shouldDeferParentWakeForSessionHistory previously had only one escape
path from the defer state: stale pending tool call. If the assistant
had unfinished text but no pending tool call (session crashed
mid-stream, model errored after partial text, network died), the
escape never fired and parent-wake deferred forever. Background-agent
completions never woke the parent.
Add a second escape: when the assistant text blocks but no tool wait
is pending, dispatch the wake after toolCallDeferMaxMs anyway. The
prompt-async-gate still defends if the assistant text turns out to be
live; we just stop deferring indefinitely.
Closes pre-publish blocker V11.
In v4.1.0+, users observed duplicate assistant streams rendering the same
content in two languages simultaneously (e.g. Chinese + English), most often
at the end of a turn.
Root cause: ParentWakeNotifier.requeueWake() unconditionally requeued ANY
wake that arrived during the background-agent-parent-wake post-dispatch
hold window. When a duplicate completion edge fired during that hold, the
same wake was replayed after the hold expired, triggering a second prompt
dispatch and a parallel assistant stream.
The fix compares the new wake against dispatchedParentWakes.get(sessionID)
and drops identical wakes during the gate hold, while preserving the existing
requeue behavior for genuinely-new wakes and failed-dispatch retries.
Regression test added in given/when/then style covering the duplicate-during-
hold scenario (TDD red-then-green).
Fixes#4256Fixes#4019
Treat resolved abort responses with a non-null error payload the same as rejected aborts. This prevents stale-timeout cancellation bookkeeping from reporting success when the child session was not actually aborted.
Plan: .omo/plans/subagent-timeout-active-output.md
Convert OpenCode v2 session.next stream events into the existing message part activity path so child sessions that are still producing text, reasoning, or tool output refresh lastUpdate before stale polling runs.
This keeps the timeout poller from cancelling active subagents and preserves tool-call progress for session.next.tool.called events.
Plan: .omo/plans/subagent-timeout-active-output.md
Cubic caught that clearing recent parent activity on session.idle could reopen the stale-idle race. Keep the fresh activity marker until its normal expiry so background completions still defer while the parent turn is actively reasoning.
Red: parent idle after fresh reasoning delta dispatched an unexpected parent wake. Green: focused parent-wake regression suite passes.
Record fresh parent session message activity before parent-wake flushing so stale idle status cannot dispatch a background completion into a live reasoning turn.
Add regression coverage for the Discord 4.2.3/OpenCode 1.15.5 duplicate-branch repro shape where a parent reasoning delta arrives before the background all-complete wake.
Refs #4212
Refs #4019
Refs #3774
Plan: plans/background-notification-active-turn-queue.md
Require the latest assistant tool-wait turn itself to be stale before a shouldReply parent wake can bypass tool-call deferral. This prevents an all-complete background wake from forking a second parent prompt loop when OpenCode has repaired the tail to a synthetic user message.
Tests:
- bun test src/features/background-agent/parent-wake-user-message-race.test.ts src/features/background-agent/task-completion-cleanup.test.ts src/hooks/shared/prompt-async-gate.test.ts src/shared/prompt-async-route-audit.test.ts --bail
- bun run typecheck
- bun test
Background fallback retry notifications were queued as bare internal user messages, so OpenCode could treat the notification as a new default-agent turn. Reuse the same parent prompt context resolver used by completion notifications for retrying and retry-ready wakes, and pin regression coverage for Hephaestus parent sessions plus missing-context fallbacks.
When a background subagent emits [ALL BACKGROUND TASKS COMPLETE], the
plugin queues a parent-wake that ultimately calls
dispatchInternalPrompt against the parent session. If the user submits
a new prompt inside the ~250 ms post-dispatch hold window, both writes
land on the same OpenCode session-storage file at the same instant.
OpenCode's @parcel/watcher (which the plugin itself does not depend on,
but does indirectly trigger) batches those events into a TSFN callback
and dispatches them into a JS env that the renderer has just torn down
because the session view re-mounted around the user's new message ->
napi_fatal_error / SIGABRT on macOS arm64. Removing the plugin removes
the parent-wake, which is why removing OmO eliminates the crash.
Mitigation:
- Before flushPendingParentWake calls dispatchInternalPrompt, inspect
the parent session's message tail. If the most recent message is a
user message added inside PARENT_WAKE_USER_MESSAGE_IN_PROGRESS_WINDOW_MS
(default 2_000 ms), reschedule instead of dispatching. The user's own
prompt will drive the model; queued notifications will be re-flushed
on the next idle.
