[nissa-seru]

A turn that the framework drops before it sends anything now ends with inference:aborted. Before, it emitted inference:started and then no terminal event, so a module that brackets a turn between the two, such as a typing indicator, kept the dropped turn open. The Discord module's typing keepalive now also stops at a turn's normal end, inference:completed, which onTrace's documentation had left out.

The gap

beginAgentTurn emits inference:started, then runs the modules' gatherContext and the MCPL beforeInference hooks, then compiles. Both waits can be long, so the framework checks the turn twice before it sends anything:

• before the compile: whether the agent was replaced or removed, or the framework is stopping;
• after the compile: the same, and whether the framework was stopped and started again.

Either check drops the turn, and both returned with no terminal event. The check before the compile also logged nothing.

onTrace's documentation (src/types/module.ts) names the events a module ends a turn's bracket on: inference:turn_ended, inference:aborted, inference:failed and inference:exhausted. It left out the normal end, inference:completed, which driveStream emits once, when the turn's stream completes. Tool-call cycles stay inside that stream, so it doesn't fire between them. The Discord module in this repository follows the documentation. Its typing keepalive stops at any of the four, and otherwise goes on refreshing until the agent's next inference:started or TYPING_MAX_MS, five minutes. So with autoTyping on, its default, typing went on after every ordinary reply, and after a dropped turn too.

The change

• Each drop logs a line and ends the turn with inference:aborted. Its reason is the logged line's text without the [turn] <agent>: prefix, for example "a turn compiled while the framework was stopping is dropped before it sends anything". Its durationMs runs from this start of the turn, as the other inference:aborted events count from their own start. Nothing failed and nothing was sent, so the turn ends as aborted rather than failed, as a turn abandoned at an operator quiesce already does (reason: 'quiesce_abandoned'). lastInference.endedAt is recorded too, as driveStream records it, so the health snapshot no longer shows a dropped turn as still running. It is recorded only while the record is still this turn's: an ephemeral run disposed while its turn waited has had its record deleted, and its name gets none back.
• Unless a newer turn of the agent's name holds the turn token. That happens when the turn was dropped because the framework stopped and started again, or because its agent object was replaced, and the new lifecycle or the replacement has begun a turn of its own meanwhile. The terminal event and lastInference carry only the agent's name, so the dropped turn's would end that newer turn: its typing loop would stop while it runs. The dropped turn then emits and records nothing, as before.
• The Discord module's typing keepalive stops at inference:completed too, and onTrace's documentation names it as the normal end. The documentation also says what else a bracket meets: a context-budget restart emits inference:started again with no terminal event for the stream it replaced, and a failing turn can emit inference:failed and then inference:exhausted.
• The web UI in ui/ now also returns an agent to idle on inference:turn_ended, inference:aborted and inference:exhausted, where it had looked only at inference:completed and inference:failed.
• inference:aborted's reason and durationMs get doc comments, since reason now has three sources: an operator's abort, a quiesce, and a dropped turn.

The provider-wait hold after the compile already ends its turn with inference:failed, as the failure sites' holds do, and this change leaves it alone. Ending it as aborted would leave connectome-host's activity module reading every held turn as running (below).

Listeners elsewhere that handle only some of the terminal events still read a dropped turn as running. connectome-host's activity module ends its indicator on inference:completed, inference:failed and inference:exhausted, and ignores inference:aborted, which this change now emits for a dropped turn. The drop's stopTyping() clears its indicator, but its own state still reads the turn as running until the agent's next end event, so a channel added to it with show_in meanwhile starts typing. Its owners are building that change.

Tests

In test/tool-result-guard.test.ts, each test records the agent's inference:* events as a module's listener sees them:

• "a framework compile from before stop cannot drive after restart", the drop after the compile when the framework has stopped and started again, checks that the turn gets exactly one terminal event, inference:aborted, with its reason.
• Ten tests drop a turn at each check, before the compile (the turn waits in a module's gatherContext) and after it (it waits in its compile). Each loses its place in one of five ways: the framework stops, the agent is removed, its run is disposed with its record and token (as runEphemeralToCompletion's disposal does), the agent object is replaced under its name, or the replacement's own turn already holds the name's turn token.
  • In the first four ways, the turn gets exactly one terminal event, inference:aborted, with the exact reason, and its durationMs covers the time it was held. lastInference.endedAt is recorded, except for the disposed run, whose name gets no record back.
  • In the fifth, the dropped turn emits nothing after inference:started, and records no end.

In test/discord-typing-keepalive.test.ts, for each of the five terminal events, a turn's keepalive shows typing at inference:started, refreshes once while the turn runs, and refreshes no more after the event. The module's interval runs on mocked timers.

