[matt2e]

Problem

Since #1601, the mention chooser in streams and forums also offers the viewer's own agents from their other communities ("Jitter · In feedback…"). The Space rule from #935 and #1623 tags a typed name only when exactly one choice has that exact name, and it counted those other-community agents as local choices.

So a viewer with two agents named Jitter, one in this community and one in another, got two exact matches for @Jitter. Space stayed a plain space, the message went out untagged, and Jitter never woke. The same happened to a member who shared a name with one of the viewer's agents elsewhere (a person Mira vs. the viewer's agent Mira in another community). An existing MessageComposer test pinned this as expected behaviour. Before #1601, @Jitter + Space tagged the Jitter here, which is why the report says "it used to work".

Change

All of this lives in the mentions plugin (src/bundled/mentions). The chooser and the headless pending-chip resolver share spaceMention, so they follow the same rules.

• Rule 1: hide same-named other-community agents. The chooser no longer offers an agent from another community when one of the viewer's unarchived agents in this community has the same normalized name, whether or not that agent is in the channel (offeredElsewhere in mention-candidates.ts). Team eligibility and the Agents page import list still use the unfiltered list. An agent archived here can still be brought back by mention.

• Rule 2: Space precedence (here → own elsewhere → directory). Among choices with the typed name:

  1. a local choice in this community (a member, or an agent offered here, including "Not in channel" rows) wins, so the viewer's agents from other communities can't be the exact name;
  2. otherwise, one of the viewer's agents from another community wins, so directory results can't be the exact name;
  3. otherwise, the plain full-set rule applies.

  In step 2, a local choice here and a directory result with the same name are still namesakes, as before.

• The instant tag stays for elsewhere-only names. When the only local match is the viewer's agent from another community (for example @Rocket), Space tags it at once, as on main. A directory result found later can't make it ambiguous, and a truncated directory search can't turn it back into prose.

• The cold-start hold covers only matching names. Rule 1 needs to know the viewer's agents here. Until they are known, other-community rows are held, but only for queries that match one of their names (elsewhereMatches). Directory rows for that query wait too, so a late-released agent never lands below a directory stranger, and Space and Enter can't pick different rows. Other names, such as @alice, tag a fully named member at once.

• Library prefetch when the composer mounts. In channels that offer such agents, the composer's mounted mention picker now loads the viewer's agent library before the first @. This avoids a cold read that shares the resolver's 8 s timeout and could send Space back to prose. It does not start for DMs, Solo chats, rosters, invite mode, or when there are no other-community agents.

• Retry rereads the library only when it failed. Retry for a directory, membership or name error no longer reloads a ready library, which used to disable shown rows until the reload landed. This also applies to invite mode.

Behaviour in a single community, with no agents elsewhere, is unchanged. elsewhereMatches is false there, so every rank reduces to the old exactMention calls.

Spec (src/bundled/mentions/README.md)

• §1, item 3: adds the exception for the viewer's agents from other communities that have the same name as one of the viewer's unarchived agents here, normalized as in §3.
• §3 (unchanged): its normalization (case, accents, outer whitespace) is now what §1's exception and §6's matching-name hold refer to.
• §5: adds the precedence paragraph (here → own elsewhere → directory) and the instant tag for an elsewhere-only name.
• §6: other-community rows show only once the viewer's agents here are known. Matching queries wait, and directory people wait with them. A client SHOULD start that read when an offering composer opens.
• Fixtures: the space fixtures now run §5 step 2 with the precedence. Members and choices marked offered or elsewhere count as local. There are new cases for each precedence tier and one for an offered non-member vs. its namesake elsewhere. offered is a new optional field (default false), and the fixture version stays 2.

docs/buzz-help.md does not describe the Space-tag rule, so it isn't updated.

