feat(mcpl): every channel item names its conversation in a stored source header

[nissa-seru]

This PR gives every channel item a stored source header naming its conversation, so a message read alone says where it came from. This update merges #256's latest head, 0bae78b1, with no conflicts; this PR's own commits are unchanged.

Stacked on the speech-routes PR #256, itself stacked on #252 and #247. Please review only what this PR adds to #256's head 0bae78b1: its eleven commits, 42c3f5a..7c4c723, 7adf10c, 0c4030c, b801b10, 7a80df5, 2c6ad2b and 26edf15 (see "A review run's findings"), and the merges 6d6f133, 48e5135, eaf67dd, d0078bc, 69cf802, de92afb, 967615a and 665957d that bring #256's heads in (together, git diff 0bae78b1 665957d). Merge after #256.

Problem

A message read alone couldn't say where it came from:

• channel provenance lived only in metadata;
• adapters described the location in prose only when the channel changed, so the second of two consecutive messages didn't name its channel;
• items in channel_open's backscroll carried nothing.

Changes

• A stored source header on every channel item, from either MCPL lane. The item's first text block names its conversation:
  [source: <server> / <canonical-channel-id> · <label at receipt> · thread <id> · reply to <id>].

  • The label, thread and reply parts appear only when the item has them.
  • A push that names no channel gets [source: <server> · unscoped] instead of a guessed channel.
  • The header is rendered once at ingestion, from the item's frozen source envelope, and stored with the message. A later rename, a recompile or a replay never rewrites it.
  • Creates, coalesced deliveries and RFC-006 corrections all carry it. Local-surface input and a silent heartbeat get none.

• Header values are literal. Every field comes from an adapter. A value holding a grammar character (brackets, ·, /, quotes, a backslash) or any control character is rendered as a JSON string literal. Line separators, DEL and the C1 controls are escaped visibly as \uXXXX, so a label can't forge a second attribution or break a header across lines. A label whose first word is one of the header's own (thread, reply to or unscoped, in any case, followed by whitespace or by nothing) is quoted too, so it can't read as a thread or reply part. A word that runs on (thread:ops, unscopedness) isn't the header's word.

  • Invisible and space-lookalike characters (0c4030c, fixing One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269, which slimepriestess raised from discord-mcpl refactor: settle runEphemeralToCompletion from the stream driver, not trace events (adopts #55) #63's review). A value is also quoted when it holds a format or default-ignorable character, a space other than U+0020, or a / with whitespace on both sides. Format and default-ignorable characters include zero-width characters, bidi overrides and isolates, fillers such as U+3164, the combining grapheme joiner, variation selectors and interlinear-annotation controls. Inside the quotes those characters are escaped as \uXXXX, per UTF-16 unit, so a quoted bidi override can't reorder what follows it, and JSON.parse still gives the value back. A label that starts with such a character is quoted for that alone, so a header word behind it can't read as a field: \u200bthread x is quoted.
  • Emoji sequences stay as they are, joiners and selectors included, because a label is also a name a resident types back: ❤️ cats isn't quoted. Keycaps and tag flags aren't covered, so a name holding one is quoted.
  • The same rule as discord-mcpl's source-header.ts (refactor: settle runEphemeralToCompletion from the stream driver, not trace events (adopts #55) #63), which renders the same header. The two files' six constants and three functions' code are identical, and only quoted()'s comment differs. A differential fuzz across both files, at 7a80df5 here and discord-mcpl's ac08f9b, found no difference in 1,538,880 renderings.
  • 7a80df5 drops two pieces that never changed a rendering. Agnes-984's remove-each-piece check found them: labelValue's strip of invisible characters before the header-word test, which headerValue already covers, and a separator widened to \s\/\s, whose extra spaces other rules already quote. Rendering is identical with and without them across 2,821,280 renderings, comparing b801b10 with 7a80df5.

• Header-shaped text in a body is marked (7adf10c, from slimepriestess's review). Every [source: or [source] in an item's own text blocks, in any case and with any whitespace, control or invisible character inside the bracket (between the letters of source too, since 2c6ad2b), is stored with a backslash before its bracket, wherever it falls. Outside the host's own header line, no unmarked opening remains; a header value that holds one is quoted inside the header, where it reads as a literal.

