* fix(lifecycle): don't split streaming messages on subagent tool calls
A subagent's internal tool calls (ParentToolCallID set) stream on the same
session concurrently with the parent agent's own text. Flushing the message
buffer on those events closed the parent's in-flight streaming message, so
each subagent tool call started a new message row — splitting one logical
message mid-sentence and breaking markdown (e.g. **bold** spans rendered as
literal asterisks) across the fragments.
Only flush on top-level tool calls; subagent-internal ones leave the parent's
streaming segment open.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(lifecycle): drop dead flush-text branch in tool call handler
flushMessageBuffer always publishes remaining content through the streaming
path and returns "", so the event.Text / history-append branch could never
execute. Call it directly instead of through a helper carrying dead code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* feat: surface agent subscription usage in settings and chat tooltip
Adds live subscription rate-limit visibility for Claude Code and Codex:
- GET /api/v1/agents/usage lists per-agent utilization windows (plan, percent,
reset time) for host agents with subscription credentials; ?fresh=true
bypasses the 5-minute cache with a 15s per-provider clamp.
- Fixes both existing usage clients: Claude parsed float utilization as int
(decode failure) and now reads the richer limits[] array incl. per-model
weekly windows; Codex read a nonexistent auth.json field and response
headers - now reads tokens.access_token, parses the body windows, and
refreshes OAuth tokens on 401. Token refresh persists without dropping
unknown credential fields.
- Settings > Agents gains a Subscription Usage section; the chat context
doughnut tooltip shows the session agent's windows, fetched fresh on hover.
- Shared UsageWindowRows component (bars, Good/High/Critical status,
reset countdowns) keeps settings, tooltip, and office surfaces consistent.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix: recompute agent billing type on every read
Subscription billing detection for Claude and Codex was cached with
sync.OnceValue for the process lifetime, so logging in with OAuth after
the backend started never flipped billing badges or office usage checks
until restart. Detection now re-reads the credential files on each call
and reuses the usage clients' HasSubscriptionCredentials checks. Also
adds a regression assertion that fractional-second resets_at timestamps
parse under the RFC3339 layout.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix: address subscription usage review feedback
- serialize empty utilization windows as arrays, not null
- sanitize per-agent usage errors in the API (raw provider errors are
logged server-side; they can carry credential paths/response bodies)
- persist refreshed CLI credentials via atomic temp-file rename
- keep previous settings usage on transient refresh failure; drop stale
tooltip usage (and its memo) when a fresh fetch fails
- add controller tests for SubscriptionUsage mapping and fresh propagation
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix: coalesce concurrent usage fetches and guard codex token refresh
- serialize usage-cache fetches per key with a double-checked lock so
concurrent fresh requests make one provider call instead of a burst
- reject codex refresh responses with an empty access_token instead of
persisting them and wiping working credentials
- pin the usage-rows reset fixture 30s under the day boundary so the
formatted value is deterministic on fast runners
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix: coalesce failed usage fetches with short negative caching
Concurrent lookups after a fetch error each retried the provider once
the per-key lock rotated. Failures are now cached for 15 seconds (never
longer than the caller's staleness bound), so bursts share one error
result and the provider gets a breather before the next retry.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix: keep stale usage through failed refreshes and skip caching cancellations
Track the last success and last failure independently in the usage
cache: a failed fresh refresh no longer evicts a still-valid success
(non-fresh callers keep the stale value while fresh callers share the
recent error), and a caller-cancelled context no longer records a
negative entry that would penalize other callers.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test: synchronize concurrent usage-cache tests with a channel barrier
Replaces the sleep-widened race window in the coalescing tests with a
started/release barrier per the backend testing guidelines, and dedupes
the two concurrent-caller tests into one helper.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
* perf(backend): cap cumulative diff bytes and file count during large rebases
* perf(web): coalesce cumulative-diff refetches and skip redundant status compares
* feat(web): warn when the changes panel hides files from a capped diff
* fix(backend): enforce cumulative diff total byte budget strictly
* fix(web): keep coalesced refetch timer shared across envKey subscribers
* fix(web): drop unsound git-status SHA fast-path that staled live diffs
* fix(web): cancel coalesced timer on mid-flight drain to avoid duplicate refetch
* refactor(web): move changes-panel banner and helper to a sibling file
* docs(backend): clarify cumulative diff truncated-files-count semantics
* fix(web): tighten useSelectedFileKey mode param to the all|file union
* fix(backend): prefer local copilot CLI over npx at launch
IsInstalled probed the copilot binary for the UI badge but launch always
ran `npx -y @github/copilot`, paying an npm registry round-trip every
start — minutes behind a private registry. Reuse the env-correct binary
check (standalone LookPath, SSH preflight probe) so when copilot is on
PATH the lifecycle runs it directly; containerized runtimes keep npx.
