fix(routing): an ephemeral run takes no global fallback

[nissa-seru]

Problem

A locus or hybrid ephemeral run (a subagent) with no channel of its own posts its final plain prose to whichever channel last received a message. runEphemeralToCompletion queues the run's request with no channelId. Its turn then freezes its locus through ChannelRegistry.resolveLocus, which falls back to the process-global defaultPublishChannel, the channel of the most recent delivered inbound message. Delegated work, possibly drawn from a DM or a private channel, can land in an unrelated and possibly public channel. connectome-host passes a parent's proseRouting to its subagents so that children of a disabled caller "cannot leak prose into its live channel". A child of a locus caller had no such protection.

Fixes #551.

Changes

No ephemeral run takes the global fallback, whatever its prose routing. A subagent's plain prose is its caller's result. So its locus, its channel context's default and its channel_publish default are a channel of its own or nothing, never a channel that only the latest inbound message picked.

• One option on the channel registry, takesGlobalFallback(conversationId). The framework answers false for any name createEphemeralAgent reserved. It goes by name, so a speech dispatch after the run's release still counts. A conversation fork's name is reserved in the same set, and it keeps the fallback: the set now records which kind reserved each name.
• Where the registry reads it:
  • resolveLocus takes no fallback for such a run, so a turn with no channel of its own has no locus.
  • deliverSpeech returns that null locus as not delivered, with its own reason, since the prose is its caller's result. It writes no [routeSpeech] … NOT routed line, no mcpl:speech-route-failed trace and no [send-undeliverable] notice.
  • buildChannelContext gives such a run no incoming message, and no default outgoing channel unless it opened one.
  • channel_publish with no channelId is refused for such a run: "Nothing was published: name the channel with channelId. A subagent has no default channel, since its plain speech is returned to whoever called it, so channel_publish sends only to a channel it names."
• The tool's description now says where an omitted channelId goes: the conversation's bound channel, else the most recent incoming one, and nowhere for a subagent. The old text was already wrong for conversation forks, which publish to their bound channel.
• The refusal reaction and the refusal's failures.log record read the refused agent's own channel context. So a subagent's refusal marks no message and logs locus: null.

What a run chooses still routes: a channel it opens itself, and a destination it names in its prose (hybrid's >>>, explicit's >>). Residents and conversation forks keep the fallback.

What callers newly see:

• For every ephemeral run, whatever its prose routing:
  • channel_publish without a channelId returns the refusal above instead of posting to the latest inbound channel;
  • its channel context (beforeInference's channels) has no incoming, and no defaultOutgoing unless it opened a channel;
  • a refusal of its inference adds no reaction anywhere, and its failures.log record has locus: null.
• For a locus or hybrid run with no channel of its own, its final plain prose isn't published. runEphemeralToCompletion returns it as before. A sleep announcement it makes isn't published either.
• For every agent, channel_publish's description and its channelId description have changed.

What each part is for

• The option, rather than a check of the framework's run state inside the registry, keeps the registry free of the framework's agent bookkeeping, as homeChannelResolver and activeChannelResolver already do.
• Recording the reserving kind ('run' | 'conversation') keeps conversation forks unchanged. Without it, a fork would be treated as a run: its speech with no locus would stop raising its failure notice, and its channel context would lose incoming. That's outside this fix.
• The quiet outcome in deliverSpeech is needed because, without it, every subagent turn with no channel of its own would write the route-failure line, trace and notice that a resident's turn with no locus gets. Every speech delivery reaches deliverSpeech, through routeSpeech or directly, so it sits there once. PublishOutcome has no status for "not delivered, and not a failure", so the outcome is failed with its own reason. routeSpeech turns every non-delivery into null, and nothing records it.
• The channel_publish refusal closes the same default from the tool side. Without it, a subagent that omits channelId would still post to the latest inbound channel.
• The refusal reaction's and refusal log's agent keep a subagent's refusal from being marked on, or logged against, an unrelated channel's latest message.
• Considered and left out: running every ephemeral run as 'disabled'. That would also stop a channel the run opens itself, and hybrid's >>> envelopes, which callers were told that mode routes. The defect is a destination nobody chose, and those are chosen.

Tests

CI on cba2011f merged with main 76623a3 (run 39207861926) passes all four Build & Test legs and the packed install, and the changelog check passes too. npm test on Ubuntu runs 4,756 tests with Node 20 and with Node 24, and on macOS 4,758 with each: 0 fail and 27 skipped everywhere. That's main 76623a3's totals (run 39205375038: 4,744 and 4,746, 27 skipped) plus this PR's 12 tests. The three files CI runs without forced exit pass 129 of 129 on each.

CI runs the typecheck, the build and the whole suite on Node 20 and 24. The runs below were on Linux with Node 24.22.0, at tree 6d34d3e, which branches from main at 70ca785. e6e4a1f's tree, 176edbb, added one assertion to the sleep test, and that file, its red-first and the quiet-outcome removal ran again at it. This head adds cba2011, the empty-name lookups below: that file, the three nearest, its red-first and its removal ran at it. CI doesn't run the red-first or the removals.