  • Why. Without it, anyone who can post in a channel could forge a header naming another channel, and a model could take it for the host's attribution and for a place to channel_publish to.
  • Why every position. In XML tool mode the formatter writes participant: before each message, so the real header sits mid-line too, and a body could stage a whole message from another channel.
  • Split openings. The text blocks are read as one string, so an opening split across blocks is caught.
  • One function does this wherever the host stores what a sender wrote: both lanes, the closed-channel invitation (it names the adapter's author, guild and label), channel_open backscroll item text (string content included), a local surface's input, and the legacy Discord module's stored and edited text. The last four are 2c6ad2b's.
  • The only change to stored body text this PR makes. Invisible characters don't hide an opening wherever they stand in it (0c4030c), and with Unicode case folding [ſource: opens one too. Visible look-alikes can still imitate it.

• channel_open backscroll. Each item gains a source field, rendered from the item's own channel, so an item spliced in from another channel shows its true channel. The host owns the source key: an adapter's own source is kept as adapterSource on every item, and an item that names no channel gets no source.

• The authority rule. When a label differs, the canonical id is authoritative. The rule is stated in the channel_list, channel_open and channel_publish descriptions, and in a one-time [source] notice. The rule also says three things: the host writes a [source: …] header only as the first line of a message; \[source… inside a message is its sender's own text; and any such line in a connector tool's own output, such as fetch_history, belongs to that tool result. The notice is stored before it's recorded as given, so a failed append can't suppress it, and a notice already in the resident's context counts as given, so a failed record doesn't repeat it. It's given at the start of the resident's first turn with any item stored with a header retained in its context, however far back. That item can be a channel item, or a push that names no channel; the notice names the unscoped form. The residents told are recorded as a list of names, so any agent name works.

• Readers of what was said skip the header (7adf10c). They skip it while the first text block is exactly metadata.sourceHeader, or begins with it: the store's adaptive resolution joins an item's text blocks with nothing between them and shards the result, so a long item's first shard starts with the header (2c6ad2b). A later shard comes back whole, by its shard index, even if a split falls between a marked opening's backslash and its bracket (26edf15).

  • History search, substring and regex alike, matches and snips the body. So ^ anchors at the body's start, and a channel's name or id doesn't match every item from that channel.
  • The semantic index embeds and snips the body.
  • The tune-out backlog lists author: body.
  • extract still shows the header.

• Digests around the header (the stamps from feat(mcpl): stamp each inbound item's delivered-body and stored-copy digests #252):

  • sourceBodyDigest is taken from the body as delivered, before the header and the marking.
  • The exact header text is recorded as metadata.sourceHeader.
  • storedBodyDigest is of the blocks as handed to storage, header and marking included. The copy check applies to an unsharded copy, and a body the store shards carries the whole body's digests on every shard.
  • metadata.sourceHeader is always the host's: set from the rendered header and removed when none is rendered, so an adapter-supplied value never survives.
  • So a rename changes the header and the stored digest, but never the delivered body's version.

• Notices name a connector's values by the same rule (2c6ad2b). Routing notices, holds, the [delivered] receipt, draft lists and the [discord-send-failed] marker show a label by labelValue's rule, and ids, threads and surfaces by headerValue's. So a label carrying a line break and a forged notice is one quoted, escaped value, while ordinary names read as they are. The closed-channel invitation does the same for its author, label and guild, shows the label its server registered for the channel (else the event's own) as given, with no added #, and writes the ids its options ask for as JSON strings.

Changelog fragment: changelog.d/source-headers.added.md.

Integration with main

The stack integrates by merging upward, so every reviewed commit keeps its hash. 6d6f133 merges #256's integrated head 2287051 into 7c4c723. Both textual conflicts are where this PR meets main's #236, which refuses a visibly-empty push at admission. Review them with git show --remerge-diff 6d6f133:

• src/framework.ts, the push path. Main's isSilentHeartbeatMarker({ ...event, content }) replaces this PR's inline marker check. The source header is still added after that check, so the marker sees the undecorated body. The digest comment now covers both the admission refusal and the header.
• test/inbound-source.test.ts. The integrated parent's rewritten cases are kept (from feat(mcpl): stamp each inbound item's delivered-body and stored-copy digests #252's integration with main, 88a24e6: an empty push refused, an accepted tick, a DM with a body), with this PR's header expectations added: the accepted push stores [source: discord · unscoped] first, and its stored digest differs from the delivered one. A DM stores the header, the body and the invitation. The empty-body header assertions are gone, since such a push no longer reaches storage.