* fix(backend): make SSH native-binary probe non-fatal on transport error
The native-binary probe is an optimisation, not a requirement. A transient
SSH transport error (MaxSessions, reset) made "Verifying agent binary"
abort the launch, blaming copilot before npx was tried. Treat probe
transport errors as a miss and fall through to the required default-command
probe, which reports failures with proper context. Also call
NativeBinaryName() once.
* fix(backend): clear stale native-binary metadata before SSH probe
A pre-existing or stale MetadataKeyNativeBinary in req.Metadata could make
the command builder prefer copilot --acp on an npx-only host even when this
probe misses. Clear it up front so only a positive probe re-sets it.
* refactor(backend): use resolved RuntimeName for native-binary check
buildExecutionFromInstance re-derived the runtime from req.ExecutorType
while promoteWorkspaceExecution read execution.RuntimeName. Use the already
resolved execution.RuntimeName (set from rt.Name()) in both, so the two call
sites are consistent.
* fix(backend): reflect local copilot CLI in profile command preview
The profile setup command preview always rendered `npx -y @github/copilot`
because PreviewAgentCommand built the command without the native-binary
probe. The preview is not bound to an executor, so assume the default
standalone host and probe PATH (mirroring the standalone launch path): show
`copilot --acp` when the CLI is installed, npx otherwise.
* feat(diff): use stored base_branch for diff stats + per-task compare picker
Task-card additions/deletions and the Changes panel were computed against a
hardcoded origin/main → master priority list inside agentctl. For tasks
branched off a release/staging branch this inflated the counts by every
commit since the actual fork point, and the Changes panel grouped unrelated
files alongside the task's real work.
This wires the per-repo task_repositories.base_branch end-to-end into the
agentctl WorkspaceTracker so each repo's diff stats resolve against the
recorded base instead of the integration-branch fallback. A picker on the
Changes panel branch hover card lets the user change the base mid-task; the
new value is persisted via PATCH /tasks/:id/repositories/:repo_id, pushed
live to the running agentctl instance, and the session's cached
base_commit_sha is cleared so the commits panel + cumulative diff refetch
against the new base without a session restart.
Falls back to the existing origin/main → master list for legacy tasks
(empty base_branch) and to the branch tip when no merge-base exists with
HEAD (unrelated histories — local rebase backups, etc.) so the stats and
commits panels stay anchored on the same SHA in every failure mode.
* fix(agentctl): sanitize user-controlled base_branch refs before git exec
CodeQL flagged the picker's user-controlled base_branch payload reaching
exec.Command("git", ..., baseBranch) inside workspace_git_cmd.go as a
command-built-from-user-controlled-source vulnerability. A malicious caller
could smuggle a leading-"-" argument like "-upload-pack=..." that git
itself reinterprets as a flag, even though exec.Command keeps args
separate from the shell.
Adds an IsSafeGitRef sanitizer at every boundary the value crosses on its
way to git: the service write path, the agentctl HTTP handlers
(/workspace/base-branches, /git/log, /git/cumulative-diff), and
WorkspaceTracker.SetBaseBranch itself. Unsafe values are rejected at the
service tier and silently downgraded to the no-override fallback inside
agentctl so a misbehaving caller can't half-break the tracker.
* fix(diff): address PR review blockers and CI lint
- Split UpdateRepositoryBaseBranch into validate/load/apply helpers to drop
cyclomatic complexity below the cyclop limit that gated Run Backend Tests.
- Replace direct m.cfg.BaseBranches reads/writes with getBaseBranches /
setBaseBranches under a dedicated RWMutex. UpdateBaseBranches used to
write under repoTrackersMu while rescan + lazy-subpath callers read under
workspaceTrackersMu, leaving an unsynchronized data race on the map.
- Drop the false 'PR target is auto-wired' claim from the MCP tool
description — GitOperator.CreatePR takes baseBranch as an explicit arg
and never reads the tracker's value.
- Picker no-op guard now compares against currentBase instead of storedBase
so selecting the displayed fallback (legacy task before any picker
change) is treated as no-op instead of triggering a spurious save.
