The UsageAccumulator sums cacheRead/cacheWrite across all API calls
within a single turn. With Anthropic prompt caching, each call reports
cacheRead ≈ current_context_size, so after N tool-call round-trips the
accumulated total becomes N × actual_context, which gets clamped to
contextWindow (200k) by deriveSessionTotalTokens().
Fix: track the most recent API call's cache fields separately and use
them in toNormalizedUsage() for context-size reporting. This makes
/status Context display accurate while preserving accumulated output
token counts.
Fixes#13698Fixes#13782
Co-authored-by: akari-musubi <259925157+akari-musubi@users.noreply.github.com>
* fix: prevent FD leaks in child process cleanup
- Destroy stdio streams (stdin/stdout/stderr) after process exit
- Remove event listeners to prevent memory leaks
- Clean up child process reference in moveToFinished()
- Also fixes model override handling in agent.ts
Fixes EBADF errors caused by accumulating file descriptors
from sub-agent spawns.
* Fix: allow stdin destroy in process registry cleanup
---------
Co-authored-by: Tak Hoffman <781889+Takhoffman@users.noreply.github.com>
Two fixes for Google Antigravity (Cloud Code Assist) reliability:
1. Forward-compat model fallback: pi-ai's model registry doesn't include
claude-opus-4-6-thinking. Add resolveAntigravityOpus46ForwardCompatModel()
that clones the opus-4-5 template so the correct api ("google-gemini-cli")
and baseUrl are preserved. Fixes#13765.
2. Fix thinking.signature rejection: The API returns Claude thinking blocks
without signatures, then rejects them on replay. The existing sanitizer
strips unsigned blocks, but the orphaned-user-message path in attempt.ts
bypassed it by reading directly from disk. Now applies
sanitizeAntigravityThinkingBlocks at that code path.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Move appendCacheTtlTimestamp() to after prompt + compaction retry
completes instead of before. The previous placement inserted a custom
entry (openclaw.cache-ttl) between compaction and the next prompt,
which broke pi-coding-agent's prepareCompaction() guard — the guard
only checks if the last entry is type 'compaction', and the cache-ttl
custom entry made it type 'custom', allowing an immediate second
compaction at very low token counts (e.g. 5,545 tokens) that nuked
all preserved context.
Fixes#9282
Relates to #12170
After a successful launchctl kickstart, the stdout.write() for the
status message may fail with EPIPE if the receiving end has already
closed. Catch and ignore EPIPE specifically; re-throw other errors.
Closes#14234
Co-authored-by: Echo Ito <echoito@MacBook-Air.local>
* fix(gateway): drain active turns before restart to prevent message loss
On SIGUSR1 restart, the gateway now waits up to 30s for in-flight agent
turns to complete before tearing down the server. This prevents buffered
messages from being dropped when config.patch or update triggers a restart
while agents are mid-turn.
Changes:
- command-queue.ts: add getActiveTaskCount() and waitForActiveTasks()
helpers to track and wait on active lane tasks
- run-loop.ts: on restart signal, drain active tasks before server.close()
with a 30s timeout; extend force-exit timer accordingly
- command-queue.test.ts: update imports for new exports
Fixes#13883
* fix(queue): snapshot active tasks for restart drain
---------
Co-authored-by: Elonito <0xRaini@users.noreply.github.com>
Co-authored-by: Tak Hoffman <781889+Takhoffman@users.noreply.github.com>
* fix: exclude maxTokens and token-count fields from config redaction
The /token/i regex in SENSITIVE_KEY_PATTERNS falsely matched fields like
maxTokens, maxOutputTokens, maxCompletionTokens etc. These are numeric
config fields for token counts, not sensitive credentials.
Added a whitelist (SENSITIVE_KEY_WHITELIST) that explicitly excludes
known token-count field names from redaction. This prevents config
corruption when maxTokens gets replaced with __OPENCLAW_REDACTED__
during config round-trips.
Fixes#13236
* fix: honor deleteAfterRun for one-shot 'at' jobs with 'skipped' status
Previously, deleteAfterRun only triggered when result.status was 'ok'.
For one-shot 'at' jobs, a 'skipped' status (e.g. empty heartbeat file)
would leave the job in state but disabled, never getting cleaned up.
Now deleteAfterRun also triggers on 'skipped' status for 'at' jobs,
since a skipped one-shot job has no meaningful retry path.
Fixes#13249
* Cron: format timer.ts
---------
Co-authored-by: nice03 <niceyslee@gmail.com>
Co-authored-by: Tak Hoffman <781889+Takhoffman@users.noreply.github.com>
- Guard against undefined/empty token in buildGatewayAuthConfig
- Automatically generate random token when token param is undefined, empty, or whitespace
- Prevents JSON.stringify from writing literal string "undefined" to config
- Add tests for undefined, empty, and whitespace token cases
Fixes#13756
Co-authored-by: Klawd Asklee <klawdebot@gmail.com>
* fix(gateway): handle async EPIPE on stdout/stderr during shutdown
The console capture forward() wrapper catches synchronous EPIPE errors,
but when the receiving pipe closes during shutdown Node emits the error
asynchronously on the stream. Without a listener this becomes an
uncaught exception that crashes the gateway, causing macOS launchd to
permanently unload the service.
Add error listeners on process.stdout and process.stderr inside
enableConsoleCapture() that silently swallow EPIPE/EIO (matching the
existing isEpipeError helper) and re-throw anything else.