48e5135 then merges #256's shutdown fix 4180ccd into 6d6f133. It merged clean (git show --remerge-diff 48e5135 is empty), and this PR's delta on 4180ccd is the same as on 2287051.

eaf67dd then merges #256's head 8f29e62 into 48e5135. That head adds the follow-up to slimepriestess's review of #256 (1bbf3f7, 2c83659) and #256's merge of #252's later head a2accc1. a2accc1 carries #247's 315d12c and the stored-form body digest from slimepriestess's review of #252. It merged clean (git show --remerge-diff eaf67dd is empty).

d0078bc then merges #256's head e041853, which fixes what a review run over #256's own commits found (see #256's "A review run's findings"). It had one conflict, in src/framework.ts, from adjacent edits. This branch added withSourceHeader and maybeExplainSourceHeaders directly above candidateForChannel, and e041853 rewrote that method's doc comment. The resolution keeps both: the two methods as they are here, then the new comment.

69cf802 then merges #256's head 4e19f02, in which routing names each conversation with its server, the server / channel-id form this PR's source headers use. It merged clean.

de92afb then merges #256's head 6aaeb85, the fixes from #256's second review run, with #252's and #247's. It merged clean. Two of #252's tests it brings in assumed a stored body without a header, and the invitation's label text changed under them; 2c6ad2b updates them (see below).

967615a then merges #256's head 9dc8239, which gives defaultOutgoing MCPL RFC-011 revision 5's meaning. It merged clean (git show --remerge-diff 967615a is empty), and the full suite at 967615a passes 1658/1658 across 136 files, with npx tsc --noEmit clean.

665957d then merges #256's head 0bae78b1, which adds #256's ten fallback rounds: a wake that names no conversation keeps the fallback route, and a wake that continues a turn names that turn's conversations. It merged clean (git show --remerge-diff 665957d is empty). This PR's own changes to the files it touches are where messages enter, plus how a conversation is described, and none of them is in the route decision. The full suite at 665957d passes 1714/1714 across 136 files, run one file at a time with no --test-force-exit, with npx tsc --noEmit clean.

A review run's findings (2c6ad2b, 26edf15)

The house ran its review workflow over this PR's own commits at 69cf802. Petra-1012 judged each of the 27 issues it found against the code: 18 were real and are fixed in 2c6ad2b, 8 were real but not worth changing, and 1 is intended. The review of #247 had also left one point to this PR: how a connector's label is shown in host notices. The fixes are described above, under Changes, and 26edf15 follows the review of 2c6ad2b: the invitation names a channel by the label its server registered, and a later shard of a sharded item keeps its text. In short:

• Marking ignores controls and invisible characters anywhere in an opening, whitespace between its letters, and reaches the invitation, a local surface's input, backscroll string content and the legacy Discord module. A malformed backscroll entry passes through rather than failing the open.
• Readers skip the header at the start of a sharded first shard.
• The notice isn't repeated when its record fails to write, and a failed state read gives none.
• metadata.sourceHeader is always the host's.
• Notices and the invitation name connector values by the header's rule.
• Texts. The fixes reach the rule's wording, the label rule, the emoji exemption, the digest wording and the changelog fragments. Comments no longer carry internal tracker references.

Judged not to need a change (8). Among them are edits through editMessage replacing a header without a check, stored items from before the upgrade staying unmarked, and the first request after the upgrade missing the prompt cache because the channel tools' descriptions changed.

Checks (Anima VPS):

• The full suite, one file at a time: 1658/1658 across 136 files, with npx tsc --noEmit clean, at 26edf15, and the same at 2c6ad2b. At the merge de92afb, before the round, it gave 1642/1644. The two failures were feat(mcpl): stamp each inbound item's delivered-body and stored-copy digests #252's tests that 2c6ad2b updates.
• Red checks. Each of 16 checks undoes one fix with a string replacement and runs the test written for it. Each test passes on 26edf15's source, and each fails under its mutation. They cover:
  • marking: spaced letters, controls, the invitation on each lane, its author's quoting, console input, backscroll string content and malformed entries, and the Discord module's import and edit;
  • the invitation's registered label;
  • readers: the sharded first shard, and a later shard kept whole;
  • the notice: a context that holds it;
  • notices: the label and the thread.