- HTTP handler now maps non-validation service errors to 500 instead of
400, classifying via the existing isValidationError helper plus the new
unsafe-git-ref message.
* fix(diff): regex-based ref sanitiser at every git subprocess sink
CodeQL kept flagging workspace_git_cmd.go:43 even after SetBaseBranch
rejected unsafe values — the taint tracker doesn't follow sanitised stores
through struct fields, so it still saw user input flowing into
exec.Command("git", ..., baseBranch).
Switches IsSafeGitRef to a single regex check (a pattern CodeQL recognises
as a sanitiser barrier) and adds SanitizeGitRef calls inline at every
agentctl call site that splices a user-controlled ref into a git arg:
- workspace_git_status.resolveStoredRef (before the rev-parse probes)
- workspace_git_status.computeBaseCommit (before merge-base/rev-parse)
- workspace_git_status.getAheadBehindCounts (before rev-list)
- agentctl/server/api/git.go runGitLogForRepo + handleGitCumulativeDiff
(before computeMergeBase/GetRevParse on the HTTP TargetBranch)
Behaviour is unchanged for valid refs — only narrows the allowlist (now
[A-Za-z0-9_./-]) so the sanitiser barrier is recognisable to static
analysis. SetBaseBranch's field-store sanitiser stays as defence in depth.
* fix(diff): transform sanitiser at every map read for CodeQL taint flow
CodeQL still flagged workspace_git_cmd.go:43 because:
- handleSetBaseBranches used an IsSafeGitRef bool guard before a verbatim
map[k] = v copy. CodeQL recognises regex-transforming sanitisers but not
bool-guarded passthroughs.
- lookupBaseBranch returned m.cfg.BaseBranches[k] verbatim. Even though
the map was sanitised on write, taint flow across map stores + field
stores loses the sanitised state in the analyser.
Switches both sites to the transforming SanitizeGitRef helper so every
read produces a freshly-sanitised value before it flows to a downstream
git argument. Behaviour is unchanged for valid refs.
* fix(diff): regex transformer sanitiser + guard live push from empty map
CodeQL still flagged the previous SanitizeGitRef because it returned the
input string unchanged on success — an identity passthrough that the taint
tracker doesn't recognise as a sanitiser barrier.
Switches to safeGitRefPattern.FindString so the output comes from the
regex match (not the parameter). IsSafeGitRef now wraps SanitizeGitRef so
the predicate and the transformer share one allowlist.
Also addresses two Claude review blockers:
- collectTaskBaseBranches returns a nil/empty map when the task has no
recorded base_branches; pushing that to agentctl wipes every tracker's
override. Now skipped — the DB row is the source of truth for the next
session launch.
- applyBaseBranchSideEffects now runs on context.WithoutCancel so an HTTP
/ WS request cancel after the task_repositories UPDATE commits can't
leave session.base_commit_sha stale + the live agentctl tracker map
un-pushed, drifting from the persisted row.
* fix(diff): share securityutil.IsValidBranchName + harden error classification
CodeQL still flagged the picker's regex sanitiser as taint passthrough.
Refactors SanitizeGitRef / isSafeBaseBranchRef to delegate to the existing
securityutil.IsValidBranchName helper that agentctl's Rebase / Merge /
RenameBranch already use without being flagged. The shared allowlist also
fixes a Claude review blocker: 'feature/@2024' used to pass the service
validator but get dropped by the agentctl-side check at push time,
silently wiping every tracker's override.
Plus the two error-classification blockers Claude raised on the previous
HEAD:
- WS handler (task_ws_handlers.go wsUpdateTaskRepository) now distinguishes
validation errors (surfaced verbatim) from internal faults (opaque
500-style message). HTTP path already did this via the existing
isValidationError helper.
- MCP handler (handlers.handleUpdateRepositoryBaseBranch) gets the same
treatment via a new isValidationError local to the mcp/handlers
package so raw service error strings stop leaking across the MCP
boundary.
* fix(diff): inline securityutil.IsValidBranchName at every git subprocess sink
CodeQL's interprocedural taint tracking wasn't recognising SanitizeGitRef
as a sanitiser even after it delegated to securityutil.IsValidBranchName
— the wrapper helper + map/field hops between source and sink kept the
flow tainted in the analyser.