Evidence

At 9d81e5d9, on main cd036e0d, with context-manager 0.14.0 and membrane 0.5.87 installed:

• The full suite, each test file in its own process, after a clean tsc and build. With --test-force-exit, as CI runs it, Node 20.20.2 passes 2,154/2,154 across 144 files. On Node 24.21.0 the forced runner counted 2,146, all passing: it cut two files short (history-module-realstore 22/28, history-module-ux 24/26), and without the force every file passes in full, 2,154/2,154.
• Removals in the drop code. Each of eight pieces, removed alone, fails the drop tests aimed at it (Node 24.19.0; the first seven at 3f4cdb64, before the last commit):
  • the turn-token check, so a dropped turn ends a newer turn's bracket: the two newer-turn tests;
  • the terminal event: seven of the nine drop tests;
  • the lastInference end, and durationMs set to 0: six each;
  • the drop before the compile ending nothing: its three tests;
  • the drop after the compile ending nothing: its three and the restart test;
  • the removal reason in place of the stopping one: the stopping test;
  • the record written whatever turn it belongs to: the two disposed-run tests, and only they (Node 24.19.0, and Ruth-2424's run on Node 24.21.0).
• The Discord keepalive: with inference:completed taken out of TURN_TERMINAL_TRACES, exactly the test's inference:completed case fails.

At this head, 2d090e3f, which merges main 92125b07: the merge's one conflict was the drop before the compile, where main now ends the dropped turn's routing (endDroppedTurn). The merge keeps both, in the order the drop after the compile already had: the reason and its log line, endDroppedTurn, then the typing stop, the event gate's end and the terminal event. The change against main is the same 9 files, +257 −12, as before the merge. Each file run on its own, the suite passes 2,244/2,244 across 145 files on Node 24.21.0, forced and unforced, and on Node 20.20.2 forced, with none cut short. Typecheck and build are clean.

Not covered

• A dropped turn still stops typing on every channel and ends the agent's event-gate liveness by name, even when a newer turn of the name is running. Both predate this change. A retry of a replaced agent removes the new agent's turn token and resets its routing state #327, a retry of a replaced agent, is the same family and has its own fix.

Ruth-2424 reviewed it, and the merge of main. Agnes-1009's probe found that onTrace's documentation left out the normal end.

— Ada-1017 (Claude Opus 5.5)

[nissa-seru]

@antra-tess @Anarchid Could you review this? It closes two gaps in how a turn's bracket ends. A turn the framework drops before it sends anything now ends with inference:aborted, and the in-repo Discord module's typing keepalive now also stops at a turn's normal end, inference:completed, which onTrace's documentation had left out. The description has the evidence, and it lists what this doesn't cover.

We'll merge it not before 2026-10-15 12:00 UTC unless you'd like changes. #333 also changes the two drop sites, and is due to merge first, so this branch will take main with both changes before it merges.

— Ada-1017

[greptile-apps]

Confidence Score: 4/5

[Medium impact] Fixes turn lifecycle events for dropped inferences and typing indicators.

Appears safe to merge, with a non-blocking cleanup issue for timed-out ephemeral agents.

Findings

1.  Disposed agents regain records ▶

Fix with agent prompt

### Issue 1
src/framework.ts:11149
When an ephemeral agent's idle watchdog expires during `gatherContext` or compilation, `runEphemeralToCompletion` removes its `lastInferenceAt` entry and turn token. If the old wait finishes later, this write restores an orphan `{ endedAt }` entry because the missing token passes the guard. No later cleanup removes it, so long-running hosts retain records for those timed-out subagents. Keep the terminal event, but only update a record that still belongs to this turn.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds inference:aborted for dropped turns, stops Discord typing at normal completion, and returns the UI to idle on the remaining end events.

• Dropped turns now end with an abort event.
• Discord typing stops when a turn completes normally.

_{Reviews (1) · Last reviewed commit: "fix(discord): the typing keepalive stops..." · Reviewed by Greptile}

[greptile-apps]

Disposed agents regain records

When an ephemeral agent's idle watchdog expires during gatherContext or compilation, runEphemeralToCompletion removes its lastInferenceAt entry and turn token. If the old wait finishes later, this write restores an orphan { endedAt } entry because the missing token passes the guard. No later cleanup removes it, so long-running hosts retain records for those timed-out subagents. Keep the terminal event, but only update a record that still belongs to this turn.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 11149

Comment:
**Disposed agents regain records**

When an ephemeral agent's idle watchdog expires during `gatherContext` or compilation, `runEphemeralToCompletion` removes its `lastInferenceAt` entry and turn token. If the old wait finishes later, this write restores an orphan `{ endedAt }` entry because the missing token passes the guard. No later cleanup removes it, so long-running hosts retain records for those timed-out subagents. Keep the terminal event, but only update a record that still belongs to this turn.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[nissa-seru]

Addressed in 9d81e5d. The drop now ends the lastInference record only while it's still this turn's, the one beginAgentTurn stamped with this turn's startedAt, so a name whose ephemeral run was disposed meanwhile gets no orphan { endedAt } back. The terminal event still goes out. The dropped-turn tests gain a disposed-run case at both checks, and with the guard taken out, those two fail and the others pass.

— Ada-1017

[nissa-seru]

Pushed 2d090e3f, which merges main (92125b07) into this pull request. The one conflict was the drop before the compile, where main now ends the dropped turn's routing. The merge keeps both, in the order the other drop already had. The change against main is otherwise as before. The suite passes 2,244/2,244 at the merge on Node 24 and Node 20, and the description's Evidence says so.

— Ada-1017