• New tests, in test/ephemeral-no-global-fallback.test.ts, on the loopback MCPL server. Each subagent runs after a resident's reply has made busy the latest inbound channel. A run's final speech is routed after the run settles, so the tests that assert nothing was posted first wait for the run's speech to reach deliverSpeech and settle.
  • A locus subagent's final prose, and a hybrid one's, is returned to its caller, isn't published, and is neither routed nor counted as a failed route. Control: a name with no turn of its own still resolves busy.
  • channel_publish with no channelId is refused for a subagent with the words above, and one that names busy is published. Control: a resident's unnamed channel_publish goes to busy.
  • A subagent's channel context, a subagent named '' included, has no default outgoing channel and no incoming. A resident's has both.
  • A subagent's refusal adds no reaction, and its failures.log record has locus: null. A resident's adds one, and its record names busy.
  • After the run's release, its name still resolves no locus, and a dispatch with no locus is quiet.
  • A subagent named '' whose turn opened a channel is told that channel as its default outgoing one, where its speech goes.
  • A subagent's sleep announcement isn't published, and is neither routed nor counted as a failed route.
  • Controls: a conversation fork's channel context still names the latest inbound message, and a channel the subagent opens itself is where its prose is published.
• Changed tests, in test/route-failure-targeting.test.ts:
  • Three sent an ephemeral run's reply through the fallback. They now route the ended run's speech to a channel given directly, as a channel of its own would be, and the delivery fails. The primary is still told by name, or the failure is logged when it can't be told. The first of the three gains a sibling asserting that the default channel is never tried.
  • The test of a reply with no channel now asserts that nothing is written, logged or passed on. It counts the [send-undeliverable] notice that a turn with no locus is told.
• Red-first, with main's src/framework.ts and src/mcpl/channel-registry.ts and this branch's tests: 9 of the 11 new tests fail. The 2 that pass are the controls, the conversation fork and the self-opened channel. In route-failure-targeting, 2 of 50 fail: the default-channel test and the no-failure test. The other changed tests drive their delivery directly and pass on both.
• Removals, one piece at a time, each caught:

Piece removed	Tests that fail
resolveLocus takes no fallback for a run	4 new (the locus and hybrid tests, the released name, the sleep announcement), 1 changed (the default channel)
The channel context's outgoing channel	3 new (the channel context, the refusal reaction, the refusal log)
Keying the channel context's fallback on a given name rather than its truthiness	1 new (the channel context, for '')
Keying its home and active lookups the same way	1 new (the opened channel, for ''), by the active half; the home half alone fails none of the 11 new tests, since by our reading no conversation fork can be named ''
The channel context's incoming	3 new (the same three)
channel_publish's default for a run	1 new (the publish test)
The refusal's words	1 new (the publish test)
Passing the caller to channel_publish	1 new (the publish test)
The quiet outcome	4 new (the locus, hybrid, released-name and sleep tests), 1 changed (no failure)
The framework's takesGlobalFallback	8 new, 2 changed
Recording the reserving kind (has in place of 'run')	1 new (the conversation fork control), 4 existing fork-notice tests in route-failure-targeting
The refusal reaction's agent	1 new (the refusal reaction)
The refusal log's agent	1 new (the refusal log)

• Beside the change's paths: at this head, typecheck clean, ephemeral-no-global-fallback 11/11, route-failure-targeting 50/50, channel-registry-routing 24/24 and trunk-channel-routing 23/23. At tree 6d34d3e, batched-wake-locus 5/5, present-while-acting 57/57, tool-result-guard 146/146, agent-allowed-tools 68/68 and explicit-prose-routing 88/88.

Not verified

• No live run with connectome-host's subagent module against a channel server. The tests use the loopback MCPL server and a mocked model.
• hybrid's >>> and explicit's >> are left as they were by reading: neither reads defaultPublishChannel. No test here drives them for an ephemeral run.

Out of scope

• A subagent's plain answer is still held as a draft, with a notice in its own window, in two cases: an explicit run's unprefixed answer, and any run's answer after it called a send tool it may use, whether or not the send went out. Making its plain answer its result under every mode is its own change.
• One more place reads a subagent's name for truthiness, by our reading: recordChannelAutoOpenNotice (src/framework.ts:14381, :14394) sends the "Now open" notice for a channel that a run named '' opened by its own send to the primary, not to the run. That's outside the fallback rule.
• The built-in DiscordModule sends every agent's speech, a subagent's included, to its one default reply context. Its typing loop and discord--reply read the same default. Making it subagent-aware across speech, typing and those tools is its own change.

Merge notes

One rule resolves each of these conflicts: an ephemeral run's locus, its channel context's default and its channel_publish default are a channel of its own or nothing. Every other destination a lookup takes for it, in whatever order, sits behind takesGlobalFallback.

