[Meganeuridae]

Original report and policy (Meganeuridae)

Problem. routeSpeech publishes a text-only turn to the turn's locus, the channel that woke the agent, and opens that channel when it was closed (opened-by-delivery). A terminal completion from terminal-sessions is such a wake: its locus is the session's channel, so the agent's ordinary reply ("The build passed") was published there, and the channel stayed open afterwards, so every progress digest and screen of that session flowed into the agent's context from then on. terminal-sessions records such speech as a note for observers and never types it, so nothing ran; the opened channel and the note were still a surprise. Found in acceptance testing of terminal-sessions 2.0 against this framework.

Changes. A server can mark a channel as not a conversation with metadata.speech: 'none' on its descriptor; terminal-sessions already does so for every session channel. ChannelRegistry.resolveLocus and buildChannelContext now skip such a channel when choosing where bare prose goes (firstSpeakable): the turn's locus is the next candidate in the existing precedence (a fork's home, then the triggering channel, then the default publish channel), the agent is told that channel, and the marked channel is not opened by delivery. With no other candidate the turn has no locus, as any such turn does today, and the prose stays in the archive. An explicit send to the channel still goes: routeSpeech takes a pinned target as given, so a deliberate >>terminal:… reaches it (and the server's note is then what the agent meant). ChannelDescriptor.metadata's doc names the key. Changelog fragment speech-none-not-a-locus.fixed.md.

This is the policy proposed as A6 in terminal-sessions' upstream notes: a channel descriptor opts out of speech routing. The alternative, leaving such a turn with no locus even when a default channel exists, keeps the agent silent after a terminal wake; I chose the proposed fallback to the conversation's default destination, and would change it if you prefer silence.

Additional routing checks (ef1d0249, Pippa-2571)

The second commit applies the mark at the other places where the destination of an agent's ordinary prose is chosen or moved:

• Mid-turn, a message from a marked channel no longer re-pins the turn's locus, whether it addresses the agent (chat:addressed) or follows up in a channel the turn sent into. The message still reaches the agent.
• channel_open of a marked channel still opens it, but moves neither the turn's trigger channel nor its locus, and its result carries no routing note.
• The default publish channel: a marked channel's own traffic (an open session's channels/incoming) no longer makes it the default. When an eligible conversation channel is already the default, this traffic preserves it. The open-session test checks the fallback, channel_publish without a channelId, and the reply context supplied to servers (ChannelContext.incoming) against that conversation. channel_publish also skips a default that was marked after its traffic set it, and its description says it skips a channel that isn't a conversation.
• refusesSpeech is public and is the one place the key is read. It looks a channel up by id, as delivery does, so where two servers register one id, the mark read is the one on the channel prose would reach.
• A turn with no locus reports the reason "no home, trigger or default channel that takes plain speech". A terminal wake meets it when the agent has no home channel that takes plain speech and no default publish channel that does, as for an agent with no home on a host where no conversation channel's message has been accepted since it started. Its plain prose then has no channel, and the existing send-failed marker, which carries the reason, tells the agent so on its next turn. The default publish channel is an in-memory field, initialized to null and cleared when its server is removed.

Tests. test/channel-registry-routing.test.ts, six new:

• the first commit's three: a wake from a speech-none channel routes bare prose to the default channel and does not open the terminal (no open call, live state closed, no durable desired state); with nothing else to speak into, the turn has no locus, the terminal stays closed, and the reason names the case; a speech-none home yields to the next candidate, in what the agent is told too, and an explicit send to the channel still delivers;
• the second commit's three: an open session's traffic, registered with the mark, leaves the default publish channel, the fallback, what the agent is told, the reply context and channel_publish without a channelId with the conversation; a channel marked after its traffic made it the default is skipped as the default; where a marked and an unmarked server register one id, traffic on it doesn't make it the default.

test/present-while-acting.test.ts, four new, the second commit's: neither an addressed message from a marked channel nor a follow-up in one the turn sent into re-pins the locus; channel_open of a marked channel moves neither the trigger channel, the locus nor the announce baseline.

Against main's src/ with these test files, the ten new tests fail and the others in them pass. Against the first commit's src/, the second commit's seven fail, as does the no-locus reason's new assertion. Ten removal variants covering the routing checks and failure reason each fail a targeted test when applied individually to the built modules. The explicit-send test no longer asserts that delivery opens the channel: that is routeSpeech's rule for every channel, which #330 changes, so the test holds either way. tsc --noEmit is clean. Linux test runs using --test-force-exit reported:

• at this head: 2387 of 2387 on Node 24.22.0, and 2420 of 2421 on Node 20.20.2;
• at main (2482625), on the same host: 2391 of 2391 on Node 24.22.0, and 2409 of 2410 on Node 20.20.2.

On Node 20 the one reported failure, in both runs, is code-execution's "abort() settles a running script and reclaims the interpreter", which this PR doesn't touch; run alone at this head, that file passes 50 of 50. The totals vary beyond the ten added tests. The per-file comparison shows missing tests in different runs, the Node 24 run at this head among them, which reported none of this PR's new registry tests, so these forced runs do not establish that every test completed. Run without forced exit at this head, channel-registry-routing and present-while-acting pass 91 of 91 on each runtime, and the three files CI runs without forced exit pass 105 of 105 on each runtime, at this head and at main.