Inlines the IsValidBranchName allowlist check (already used inline by
Rebase/Merge in git.go without being flagged) at every call site that
splices a user-controlled ref into a git subprocess argument:
- workspace_git_status.resolveStoredRef (before rev-parse --verify)
- workspace_git_status.computeBaseCommit (before merge-base / rev-parse)
- workspace_git_status.getAheadBehindCounts (before rev-list)
- agentctl/server/api/git.go runGitLogForRepo + handleGitCumulativeDiff
- agentctl/server/api/workspace_base_branches.go handleSetBaseBranches
Behaviour unchanged: same allowlist as Rebase / Merge / RenameBranch
already use. SanitizeGitRef / IsSafeGitRef survive as defence-in-depth
helpers for callers that want a one-line check; they delegate to the same
securityutil helper so the rules stay aligned.
* fix(diff): inline regexp.MatchString at every git subprocess sink
CodeQL still flagged workspace_git_cmd.go:43 even with the inline
securityutil.IsValidBranchName helper — the taint tracker apparently does
not follow into helper functions for the go/command-injection rule on
this path.
Hoists the regex to package-scope vars (safeBranchRefPattern in both
process and api packages) and calls regexp.MatchString directly at each
sink site. Behaviour is identical to securityutil.IsValidBranchName; only
the call form changes.
* fix(diff): SHA-shape allowlist on BaseCommit before downstream git diff
CodeQL still flagged workspace_git_cmd.go:43 because update.BaseCommit
is the OUTPUT of an earlier merge-base / rev-parse call whose argument
list contained a user-controlled branch ref — the analyser treats that
command's output as still tainted even when the input ref was sanitised.
Adds an inline 40-hex SHA-shape allowlist check in enrichWithBranchDiff
right before `git diff --numstat <update.BaseCommit>` so the sanitiser
barrier sits in the same function as the next subprocess call. Behaviour
is unchanged: merge-base / rev-parse already produce SHA-shaped output.
* fix(diff): async tracker refresh + tighten init spread order
Two CodeRabbit/Claude review fixes:
- UpdateBaseBranches stamps each tracker's new baseBranch synchronously
then runs RefreshGitStatus on a background goroutine instead of inside
the picker-save HTTP handler. Per-tracker git subprocesses can each
take ~3-5 calls; serialising them on the request goroutine made the
PATCH response wait for every refresh to finish. The picker UI re-fetches
via the existing WS stream once Refresh emits, so the handler doesn't
need to block.
- updateTaskRepositoryBaseBranch now spreads options.init BEFORE the
required method/body fields so an upstream caller's init overrides
can't replace the PATCH method or the base_branch payload. Matches
the convention deleteTask / moveTask already follow.
* fix(diff): narrow mcp isValidationError + opaque 500 + post-sanitize log count
Three final Claude review threads on the latest commit:
- mcp/handlers isValidationError drops the broad 'invalid' substring so
DB / IO errors whose messages happen to contain 'invalid' don't
misclassify as ErrorCodeValidation.
- task_http_handlers httpUpdateTaskRepository now logs the raw service
error server-side and returns an opaque 'failed to update task
repository' on the 500 path. Stops potential DB / IO detail leakage
through the HTTP boundary.
- workspace_base_branches debug log reports accepted + rejected counts
(post-sanitisation) instead of the pre-filter total so the metric
actually reflects what hit UpdateBaseBranches.
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
* fix(agentctl): order workspace stream wg.Add before publish to fix Close race
Client.StreamWorkspace published the new WorkspaceStream pointer to
c.workspaceStream under the lock, then released the lock and called
stream.wg.Add(2) afterwards. A concurrent Client.Close captured
c.workspaceStream under the same lock and then called ws.Wait() outside
of it, racing the subsequent Add(2) on a WaitGroup whose counter was
still 0. The race detector flagged this as Add-after-Wait on the same
counter address — reproduced in CI in
TestInitializeAndPrompt_StreamBeforeInitialize and
TestManager_RestartAgentProcess_StopErrorIsNonFatal, both of which use a
deferred client.Close after the StreamManager spawns the workspace
stream goroutine.
Move stream.wg.Add(2) under the same lock that publishes the pointer.
The unlock/lock pair now carries the Add into Close's view, so any
caller that observes c.workspaceStream is guaranteed to see the counter
already bumped before invoking Wait.