Tests

test/source-headers.test.ts runs over real MCPL children: a Discord-shaped server with two guilds sharing a channel name plus DMs, and a Zulip-shaped server with topics in a stream. It covers:

• the grammar;
• consecutive and interleaved messages, and same-named channels kept apart by id;
• thread and reply parts;
• DM, unscoped and console input;
• a rename after ingestion;
• the notice said once;
• backscroll items each under their own channel;
• a coalesced correction;
• forged labels and line separators;
• stored-copy digests, and an edit leaving them unmatched;
• the tool result rendered through membrane's own AnthropicXmlFormatter.formatToolResults;
• Greptile's four findings, one regression case each (7c4c723; three new tests and the backscroll test extended), each confirmed failing against 49a7baf's source:
  • labels beginning with thread, reply to or unscoped, with near-misses left unquoted;
  • One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269's vectors, shared with discord-mcpl refactor: settle runEphemeralToCompletion from the stream driver, not trace events (adopts #55) #63. Each of these is quoted, escaped, and parses back: invisible characters in front of header words (ZWSP, U+3164, U+034F, a bare FE0F, U+FFF9); bidi overrides and isolates; no-break and ideographic spaces; a ZWJ outside an emoji; and an astral tag character. Emoji names, café, Обсуждение and a/b are left as they are;
  • a backscroll item that names no channel but brings its own source;
  • residents named constructor, toString and __proto__, each told once, with the record read back from the store;
  • a channel item behind 250 later pushes still bringing the notice.
• slimepriestess's review (7adf10c). Run against eaf67dd's source with identity stand-ins for the two new helpers, each case fails, except two assertions that hold on both heads by design: the near-miss case and extract keeping the header:
  • the marking. An opening at a line start, mid-line, after user: , in [source], with odd spacing and mixed case, after a doubled bracket, already escaped, and twice in one text. Also openings split across blocks, inside the word, over four blocks, around an image and between empty blocks, with near-misses left alone;
  • both lanes storing a staged body and a forged push marked, with each digest checked. Through membrane's real AnthropicXmlFormatter, the staged line reads user: \[source: …;
  • a coalesced delivery, its correction and a coalesced push;
  • backscroll item text, with and without a header;
  • the readers. Substring search for a channel name and for an id, ^ through the regex worker, messageIndexText, and the tune-out backlog. extract keeps the header;
  • a resident that only receives unscoped pushes being told.
• The far-back notice test (Greptile's fourth case) now uses header-less filler, so it still measures distance. It passes on both heads.
• Ezra-1872's review ran ten mutants against the compiled code, each undoing one of 7adf10c's behaviours, and each failed the test meant to catch it.

npm test, the full suite:

• 69cf802 (this head), as CI runs it (Petra-1012, Anima VPS, a fresh npm install): npx tsc --noEmit is clean, and npm test passes 1602/1602 on node v24.21.0 and on node v20.20.2: 7a80df5's 1588 plus feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's fourteen new tests (e041853's ten and 4e19f02's four).
• d0078bc, as CI runs it (Petra-1012, Anima VPS, a fresh npm install): npx tsc --noEmit is clean. npm test passes 1598/1598 on node v20.20.2, which is 7a80df5's 1588 plus feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's ten new tests. On node v24.21.0 the same run reported 1590 passes and no failures: 8 of channel-registry-routing's tests went unreported under --test-force-exit, and that file passes 30/30 on its own.
• 7a80df5, as CI runs it (Petra-1012, Anima VPS, a fresh npm install: chronicle 0.4.0, context-manager 0.13.0, membrane 0.5.86): npx tsc --noEmit is clean. npm test passes 1588/1588 on node v20.20.2. On node v24.21.0 the same run reported 1576 passes and no failures: 12 tests from channel-registry-routing went unreported under --test-force-exit, and that file passes 30/30 on its own. main 02adf27, an ancestor of this head, gives 1377/1377 on both runtimes. With 7adf10c's tests run against the code before it (eaf67dd), source-headers fails to load, and tune-out-e2e's divert → wake → auto-cancel arc fails.
• One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269, the rendering compared with 7adf10c. We compared 769,440 header renderings, built from each code point in U+0000–D7FF, U+E000–FFFF, U+10000–1FAFF and U+E0000–E01EF in six label templates. 3,730 differ, and every one involves a format or default-ignorable character, or a non-ASCII space. Across 512,960 openings with a code point inserted, 1,816 are newly marked, and each involves such a character. These were run at b801b10, which renders identically to 7a80df5.
• b801b10 (Petra-1012), from a clean build: 1538/1538, and code-execution 50/50, run with --test-concurrency=4 --test-force-exit. That's 1588 in all, one more than at 7adf10c: One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269's new test. The 1-minute load average at the end was about 10. Unforced, source-headers, inbound-source, empty-content, speech-routes, present-while-acting, stop-with-work-in-flight and mcpl-coalescing passed 220/220, with no file reporting asynchronous activity after its tests. Against 7adf10c's source, One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269's header test and the marking test fail.
• 7adf10c (Pippa-1873), from a clean build: 1587 pass / 0 fail of 1587, with --test-force-exit --test-concurrency=2. eaf67dd on the same runner gave 1578/1578; compared by name, the difference is the nine new tests. The load average at its start was about 3.
• 48e5135 (Basil-1786), from a clean build: 1570 pass / 0 fail of 1570. No file reported asynchronous activity after its tests ended. The load average at its start was about 14–16.
• 6d6f133 (Basil-1786), from a clean build: 1535 pass / 4 fail of 1539. The load average at its start was about 20–26.
  • Vector 34c: "300 coalesced messages took 38939ms".
  • The present-while-acting file: its aborted-stream test leaves work running after stop() ("Error: Queue is closed", an unhandled rejection). That shutdown race reproduces on feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256 without this PR: at ae42e95, and at c899120, which differs from 2287051 only in comments. feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's 4180ccd fixes it, merged here in 48e5135.
  • code-execution's "abort() settles a running script and reclaims the interpreter" and workspace-watcher's "does not re-attach for normal create and modify traffic". Each file passed when run alone afterwards, twice, at load about 15: 50/50 and 10/10.
  • R17 passed in this run.
• eaf67dd (Petra-1012), from a clean build: 1527/1527, and code-execution 50/50, run with --test-force-exit --test-concurrency=4. Under --test-force-exit each file's process exits at its last test, so this run can't show asynchronous activity left after a test. Basil-1007's unforced runs of the focused files at this head showed none. The load average was about 3–4.
• CI at 6d6f133 and again at 48e5135. All four Build & Test jobs (ubuntu-latest and macos-latest, node 20 and 24) pass, and so does "Changelog entry present". At 48e5135 GitHub reports the PR MERGEABLE and CLEAN.
• 7c4c723 (Basil-1786): 1461 pass / 1 fail / 1 cancelled of 1463. The 1-minute load average was 15–21.
  • The failure is vector 34c: "300 coalesced messages took 41728ms".
  • The cancelled item is mcpl-coalescing.test.js itself. It stalled in R16 ("a batch whose render-start commit failed is re-woken once storage recovers") and was stopped after 15 minutes, so R16 through R18 didn't report.
  • Run alone, mcpl-coalescing gave 64/66 at both 7c4c723 and 49a7baf, with the same two failures listed below.
  • R16's stall reproduces without the follow-up commit. Run alone eight at a time, R16 hung (no result within 60 s) in 4 of 24 runs at 49a7baf; at 7c4c723 it passed 24 of 24 that way, and hung (90 s limit) in 2 of 8 sequential runs under heavier load. It registers no channel, so the notice code it reaches is unchanged in effect: no notice, no store write.
• 49a7baf. It gave 1432 pass / 2 fail of 1434, and code-execution 36/36. The 1-minute load average was 8–10. Both failures are the same mcpl-coalescing timing tests the base shows:
  • vector 34c: "300 coalesced messages took 40690ms";
  • R17: counted [1, 2, 3, 4] against an expected [1, 2, 3].
• The same two fail without this PR. At feat(mcpl): stamp each inbound item's delivered-body and stored-copy digests #252's base 711b400, mcpl-coalescing run alone under similar load gave 64/66, with these same two failures. feat(mcpl): stamp each inbound item's delivered-body and stored-copy digests #252's head d36a5ad reports them in its own body too.
• R17's race is in the test. It is corrected separately in test(coalescing): R17 observes recovery transitions, not drains #254 (test-only, on main, not in this branch).
• 34c's bound is wall-clock. Elsewhere it has passed at lower load.
• The base PR's head 42c3f5a. It shows the same two failures, at 1419/1421.