- Best-effort unref() of the long-lived pending-retry and dispatched-
wake bookkeeping setTimeouts. They previously pinned the host event
loop and prolonged the teardown window during which the watcher race
can fire.
The new option userMessageInProgressWindowMs is wired through
BackgroundManager via a module-level constant and is independently
testable.
Regression test parent-wake-user-message-race.test.ts covers:
- fresh user message -> dispatch deferred
- latest message is assistant -> dispatch proceeds
- user message older than window -> dispatch proceeds
- window=0 disables the guard
This is a surface-level mitigation of the most-likely root cause from
the audit; a deeper fix (singleton guard against plugin
double-instantiation under @opencode-ai/plugin@local reload, dispose
lifecycle for OpenCode plugin reload) is out of scope here.
Root cause of the user-visible `/init-deep ulw` hang (session
`ses_1cb9c3013ffesUOy5H3QOIya4K`): the plugin's `unhandledRejection` and
`uncaughtException` listeners were calling
`scheduleForcedExit(handler(error), 1, true)`, which both ran the entire
`cleanupAll()` chain (BackgroundManager shutdown, tmux pane closure,
team-mode teardown) and then `process.exit(1)`'d the host. Under heavy
slash commands like `/init-deep ultrafucking deep`, a single mid-stream
error (e.g. opencode's own `session.processor` Aborted-process condition,
or a transient socket reset) would:
1. trigger the listener,
2. abort the in-flight background tasks (`session.error
MessageAbortedError` for both child sessions in the log),
3. close the tmux panes the user was watching,
4. immediately kill the host via `process.exit(1)`.
From the user's seat that looked like a frozen TUI, which is what they
reported as "ulw 여전히 멈추는데". The error blob also logged as `{}`
because `JSON.stringify(new Error(...))` strips non-enumerable Error
fields, so the previous log line carried no diagnostic value.
This change makes the global `uncaughtException` / `unhandledRejection`
listeners log-only:
* New `describeProcessCleanupError()` extracts `{name, message, stack}`
from Error instances, falls back to a structured `{raw: ...}` payload
for plain objects / primitives, so the log now actually says what
failed.
* `registerErrorEvent()` no longer runs cleanup and no longer calls
`scheduleForcedExit`. It detaches itself, logs a single explanatory
line, and returns. Bun's default crash behaviour is already suppressed
for these events when a listener is present, so the host now genuinely
survives transient streaming errors instead of being killed by our
own helper.
* Signal handlers (`SIGINT` / `SIGTERM` / `SIGBREAK` / `beforeExit` /
`exit`) keep their existing behaviour and still run `cleanupAll()`
before the host terminates — that is now the only path that tears
down background tasks and tmux panes.
Tests are updated to lock in the new contract:
* New regression `#given scheduleForcedExit enabled AND unhandledRejection
fires #when the listener runs #then process.exit is NOT called AND
process.exitCode stays 0 AND no cleanup runs` (and the
uncaughtException twin) re-enables `scheduleForcedExit`, spies on
`process.exit` plus `globalThis.setTimeout`, and asserts none of them
are touched. Without the fix this test failed exactly like the
observed hang (exit called once, exitCode set to 1).
* The existing "manager shuts down before process exits" tests are
rewritten to assert the opposite: cleanup is NOT invoked from the
error path.
* A complementary `'exit'` listener test pins the real shutdown
contract (`exit` event still triggers `cleanupAll`).
* A new `#given describeProcessCleanupError` block covers the four
shapes (Error, plain object with own fields, empty object, primitive).
* The unregister assertion is tightened to check the listener count
drops back to baseline (previously it relied on a side effect of the
old cleanup-on-error path).
Manual QA: `bun /tmp/process-cleanup-smoke-test.ts` (out-of-tree smoke
driver) emits five back-to-back unhandledRejection/uncaughtException
events with various payload shapes and prints
`SMOKE_TEST_OK survived 5 emissions; exitCode=0; shutdownInvocations=0`,
confirming the host survives and no spurious cleanup runs.
`bun test` runs green for the affected modules:
- src/features/background-agent (528 tests)
- src/create-managers + src/plugin (218 tests)
- src/hooks/{unstable-agent-babysitter,ralph-loop,keyword-detector}
(225 tests)
YeonGyu-Kim