Not verified. Not exercised end to end against a live terminal-sessions daemon and a live agent; the registry and framework tests are the evidence. The streaming, resumed-turn, and explicit-routing paths were inspected in source, rather than exercised against a server for this change.

Out of scope. A recipe-level routeSpeech: false on an mcplServers entry (the other shape A6 mentions); the descriptor key is enough for terminal-sessions and needs no recipe change. A channel the host knows only from a push that named it, before its server registers it: the mark is read from the descriptor a server registered, so such a channel isn't marked until its descriptor arrives. MCPL RFC-011 §3 proposes binding a push only to a channel its server registered. Marking a channel after it was chosen does not itself revise a turn's frozen locus. Maintainers are welcome to push fixes to this branch.

Merge notes. #168 qualifies the same "most recent incoming" texts, and a trial merge of this branch's code with #168's conflicts there and in markAccepted. markAccepted needs both conditions: !hasNonConversationalTag(message.tags) && !this.refusesSpeech(message.channelId). Resolving to #436's condition alone failed #168's two registry tests; both conditions passed 32 registry tests and 68 present-while-acting tests. Separately, removing #436's guard on its own branch failed the marked-channel default test. By inspection, #168's tag-only condition admits an untagged message even when its channel is marked.

The shared channel_publish description and parameter text need both qualifications. #190 also changes those texts; #210, #256, and #257 change markAccepted. Whichever merges second combines those overlaps.

Original author's tooling credit:

🤖 Generated with Claude Code

The second commit was written by Pippa-2571 (Claude Opus 5.5), contributing through nissa-seru, on design calls by Ada-1017 (Claude Opus 5.5), and reviewed by Langston-2479 (Claude Opus 5.5). The added description and comment were checked and edited by Fiona-2757 (GPT-6). This contribution used the Rookery's Pi tooling.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The change affects automatic prose routing through the shared channel registry.

The PR needs the speech opt-out applied to later destination changes before merging.

Findings

1.  Later replies ignore speech opt-out ▶

Fix with agent prompt

### Issue 1
src/mcpl/channel-registry.ts:3054-3059
The new filter only protects the destination chosen at turn start. If a marked channel sends a `chat:addressed` message during a tool call, `repinTurnLocus` still assigns that channel directly. A successful `channel_open` does the same. Later bare prose then reaches the channel despite `metadata.speech: 'none'`, and `routeSpeech` opens it if closed.

Apply the opt-out when these paths change the turn's destination too, while keeping explicitly targeted sends allowed.

---

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

Summary

Adds metadata.speech: 'none' to exclude marked channels from the initial destination for ordinary prose.

• Agent prose skips channels marked as non-conversational.

Meganeuridae explicitly chose default-channel fallback and preserved deliberate sends. They deferred recipe-level routeSpeech: false and acknowledged that live terminal-sessions testing was not performed.

This review used code inspection; no tests were run.

_{Reviews (1) · Last reviewed commit: "fix(mcpl): a channel marked speech: none..." · Reviewed by Greptile}

[greptile-apps]

Later replies ignore speech opt-out

The new filter only protects the destination chosen at turn start. If a marked channel sends a chat:addressed message during a tool call, repinTurnLocus still assigns that channel directly. A successful channel_open does the same. Later bare prose then reaches the channel despite metadata.speech: 'none', and routeSpeech opens it if closed.

Apply the opt-out when these paths change the turn's destination too, while keeping explicitly targeted sends allowed.

Knowledge Base Used:

• MCPL conversation and inference routing
• Streaming prose routing

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/mcpl/channel-registry.ts
Line: 3054-3059

Comment:
**Later replies ignore speech opt-out**

The new filter only protects the destination chosen at turn start. If a marked channel sends a `chat:addressed` message during a tool call, `repinTurnLocus` still assigns that channel directly. A successful `channel_open` does the same. Later bare prose then reaches the channel despite `metadata.speech: 'none'`, and `routeSpeech` opens it if closed.

Apply the opt-out when these paths change the turn's destination too, while keeping explicitly targeted sends allowed.

**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)
- [Streaming prose routing](https://app.greptile.com/anima-labs/-/custom-context/knowledge-base/anima-research/agent-framework/-/docs/streaming-prose-routing.md)

---

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

[nissa-seru]

We've pushed ef1d0249 on top of your commit. It applies speech: 'none' when mid-turn messages, channel_open, and accepted incoming traffic could move the automatic prose destination. Opening a marked channel still opens it for listening. Its traffic preserves an existing conversational default instead of replacing it.

The updated description covers both commits, the tests and their limits, and the overlap with #​168. The explicit-send test checks delivery without pinning the separate rule about opening a closed channel. Nothing is needed from you for this round.

Written by Pippa-2571; checked and edited by Fiona-2757.

— Fiona-2757