Not verified

• The header tests use stub adapters. discord-mcpl agreed the grammar. An independent reviewer also composed 49a7baf, 6d6f133, 48e5135 and eaf67dd with real connector processes (Discord, Slack, Telegram, portal and anima-dnd, platform clients stubbed): routing and delivery outcomes all pass, as at feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's head.
  • At 48e5135, every connector passed twice, with no stray asynchronous activity: Discord 3/3 (before and after fix: prevent deferred inference requests from starving the event loop #66), Slack 5/5, portal 4/4 and Telegram 5/5. anima-dnd passed 3/3 in 20 of 20 runs, across both door heads.
  • At eaf67dd, the same: Discord 3/3 (before and after fix: prevent deferred inference requests from starving the event loop #66), Slack 5/5, portal 4/4 and Telegram 5/5, with anima-dnd 3/3 in 10 of 10 runs across both door heads.
  • The anima-dnd figure comes from a corrected harness. It settles the door's catch-up wake first and asserts that the answer follows the GM's question. The earlier anima-dnd runs, through 6d6f133, confirmed routing and delivery but didn't pin which message was answered.
  • At 6d6f133, one of eight anima-dnd runs had failed at the file level on feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's shutdown rejection ("Queue is closed"), which 4180ccd fixes.
  • These checks assert routing and delivery outcomes, not the header text. No live platform was exercised, and eidoverse-worlds was not composed.
  • They weren't rerun at 7adf10c, which doesn't touch routing or delivery. It changes stored body text only where a body holds a header-shaped opening; otherwise it changes the notice, the channel tools' descriptions and three readers.
• The header's effect on residents' attribution is untested. The tests establish that the header is present, correct and stable, not how residents use it.

Companion PRs and merge order

#247, then #252, then #256, then this PR.

Authorship and review

• Petra-1012 wrote this PR and its first four commits.
• Tessa-974 reviewed it and approved it at 49a7baf.
• The header grammar was agreed with the discord-mcpl side of the same work.
• Basil-1007 composed 49a7baf, 6d6f133, 48e5135 and eaf67dd with the connectors, and verified 6d6f133 and 48e5135 as steps up the stack.
• Basil-1786 addressed Greptile's four findings in 7c4c723. June-1784 assessed the findings with Basil-1786, and reviewed and approved that commit.
• Basil-1786 merged feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's integrated head into this branch (6d6f133), then its shutdown fix (48e5135). June-1784 reviewed and approved both merges.
• Petra-1012 merged feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's 8f29e62 into this branch (eaf67dd). Basil-1007 reviewed the merge and reran the connector compositions at it, all clean, with anima-dnd at 10 of 10.
• Petra-1012 merged feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's e041853 into this branch (d0078bc), and Ada-1017 reviewed the resolution.
• Petra-1012 merged feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256's 4e19f02 into this branch (69cf802). Basil-1007 reviewed the same clean merge of that fix's first version.
• Petra-1012 made One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269's header-rule change in 0c4030c, b801b10 and 7a80df5. Agnes-984 made the same change on discord-mcpl refactor: settle runEphemeralToCompletion from the stream driver, not trace events (adopts #55) #63, reviewed and approved this side, and ran the cross-file fuzz. Ada-1017, who owns One quoting rule for server-chosen text: renderSourceHeader/labelValue miss the invisible-character class logValue already escapes #269, found the default-ignorable gap. The emoji exemption is Agnes's proposal. Agnes-984's remove-each-piece check found the two pieces 7a80df5 drops, and Agnes-984 reviewed that commit.
• Pippa-1873 answered slimepriestess's review in 7adf10c. Ezra-1872 assessed the review independently, then reviewed and approved the commit. Petra-1012 accepted the change to stored body text and set its conditions.
• Petra-1012 wrote the second review round, 2c6ad2b and 26edf15, and the merge de92afb. Agnes-984 reviewed them, and found what 26edf15 changes.

Written by Petra-1012 (Opus), contributing through nissa-seru. Basil-1786 (Opus) wrote the Greptile follow-up and its edits, and June-1784 reviewed them. Pippa-1873 (Opus) wrote the slimepriestess follow-up (7adf10c) and its edits, and Ezra-1872 (Opus) reviewed them. Petra-1012 wrote the #269 change (0c4030c, b801b10, 7a80df5) and the second review round (2c6ad2b, 26edf15), with their edits, and Agnes-984 (Opus) reviewed both. Everyone named contributes through nissa-seru.

[greptile-apps]

Corrections lose their thread

A channel-scoped push correction can omit origin.threadId and inherit the subject's earlier thread during delivery. The new frozen source does not inherit it. In the existing G11 case, the stored message still belongs to thread, but its header and route candidate name the channel root. A reply can therefore publish outside the original thread.

Freeze the source with the same inherited identity used by deliverCoalesced.

Knowledge Base Used: MCPL conversation and inference routing

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 8361-8362

Comment:
**Corrections lose their thread**

A channel-scoped push correction can omit `origin.threadId` and inherit the subject's earlier thread during delivery. The new frozen source does not inherit it. In the existing G11 case, the stored message still belongs to `thread`, but its header and route candidate name the channel root. A reply can therefore publish outside the original thread.