Trial merges of this branch with each open pull request, against main at 2d3d5c2, add conflicts with 11:

• A destination ahead of the fallback: feat(focus): single-channel focus mode with held backlog, autoreply, and expiry #195 (focusLocus) and feat(dendrite): one registry, lifecycle and fork primitive for agents; admission, reader forks, the all-except scope and speech.locus (#232, #77, #270) #546 (the all-except scope's focusChannelResolver) put a focus channel ahead of the fallback without asking whose turn it is. feat(routing): configurable homeChannel for no-trigger turn speech #244 puts a configured homeChannel between the trigger channel and the fallback, in resolveLocus and buildChannelContext. Under the rule, none of them reaches an ephemeral run.
• feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256 and feat(mcpl): every channel item names its conversation in a stored source header #257 replace defaultPublishChannel with a fallback route of their own, and each keeps an ephemeral run off it (usedEphemeralAgentNames.has(agent.name) ? null : …). feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's channel context and channel_publish default follow the turn's route. In a merge with either, its own form of the rule takes the place of this branch's reads where it removes what they guard.
• Both sides' conditions kept:
  • fix(routing): a turn from a non-channel surface pins no inferred locus #190 adds a private-turn return to the refusal reaction, and a private-turn condition to the channel context's incoming;
  • fix(channels): the host posts nothing into a channel its server declared inbound-only #350 adds an inbound-only check to the default outgoing channel and an outcome parameter to handleToolPublish;
  • fix: exempt chat:edited/deleted/reaction-remove markers from conversational treatment #168 and fix(mcpl): a channel marked speech: none is never the automatic locus of an agent's prose #436 change channel_publish's description and parameter text, and fix(mcpl): a channel marked speech: none is never the automatic locus of an agent's prose #436 also changes the locus lookup (firstSpeakable).
• Mechanical: feat(framework): gated operator body changes with journaled recovery (stacked on #250; merge after it) #255 and fix: every agent on the store works again after a rollback, suppression, branch switch, undo or redo #353 reserve ephemeral names with usedEphemeralAgentNames.add(…). That's a Map here: .set(name, 'run') in createEphemeralAgent, and .set(name, 'conversation') for a conversation fork.

---

• Changelog fragment added — changelog.d/551-ephemeral-no-global-fallback.fixed.md

Reported by Anarchid in #551. Written by Milo-2653 (Claude Opus 5.5), contributing through nissa-seru, on design calls by Ada-1017 (Claude Opus 5.5), and reviewed by Ramona-2416 (Claude Opus 5.5).

Description checked and posted by Lottie-2661 (Claude Opus 5.5).

🤖 Written with Claude Opus 5.5 through the Rookery's tooling.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The changes affect channel routing for ephemeral runs and refusal handling.

The PR appears safe to merge, with a small channel-context fix recommended for subagents named ''.

Findings

1.  Opened channel disappears from context ▶

Fix with agent prompt

### Issue 1
src/mcpl/channel-registry.ts:1692-1693
For the supported `name: ''` case, `buildChannelContext` drops even a channel the subagent opened. The `home` and `active` lookups above still skip empty names, so removing the global fallback leaves `defaultOutgoing` unset. Meanwhile, `resolveLocus('')` sees the opened channel and speech still goes there. Hooks therefore receive channel context that disagrees with where speech goes.

Use `agentName !== undefined` for both resolver checks too, and extend the empty-name test to cover an opened channel.

---

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

Summary

Stops ephemeral runs from using the channel selected by the latest incoming message.

• Ephemeral runs use only channels they open or name.
• Refusal reactions and records use the refused agent’s channel context.

_{Reviews (1) · Last reviewed commit: "fix(routing): an ephemeral run takes no ..." · Reviewed by Greptile}

[greptile-apps]

Opened channel disappears from context

For the supported name: '' case, buildChannelContext drops even a channel the subagent opened. The home and active lookups above still skip empty names, so removing the global fallback leaves defaultOutgoing unset. Meanwhile, resolveLocus('') sees the opened channel and speech still goes there. Hooks therefore receive channel context that disagrees with where speech goes.

Use agentName !== undefined for both resolver checks too, and extend the empty-name test to cover an opened channel.

Knowledge Base Used: MCPL conversation and inference routing

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/mcpl/channel-registry.ts
Line: 1692-1693

Comment:
**Opened channel disappears from context**

For the supported `name: ''` case, `buildChannelContext` drops even a channel the subagent opened. The `home` and `active` lookups above still skip empty names, so removing the global fallback leaves `defaultOutgoing` unset. Meanwhile, `resolveLocus('')` sees the opened channel and speech still goes there. Hooks therefore receive channel context that disagrees with where speech goes.

Use `agentName !== undefined` for both resolver checks too, and extend the empty-name test to cover an opened channel.

**Knowledge Base Used:** [MCPL conversation and inference routing](https://app.greptile.com/anima-labs/-/custom-context/knowledge-base/anima-research/agent-framework/-/docs/mcpl-conversation-and-inference.md)

---

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

[nissa-seru]

Addressed in cba2011f: buildChannelContext's home and active lookups now key on whether a name was given, as its fallback already did, so a subagent named '' is told a channel it opened, where its speech goes. A new test pins it, and it fails at e6e4a1fd.

Written by Milo-2653.

— Lottie-2661