Closes#13367
* guard stream error listeners against repeated enableConsoleCapture() calls
Use a separate streamErrorHandlersInstalled flag in loggingState so that
test resets of consolePatched don't cause listener accumulation on
process.stdout/stderr.
When internal tools (e.g. TTS) emit MEDIA:/tmp/... with absolute paths,
isValidMedia() correctly rejects them for security. However, the rejected
MEDIA: line was kept as visible text in the output, leaking the path to
the user.
Now strip MEDIA: lines that look like local paths even when the path
is invalid, so they never appear as user-visible text.
Closes#14365
Co-authored-by: Echo Ito <echoito@MacBook-Air.local>
When the gateway is installed as a macOS launch agent and no token is
configured, the service enters an infinite restart loop because launchd
does not inherit shell environment variables. Auto-generate a token
during `gateway install` when auth mode is `token` and no token exists,
matching the existing pattern in doctor.ts and configure.gateway.ts.
The token is persisted to the config file and embedded in the plist
EnvironmentVariables for belt-and-suspenders reliability.
Relates-to: #5103, #2433, #1690, #7749
* fix(whatsapp): convert Markdown bold/strikethrough to WhatsApp formatting
* refactor: Move `escapeRegExp` utility function to `utils.js`.
---------
Co-authored-by: Luna AI <luna@coredirection.ai>
* fix(cron): pass agentId to runHeartbeatOnce for main-session jobs
Main-session cron jobs with agentId always ran the heartbeat under
the default agent, ignoring the job's agent binding. enqueueSystemEvent
correctly routed the system event to the bound agent's session, but
runHeartbeatOnce was called without agentId, so the heartbeat ran under
the default agent and never picked up the event.
Thread agentId from job.agentId through the CronServiceDeps type,
timer execution, and the gateway wrapper so heartbeat-runner uses the
correct agent.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* cron: add heartbeat agentId propagation regression test (#14140) (thanks @ishikawa-pro)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Tak Hoffman <781889+Takhoffman@users.noreply.github.com>
Previously, if one cron job had a malformed schedule expression (e.g. invalid cron syntax),
the error would propagate up and break the entire scheduler loop. This meant one misconfigured
job could prevent ALL cron jobs from running.
Changes:
- Wrap per-job schedule computation in try/catch in recomputeNextRuns()
- Track consecutive schedule errors via new scheduleErrorCount field
- Log warnings for schedule errors with job ID and name
- Auto-disable jobs after 3 consecutive schedule errors (with error-level log)
- Clear error count when schedule computation succeeds
- Continue processing other jobs even when one fails
This ensures the scheduler is resilient to individual job misconfigurations while still
providing visibility into problems through logging.
Co-authored-by: Marvin <numegilagent@gmail.com>
* fix(cron): re-arm timer when onTimer fires during active job execution
When a cron job takes longer than MAX_TIMER_DELAY_MS (60s), the clamped
timer fires while state.running is still true. The early return in
onTimer() previously exited without re-arming the timer, leaving no
setTimeout scheduled. This silently kills the cron scheduler until the
next gateway restart.
The fix calls armTimer(state) before the early return so the scheduler
continues ticking even when a job is in progress.
This is the likely root cause of recurring cron jobs silently skipping,
as reported in #12025. One-shot (kind: 'at') jobs were unaffected
because they typically complete within a single timer cycle.
Includes a regression test that simulates a slow job exceeding the
timer clamp period and verifies the next occurrence still fires.
* fix: update tests for timer re-arm behavior
- Update existing regression test to expect timer re-arm with non-zero
delay instead of no timer at all
- Simplify new test to directly verify state.timer is set after onTimer
returns early due to running guard
* fix: use fixed 60s delay for re-arm to prevent zero-delay hot-loop
When the running guard re-arms the timer, use MAX_TIMER_DELAY_MS
directly instead of calling armTimer() which can compute a zero delay
for past-due jobs. This prevents a tight spin while still keeping the
scheduler alive.
* style: add curly braces to satisfy eslint(curly) rule
The `computeNextRunAtMs` function used `nowSecondMs - 1` as the
reference time for croner's `nextRun()`, which caused it to return the
current second as a valid next-run time. When a job fired at e.g.
11:00:00.500, computing the next run still yielded 11:00:00.000 (same
second, already elapsed), causing the scheduler to immediately re-fire
the job in a tight loop (15-21x observed in the wild).
Fix: use `nowSecondMs` directly (no `-1` lookback) and change the
return guard from `>=` to `>` so next-run is always strictly after
the current second.
Fixes#14164
Replace the full manual mock with importOriginal spread so new SDK
exports are available automatically. Only ChannelType, MessageType,
and Client are overridden — the rest come from the real module.
Prevents CI breakage when @buape/carbon adds exports (e.g. the
recent StringSelectMenu failure that blocked unrelated PRs).
Closes#13244
* fix(cli): exit with non-zero code when configure/agents-add wizards are cancelled
Follow-up to the onboard cancel fix. The configure wizard and
agents add wizard also caught WizardCancelledError and exited with
code 0, which signals success to callers. Change to exit(1) for
consistency — user cancellation is not a successful completion.
This ensures scripts that chain these commands with set -e will
correctly stop when the user cancels.
* fix(cli): make wizard cancellations exit non-zero (#14156) (thanks @0xRaini)
---------
Co-authored-by: Rain <rain@Rains-MBA-M4.local>
Co-authored-by: Sebastian <19554889+sebslight@users.noreply.github.com>
