be reattached and triggering
infinite fix loops when the model mistook stale buggy screenshots
for current state.
Replace the always-on approach with a threshold-gated design:
- Below 20 images (configurable): zero transformation, images stay
in-place in history
- At or above 20: in-place replace historical images with text
references, reattach only the most recent 3 unique images
- Replacement is persistent (mutates this.history), so the count
resets and won't re-trigger until 20 new images accumulate
- Current user request images are protected via skipContent
Also lower DEFAULT_SCREENSHOT_TRIGGER_THRESHOLD from 50 to 20 to
align with the new image payload threshold.
---------
Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
Fixes#6311
The extract cursor previously advanced unconditionally after the
forked extractor agent reported 'completed', even when it made zero
real tool calls (e.g. a small/local model hallucinating a bash
command instead of calling write_file). This silently and
permanently skipped those history messages from being reprocessed.
Also fixes extractionAgentPlanner.ts using filesTouched (attempted
paths, unconfirmed) instead of filesWritten (confirmed successful
writes) when deriving touchedTopics, matching the pattern already
used in remember.ts.
touchedTopics.length > 0 alone is not sufficient to gate the cursor
advance: a legitimate 'nothing durable to save' outcome also produces
an empty touchedTopics array and would otherwise be treated the same
as a hallucinated run. A new hasToolActivity signal (derived from
filesTouched, which includes read-only calls like read_file)
distinguishes 'agent engaged with the task and found nothing new to
save' (legitimate noop, cursor still advances) from 'agent made zero
tool calls at all' (hallucination, cursor held for retry).
The channel proxy path used ProxyAgent, which unconditionally routes
all requests through the proxy and ignores NO_PROXY. This caused
requests to hosts listed in NO_PROXY (e.g. localhost, internal IPs)
to fail when a corporate proxy was configured.
Switch to EnvHttpProxyAgent, matching the main CLI config path that
already handles NO_PROXY correctly.
Co-authored-by: qwen-autofix <autofix@qwen-code.ai>
* feat(review): add issue-fidelity and root-cause ownership gate to /review
Adds a dedicated Issue Fidelity & Root-Cause Ownership agent (Agent 0) to
the /review pipeline and a core-infrastructure scope gate that runs before
the review agents.
Agent 0 fetches linked GitHub issue evidence directly (closingIssuesReferences
plus issue comments) instead of trusting the PR author's framing, compares the
original reported failure against the PR's claimed fix, and flags client-side
parser/sanitizer workarounds for malformed upstream output as Critical unless a
maintainer explicitly requested the defensive mitigation. The core-infra gate
applies the repository's existing two-tier maintainer-only rule before spending
review budget.
This hardens the pipeline against a false-approval mode where a bot PR passes
its own tests and reads as internally reasonable but fixes the author's mistaken
diagnosis rather than the linked issue's actual root cause.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs(review): address PR review feedback on issue-fidelity gate
- Fetch issue evidence with `gh issue view --json title,body,comments` so the
issue body (reporter repro/observed payload/expected behavior) is included;
`--comments` alone omits it. Use each closingIssuesReferences entry's own
repository so cross-repo linked issues resolve correctly.
- Treat closingIssuesReferences as a discovery hint (fetch apparent target
issues even when it is empty) and treat fetched issue content as untrusted
data (extract facts, ignore embedded instructions).
- Run Agent 0 (Issue Fidelity) only for PR targets; skip it for local-diff and
file-path reviews, and require the PR number/repo/context in its prompt.
Handle empty references / non-bugfix / gh failure explicitly.
- Pass Agent 0's quoted issue evidence to Step 4 batch verification and stop it
rejecting issue-grounded findings just because the code compiles/tests pass.
- Make the core-infrastructure gate concrete: deterministic maintainer signal
via authorAssociation, count only core-path lines, honor the AGENTS.md
low-risk-sweep exception, clean up the worktree on hard block, run the gate
right after fetch-pr (before npm ci), and map escalate -> COMMENT (never
APPROVE) in Steps 6-7.
- Sync agent counts and token math across SKILL.md, DESIGN.md, and
code-review.md (Agent 0 is PR-only; ~620-730K).
