SQLite TIMESTAMP columns store UTC without a zone marker. `Date.parse`
treats timezoneless ISO strings as local time, so on any non-UTC host
every claim and processAfter looks (TZ offset) hours stale. That makes
fresh claims trip the kill-claim path on the first sweep tick — every
container gets killed within seconds of spawn.
Two affected sites in host-sweep.ts:
- decideStuckAction reads claim.status_changed and computes claimAge.
On a TZ=Europe/Madrid host (UTC+2), a claim made 5s ago looks
7205s old and exceeds CLAIM_STUCK_MS (60s).
- The orphan retry loop reads msg.processAfter and skips messages
rescheduled into the future. On the same host, future timestamps
look (TZ offset) hours in the past, so the skip is missed and
tries gets bumped on every tick.
Fix: introduce parseSqliteUtc(s) which appends "Z" only when no zone
marker is present, then call it from both sites. Behavior under
TZ=UTC is unchanged.
Verified on a production v2 install on TZ=Europe/Madrid: with the
patch applied, an idle container survived 30+ minutes without being
killed (previously: killed within 60s of spawn).
Tests: 5 new cases covering the bare/Z/+offset/invalid input matrix
and a TZ-independence check. All 19 host-sweep tests pass and tsc
clears against main.
When the host kills a container (absolute-ceiling, claim-stuck, or crashed),
resetStuckProcessingRows reset messages_in but left orphan rows in
processing_ack. The next sweep tick spawned a fresh container and, on the
same tick, ran enforceRunningContainerSla against outbound.db that still
contained the previous container's claim with a hours-old status_changed
timestamp — instant kill-claim, before the agent-runner could open
outbound.db to run its own clearStaleProcessingAcks(). Loop until tries
hit MAX_TRIES.
Add deleteOrphanProcessingClaims() in session-db and call it at the end of
resetStuckProcessingRows. Safe to write outbound.db here because the host
only enters this path after killContainer (or when no container is running).
Tests in host-sweep.test.ts cover the helper plus the regression: orphan
claim from a 2h-old kill is now removed atomically with the messages_in
reset, so the next sweep tick sees an empty claims list and the freshly
respawned container survives long enough to start its agent-runner.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
decideStuckAction treated a missing heartbeat file as heartbeatAge =
Infinity, which always exceeded the 30-minute ceiling. Result: every
freshly-spawned container got killed within seconds of spawn on the
first sweep pass because it hadn't produced an SDK event yet (heartbeat
is only touched on SDK events inside processQuery, not on boot).
Skip the ceiling branch when heartbeatMtimeMs === 0. Containers that
genuinely never wrote a heartbeat because they died are caught by the
separate "container process not running" cleanup path. Containers that
boot, claim a message, but hang at the gate are caught by the
claim-stuck check below — which correctly fires regardless of heartbeat
presence once claimAge exceeds tolerance.
Updates the "absent heartbeat → kill-ceiling" test (which was encoding
the bug) and adds a companion that the claim-stuck path still fires for
absent-heartbeat containers with aged claims.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the two overlapping old mechanisms (30-min setTimeout kill in
container-runner, 10-min heartbeat STALE_THRESHOLD reset in host-sweep)
with message-scoped stuck detection anchored to the processing_ack claim
age + an absolute 30-min ceiling that extends for long-declared Bash
tools.
Old model problems:
- IDLE_TIMEOUT setTimeout fired on plain wall-clock time; slow-but-alive
agents got killed at 30min regardless of activity
- 10-min STALE_THRESHOLD in the sweep was unreliable — the heartbeat is
only touched on SDK events, so legitimate silent tool work (sleep 30,
long WebFetch, npm install) looked identical to a hung container
- Two overlapping sources of truth for "when to let go of a container"
New model:
- Host sweep is the single source of truth.
- Container exposes a new `container_state` single-row table in outbound.db
(schema added; container writes, host reads). PreToolUse hook writes
current_tool + tool_declared_timeout_ms (read from Bash's tool_input);
PostToolUse / PostToolUseFailure clear it.
- Sweep decides with a pure helper `decideStuckAction`:
* absolute ceiling — kill if heartbeat age > max(30min, bash_timeout)
* per-claim stuck — kill if any processing_ack row has claim_age >
max(60s, bash_timeout) AND heartbeat hasn't been touched since claim
* otherwise ok
Kill paths reset leftover processing rows with exponential backoff,
reusing the existing retry machinery.
Tool blocklist expanded:
- AskUserQuestion (SDK placeholder; we have mcp__nanoclaw__ask_user_question)
- EnterPlanMode, ExitPlanMode, EnterWorktree, ExitWorktree (Claude Code UI
affordances; would hang in headless containers)
PreToolUse hook is also defense-in-depth: if a disallowed tool name slips
through, it returns `{ decision: 'block' }` so the agent sees a clear
error instead of appearing stuck.
Removed:
- container-runner.ts: IDLE_TIMEOUT setTimeout, resetIdle callback on
activeContainers entry, resetContainerIdleTimer export.
- delivery.ts: the resetContainerIdleTimer call on successful delivery.
- poll-loop.ts: IDLE_END_MS + its setInterval. Keeping the query open is
cheaper than close+reopen (no cold prompt cache). Liveness is now a
host-side concern.
- host-sweep.ts: 10-min STALE_THRESHOLD_MS + getStuckProcessingIds in the
stale-detection path (still exported for kill reset).
Tests:
- src/host-sweep.test.ts — 9 tests for decideStuckAction covering: fresh
heartbeat, absolute ceiling, absent heartbeat, Bash-timeout extension
(both ceiling and per-claim), claim age below tolerance, heartbeat
touched after claim, unparseable timestamps.
Ref: docs/v1-vs-v2/ACTION-ITEMS.md items 9, 6a, 10.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