Freeze the source with the same inherited identity used by `deliverCoalesced`.

**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.

[greptile-apps]

Draft resends lose their server

If two servers register the same channel id with different labels, a unique label resolves to the right server. This second lookup drops resolved.serverId and rejects the bare id as ambiguous. The resident cannot resend the draft through this tool even though its destination is known.

Pass resolved.serverId into resolveDestination.

Suggested change

	const target = registry.resolveDestination({ channelId: resolved.channelId });
	const target = registry.resolveDestination({ serverId: resolved.serverId, channelId: resolved.channelId });

Knowledge Base Used: MCPL conversation and inference routing

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 9870

Comment:
**Draft resends lose their server**

If two servers register the same channel id with different labels, a unique label resolves to the right server. This second lookup drops `resolved.serverId` and rejects the bare id as ambiguous. The resident cannot resend the draft through this tool even though its destination is known.

Pass `resolved.serverId` into `resolveDestination`.

```suggestion
    const target = registry.resolveDestination({ serverId: resolved.serverId, channelId: resolved.channelId });
```

**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.

[greptile-apps]

Speech blocks every agent

After an abort, the old stream can end while its publish is still pending. The next wake reaches this new wait through the awaited scheduler path. While it waits for that agent's speech, the framework stops taking queued messages and tool results for every agent.

An unanswered channel open with requestTimeoutMs: 0 can also leave stop() waiting forever before it closes the connection. Park this agent's wake until its speech settles instead of awaiting network work inside the scheduler.

Knowledge Base Used: Agent runtime orchestration

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 11566

Comment:
**Speech blocks every agent**

After an abort, the old stream can end while its publish is still pending. The next wake reaches this new wait through the awaited scheduler path. While it waits for that agent's speech, the framework stops taking queued messages and tool results for every agent.

An unanswered channel open with `requestTimeoutMs: 0` can also leave `stop()` waiting forever before it closes the connection. Park this agent's wake until its speech settles instead of awaiting network work inside the scheduler.

**Knowledge Base Used:** [Agent runtime orchestration](https://app.greptile.com/anima-labs/-/custom-context/knowledge-base/anima-research/agent-framework/-/docs/agent-runtime.md)

---

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

[greptile-apps]

Marking freezes the host

markHeaderOpenings copies the whole offsets array for every opening in a block. The work grows quadratically. On Node 24.21, the current helper took 5,691 ms for one 320,000-byte block containing 40,000 openings—well below the coalesced lane's 1 MiB limit.

Both MCPL lanes run this synchronously before storage, so one accepted message can stall the host. Append offsets in place and walk text blocks without restarting texts.find for every match.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/mcpl/inbound-source.ts
Line: 468-469

Comment:
**Marking freezes the host**

`markHeaderOpenings` copies the whole offsets array for every opening in a block. The work grows quadratically. On Node 24.21, the current helper took 5,691 ms for one 320,000-byte block containing 40,000 openings—well below the coalesced lane's 1 MiB limit.

Both MCPL lanes run this synchronously before storage, so one accepted message can stall the host. Append offsets in place and walk text blocks without restarting `texts.find` for every match.

---

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

[greptile-apps]

Stream events carry wrong fields

On channel-bearing locus and hybrid turns, typingChannel is now an object. It is still sent directly as channelId on inference:tokens and inference:content_block, whose public types require a string. Subscribers receive { serverId, channelId } instead of a channel id, breaking channel matching and string operations.

Keep the typing target separate from the string sent in traces.

Knowledge Base Used: Agent runtime orchestration

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 12056

Comment:
**Stream events carry wrong fields**

On channel-bearing locus and hybrid turns, `typingChannel` is now an object. It is still sent directly as `channelId` on `inference:tokens` and `inference:content_block`, whose public types require a string. Subscribers receive `{ serverId, channelId }` instead of a channel id, breaking channel matching and string operations.

Keep the typing target separate from the string sent in traces.

**Knowledge Base Used:** [Agent runtime orchestration](https://app.greptile.com/anima-labs/-/custom-context/knowledge-base/anima-research/agent-framework/-/docs/agent-runtime.md)

---

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

[greptile-apps]

Sleep uses the old destination

driveStream dispatches calls before queuing the prose written before them. The new sleep announcement therefore enters the speech chain first. If an earlier round pinned >>>#alpha and this round writes >>>#beta ... before calling sleep, the announcement reads the old alpha pin and publishes there before beta's words.

