fix(messaging): bind completion delivery to the run's own destination

[JSap0914]

A Slack mention landed in #ceo_lead_align_input at 2026-09-11 20:10 KST while an unrelated web run was working. Three minutes later that web run's answer — an internal ReelBrain ticket summary — was posted into the mention's thread. The run had never been addressed to Slack at all.

What was wrong

A channel forwarder answered "where does this answer go?" at send time, by reading the per-channel last-active slot:

const target = opts.getLastTarget();   // getLastActiveTarget('slack')

That slot holds whoever spoke to the bot most recently. The mention overwrote it (src/slack/bot.ts:1508), the web run exited, and its agent_done — which carried origin and text but no destination — followed the slot into a conversation that had asked something else. The DB shows the split plainly: the posted body is stored under the web session b25ca0b8, not the mention session 6c6e4301.

Heartbeat jobs had the same shape one layer down. A job with no destination fell through to sendChannelOutput({channel:'active'}); a Slack destination with a channel but no thread posted to the channel root, indistinguishable from a form nobody finished; an empty-string threadId passed validation and then read as falsy at send time. #437 closed the malformed case and left the rest.

And authorizeExplicitTarget vouched for an explicitly addressed channel-root send by returning the last-active target — thread included. A caller that named its destination correctly still had it moved.

What this does

The destination travels with the run. src/messaging/run-pin.ts captures origin, requestId, scope, sessionId, remoteKey and the admitted target once at spawn, and every agent_done carries it. resolveForwarderTarget reads the destination off the event and returns null when it is absent or names another channel; null means do not send. getLastTarget/getLastChatId are removed from all three forwarder option types so the lookup cannot come back.

Heartbeat destinations are complete or held. Complete means a thread, or an explicit scope: "channel_root". Held means the tick logs a reason and sends nothing, GET /api/heartbeat surfaces it, and PUT refuses to write a new half-filled Slack destination while still inheriting an existing one. Authorization now decides whether a send is allowed without rewriting where it lands.

Behaviour this removes

Web and CLI turns are no longer mirrored into chat rooms. That was the forwarder's stated purpose, and it is the bug: "the conversation the user is watching" was only ever a guess from a global slot any concurrent conversation could move. A turn with no remote destination is not homeless — its answer is already on the surface that asked for it. Affects Classic/web answers appearing in Slack, terminal jaw results appearing in Slack, Discord forwardAll, and the Telegram getLastChatId fallback.

Live heartbeat jobs with an incomplete destination stop delivering until migrated. On this install five of nine jobs name a channel with no thread, one of them enabled (hb_daily_scrum_morning). They need scope: "channel_root" or a thread in heartbeat.json before this ships, or the next run is silent. Mention-watch jobs are unaffected — they answer the thread they find.

Testing

root,unit = 13458 tests, 13436 pass. Two failures (P13 TERM-ignoring owned handles, Pi model discovery) pass in isolation on this branch and on the base commit; they are load-sensitive under full-suite concurrency, not caused by this change. npm run gate:all — 23/23.

New: tests/unit/heartbeat-destination-binding.test.ts (HDB-001..008). Updated to the new contract: runtime-messaging-isolation (a target-less final now reaches no channel; a pinned one reaches exactly one), heartbeat-runner-modes, heartbeat-file, send-validation, and the Slack/Discord/Telegram forwarder fixtures, which now put the destination on the event instead of supplying it through a lookup.

Also repairs codex-app-multiplex-spawn, which was already failing on dev: its partial namedExports mock of runtime/events.js replaces the whole module, so projection.ts could not import recordRuntimeProjectionLoss and the file threw before its first assertion.

Not in this PR

#743's fail-closed completion matcher. matchesSlackReply still treats an absent field as agreement, and flipping it before request_settled, queued_run_started and queue_update carry the same identity would silence #655 steer tracking and #407 boot drain. This PR ships the additive half — every agent_done now carries the full pin — which is that change's precondition.

The progress-card retry storm is #744, sent separately.

Refs #742 #745 #437

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 89650401-b43c-4f89-bc2a-a10a132eee7f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}