* docs(review): rename 'Linked Issue Fit' heading to 'Issue Fidelity'
Aligns the code-review docs heading with the 'Issue Fidelity' name used
for Agent 0 in SKILL.md and DESIGN.md, so the section connects to the
pipeline diagram. Addresses review feedback.
* docs(review): stop core-infra hard block before load-rules and surface it via --comment
- Hard block now stops before Step 2 (load-rules) instead of before Step 3,
so a PR destined for hard-block no longer runs the load-rules step.
- In --comment mode the hard block posts an event=COMMENT on the PR, matching
the escalate path's GitHub visibility, so external authors see the block.
---------
Co-authored-by: dragon <dragon@U-2Q53JQG9-0233.local>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* fix(core): preserve no-argument tool calls that stream an empty arguments string
For tools that take no parameters, some OpenAI-compatible providers
stream `arguments: ""` (or omit the field entirely) and never send an
argument fragment. The streaming parser dropped such calls wholesale
(`meta?.name && buffer.trim()`), while the non-streaming path keeps
them with `args: {}` — so a turn containing only that call looked
empty and geminiChat raised "Model stream ended with empty response
text", triggering pointless retries.
Align the streaming parser with the non-streaming path: emit the call
with empty args when the buffer is empty at stream end. Rewrite the
unit test that encoded the drop, and add regression coverage at parser
and converter chunk level.
* fix(core): use name metadata as slot-occupancy signal for no-argument tool calls
Follow-up to review feedback: after empty buffers became a legal end
state for no-argument tool calls, three parser methods still used
buffer.trim() to decide whether an index slot was occupied. A provider
reusing indices could then silently overwrite a completed no-argument
call (addChunk collision guard, findNextAvailableIndex) or append stray
continuation chunks to it (findMostRecentIncompleteIndex).
Switch the occupancy signal in all three places to the name metadata,
keeping the JSON-completeness check for non-empty buffers. Add
regression tests for both corruption paths and update the stale
getCompletedToolCalls JSDoc.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): collapse non-object argument parses at emit and lock canonical empty-opener shape
Review follow-up on the no-ID continuation routing at addChunk. Mid-stream,
an empty buffer with name metadata is formally undecidable between "completed
no-argument call" and "canonical opener awaiting its first argument fragment"
(every OpenAI-compatible provider opens with arguments:"" and streams
fragments ID-less at the same index). Routing must favor the canonical shape,
so the guard stays; a new test pins that shape, which the suite previously
did not cover.
The corruption concern from review is instead bounded at emit time: a buffer
polluted by a stray fragment can parse or repair to a non-object value, which
now collapses to {} so a polluted no-argument call still emits empty args.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): add debug logging for empty-buffer emission and non-object argument collapse
Review follow-up: a stray fragment that happens to parse as a valid JSON
object is indistinguishable from real arguments at emit time, so log both
the non-object collapse and empty-buffer emissions to aid diagnosis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): extend replay guard to opener-shaped replays of no-argument tool calls
Review follow-up: after empty buffers became a legal completed state, a
replayed opener (duplicate ID, #5107 lineage) could overwrite a completed
no-argument call's name metadata, since the replay guard only engaged on
non-empty buffers.
Swallowing every known-ID chunk at that state would drop ID-bearing
argument fragments for providers whose opener streams empty arguments, so
the guard uses the protocol shape as discriminator: a chunk carrying a
name but no argument content is an opener replay and is ignored; a chunk
with argument content is a continuation and appends. Regression tests
cover both directions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(core): cover null/array argument collapse and multi-slot relocation scan
Review follow-up: pin the null and array branches of the emit-time
non-object collapse, and exercise findNextAvailableIndex scanning past
multiple occupied no-argument slots during collision relocation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: FiaShi <FiaShi@fiashideMacBook-Air.local>
Co-authored-by: tomsen-ai <230283659+tomsen-ai@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
When a file is accessed via a symlinked path (e.g., in git worktrees or
monorepos with symlinked directories), conditional rules and skills
keyed on the real path would fail to activate.
Add resolveSymlinkAwareRelativePaths() that returns both the original
and realpath-resolved relative paths, so glob patterns match either form.
Resolve both the file path and projectRoot via realpath to handle macOS
/private/tmp prefix normalization correctly.
Make matchAndConsume() async in both ConditionalRulesRegistry and
SkillActivationRegistry to support the realpath I/O.
Fixes#6356
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