Adds TestClientClose_StreamWorkspaceAddWaitRace as a deterministic
regression: it waits for the server to observe the WS upgrade, then
fires Close in parallel with StreamWorkspace — pre-fix this triggers
the race within a handful of iterations under -race.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(agentctl): defer ctx cancel in StreamWorkspaceAddWaitRace iter
Wrap the per-iteration body in a runIter closure so `defer cancel()`
covers every exit path (timeout fatals as well as the normal flow),
replacing the explicit cancel() calls. Pure test hygiene — addresses
Claude review comment on PR #1272.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
* fix(lifecycle): drain StreamManager goroutines in tests
Goroutines spawned by `agentctl.Client.StreamUpdates` /
`StreamWorkspace` and by `lifecycle.StreamManager.connectWorkspaceStream`
could outlive the tests that created them on slow CI runners, causing
intermittent `goleak.VerifyTestMain` failures in
`internal/agent/runtime/lifecycle` even though every individual test
passed. The leak required the test to race: `client.Close()` could
return while a workspace dial was still in flight, leaving the
just-spawned WS read/write loops with nobody to drain them, and
`StreamManager.Wait()` only fired the drain when the external `stopCh`
had been closed first.
Make the close paths absolute drain barriers:
- `Client.Close()` now tracks every stream goroutine it spawns
(`streamWG` + a per-stream `WorkspaceStream` reference) and blocks
until all of them have exited. A `closed` flag flipped under the
client mutex makes subsequent `StreamUpdates` / `StreamWorkspace`
calls reject the dial, so a Close that races a dial in flight cannot
leave behind a stranded WS connection. `manager_interaction.go`
switches the restart path to per-stream Close helpers so it can keep
reusing the client after teardown.
- `StreamManager.Wait()` closes a new internal `waitCh` that the retry
backoff (`sleepOrStop`) and the connected `<-ws.Done() / <-stop>`
select also observe. This makes `Wait` an absolute drain barrier
regardless of whether the caller ever closes the external `stopCh`.
`streamContext` carries both stops through `stopChannelContext` so
in-flight WebSocket dials cancel on either signal.
Local stress (`make test-lifecycle-goleak LIFECYCLE_GOLEAK_COUNT=20`)
passes 20/20 with `-race` after the fix. The new
`TestClientClose_DrainsStreamGoroutines` covers the drain barrier and
the post-Close rejection path. Added a `test-lifecycle-goleak`
Makefile target as the repro hook for the CI-only flake.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(lifecycle): track mergeStops goroutine and guard workspace stream races
Address PR review feedback:
- stopChannelContext now registers the mergeStops goroutine on an optional
*sync.WaitGroup so sm.wg.Wait stays a true drain barrier even for the
connectUpdatesStream path (outer goroutine returns immediately after dial).
- StreamWorkspace re-checks workspaceStreamConn after dial to seal the
concurrent-callers race that could orphan a conn.
- readWorkspaceStream's defer now guards the workspaceStreamConn reset by
identity so an old read goroutine cannot zero out a newer conn pointer.
- Clarify comments on streamWG, Client.Close, and the mock.connected wait
in the close-barrier test.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* fix(lifecycle): drop agent-stream drain barrier to unblock workflow cascade
The previous Close drain barrier for the agent (updates) stream blocked
the on_turn_complete cascade flow. completeAndStopSession calls Close
synchronously on the current session's client, and Close was waiting on
streamWG (which tracked the read goroutine). The wait stalled the
cascade long enough that Step2's auto-started session never got its
profile name into the tab — surfacing as a flaky "Profile B tab not
visible" failure in workflow-agent-switch E2E shard 7.
Drain barrier is still in place for the workspace stream (where the
original goleak flake lived). Agent stream is closed but not awaited:
the cascade legitimately stops + restarts the agent stream on the same
client, and the workspace stream's existing barrier is sufficient for
the leak coverage.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Kandev Agent <agent@kandev.dev>
* fix(agents): emit --dangerously-skip-permissions for Claude CLI passthrough (#1261)
Claude ACP declared empty PermSettings in passthrough mode, so the launch path's
PermissionValues map (which carries the profile's DangerouslySkipPermissions
toggle) had no matching CLI-flag mapping to emit. Bypass / auto-approve was
silently ignored. Wire the permission setting to --dangerously-skip-permissions
and skip the key in seedCLIFlags so the toggle does not double-surface in the
curated cli_flags list.
* test(agents): use PermissionKeyDangerouslySkipPermissions constant in claude_acp_test
* test(agents): use constant in claude_acp_test failure message too
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>