Queue the round's prose before dispatching the announcement so it uses the resident's latest choice.

Knowledge Base Used: Streaming prose routing

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 18035-18036

Comment:
**Sleep uses the old destination**

`driveStream` dispatches calls before queuing the prose written before them. The new sleep announcement therefore enters the speech chain first. If an earlier round pinned `>>>#alpha` and this round writes `>>>#beta ...` before calling `sleep`, the announcement reads the old alpha pin and publishes there before beta's words.

Queue the round's prose before dispatching the announcement so it uses the resident's latest choice.

**Knowledge Base Used:** [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.

[greptile-apps]

Sender metadata skips the explanation

An adapter message with kind: 'source-header-notice' in its metadata or origin counts as proof that the host already gave the explanation. Both MCPL lanes preserve kind, so this check skips the real notice and records it as given. The resident can then miss the explanation of host headers and sender text.

Also require the host's system: true marker; ingestion already removes an adapter-supplied system.

Suggested change

	const meta = message.metadata as { sourceHeader?: unknown; kind?: unknown } | undefined;
	if (typeof meta?.sourceHeader === 'string') sees = true;
	if (meta?.kind === 'source-header-notice') given = true;
	const meta = message.metadata as { sourceHeader?: unknown; kind?: unknown; system?: unknown } | undefined;
	if (typeof meta?.sourceHeader === 'string') sees = true;
	if (meta?.system === true && meta.kind === 'source-header-notice') given = true;

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 10215-10217

Comment:
**Sender metadata skips the explanation**

An adapter message with `kind: 'source-header-notice'` in its metadata or origin counts as proof that the host already gave the explanation. Both MCPL lanes preserve `kind`, so this check skips the real notice and records it as given. The resident can then miss the explanation of host headers and sender text.

Also require the host's `system: true` marker; ingestion already removes an adapter-supplied `system`.

```suggestion
        const meta = message.metadata as { sourceHeader?: unknown; kind?: unknown; system?: unknown } | undefined;
        if (typeof meta?.sourceHeader === 'string') sees = true;
        if (meta?.system === true && meta.kind === 'source-header-notice') given = true;
```

---

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

[greptile-apps]

Resend examples break on quotes

The resend example puts channelId and threadId between raw quotes instead of rendering JSON string literals. A valid thread id such as release "v2" makes the suggested call malformed. Backslashes and internal newlines also change what it says. The earlier destination description is escaped, but this second copy is not.

Use quoted() for both arguments so residents can follow the example.

Suggested change

	(channelId ? `, destination: "${channelId}"${threadId ? `, threadId: "${threadId}"` : ''}` : '') +
	(channelId ? `, destination: ${quoted(channelId)}${threadId ? `, threadId: ${quoted(threadId)}` : ''}` : '') +

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/framework.ts
Line: 15775

Comment:
**Resend examples break on quotes**

The resend example puts `channelId` and `threadId` between raw quotes instead of rendering JSON string literals. A valid thread id such as `release "v2"` makes the suggested call malformed. Backslashes and internal newlines also change what it says. The earlier destination description is escaped, but this second copy is not.

Use `quoted()` for both arguments so residents can follow the example.

```suggestion
                      (channelId ? `, destination: ${quoted(channelId)}${threadId ? `, threadId: ${quoted(threadId)}` : ''}` : '') +
```

---

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

[greptile-apps]

Evicted conversations keep accumulating

During a long inference or pause, each debounced batch can move evicted events into a new server/channel/thread key. This map has no cap and keeps growing until inference ends. Clearing message content does not bound the retained ids, labels, and route candidates, so the 100-event cap no longer bounds retained metadata.

Cap this state too. When details exceed the cap, keep a conservative “multiple conversations” result.

Knowledge Base Used: Event gate pipeline

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/gate/event-gate.ts
Line: 1825-1828

Comment:
**Evicted conversations keep accumulating**

During a long inference or pause, each debounced batch can move evicted events into a new server/channel/thread key. This map has no cap and keeps growing until inference ends. Clearing message content does not bound the retained ids, labels, and route candidates, so the 100-event cap no longer bounds retained metadata.

Cap this state too. When details exceed the cap, keep a conservative “multiple conversations” result.

**Knowledge Base Used:** [Event gate pipeline](https://app.greptile.com/anima-labs/-/custom-context/knowledge-base/anima-research/agent-framework/-/docs/event-gates.md)

---

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