Validation

Unit and component suites

• At 2fadf38e4: vitest run src/bundled/mentions MessageComposer.test.tsx SoloDraft.test.tsx person-match.test.ts relay/mentions.test.ts passed all 946 tests in 11 files.
• At the head 5c9dde762 (adds tests only), on a heavily loaded machine (load average about 70–120 on 16 cores): 947 of 949 passed. Both failures were in tests this branch doesn't touch. "asks a mentioned agent of the viewer's for a model and effort on the next run only" hit the 5 s timeout, and a SoloDraft retry test lost its chip. Both passed when rerun alone, and SoloDraft.test.tsx then passed as a whole (89 tests).
• tsc --noEmit, biome check on the changed files, and git diff --check passed.
• Fail-before: each commit's new cases fail against the previous production files or a targeted mutation (5c9dde762 has no production change). Each commit message lists which change fails which case.

AI review rounds

1. Review of 73ea38cc8 → c39512ac6. The review found that a non-member directory result could beat the viewer's own other-community agent on Space, so Space and the highlighted row disagreed. An elsewhere-only name also waited for the directory and fell back to prose on a truncated search, and a row could appear and then go disabled on a cold start. Changed: the three-tier precedence, the instant tag restored, other-community rows held until the viewer's agents here are known, and a test for the archived-agent filter.
2. Review of c39512ac6 → 2692f6e54. A late-released elsewhere row could land below a directory stranger, so Enter and Space picked different rows. The hold also delayed every query, @alice included, and Retry never reread a failed library. Changed: directory rows wait while held and each batch puts elsewhere rows first, the hold covers only matching names, Retry rereads the library, and the §6 wording was tightened.
3. Review of 2692f6e54 → 2fadf38e4. The ship gate: the first matching @ of a session waited for a cold library read that shares the 8 s resolver timeout. Retry also reread a ready library. Changed: the library loads when an offering composer mounts, and Retry rereads only on failure.
4. Review of 2fadf38e4. "I'd ship this: prefetch lifecycle and Retry gate are sound, in-app check pending." It verified the prefetch effect's retain/release lifecycle, that it is gated to offering channels, the StrictMode double call, and the invite-mode Retry change. No blocking findings; the in-app check is still pending.

5c9dde762 is a test-only commit, not a review round. It pins a case the author raised: the viewer owns an agent here that isn't in the channel, plus a namesake elsewhere. That already works, through rule 1 and the "Not in channel" row counting as local. MessageComposer tests and a space fixture now cover it.

Not run, and still needed

• The browser journeys (mention-rules, completion-work, typeahead) were not run.
• The change was not tested in the installed app or with a real account. Nobody has measured how long the cold-start library read takes in the app. The prefetch should make the hold rare, but that isn't verified.
• Not covered by tests: release of the library demand when the composer unmounts (the fixture stubs retain), and the picker's "Retry agent list" (it shares its gate with the tested inline Retry).
• The demo below ran the real composer and mentions plugin in headless Chromium against an in-memory relay, with other-community agents from a stub: Chromium only, light theme, one width, default text size.
• At the author's request, this branch was pushed with --no-verify, so the pre-push hooks did not run on the push. It is based on ebbdb5a69 and has not been rebased onto current main.
• A human test is still needed. In a stream where you can add members, with an agent Jitter here and another Jitter in another community: type @Jitter, and only one Jitter row should show (no "In …" row). Press Space and it should become a chip; send, and Jitter should wake. Control: @ plus a name that only your agent elsewhere has should chip at once, and sending should ask "Bring … into …?".

Demo

Left: merge-base ebbdb5a69. Right: branch head 5c9dde762. Same typing in a #launch stream; captions name who owns each row.

elsewhere-before-after.mp4

#	Time	Rows and owners	Typed	Before	After
1	0:04–0:09	"Jitter": yours here (a member). "Jitter · In feedback…": yours elsewhere	@Jitter	Both rows	Only yours here
2	0:09–0:17	As in 1	Space, please look, Enter	Space stays a space. Sent with no one tagged	Chip. Sent, tagging Jitter here
3	0:17–0:27.5	"Mira": a person here. "Mira · In feedback…": your agent elsewhere	@Mira, Space, can you review, Enter	Space stays plain text. Sent with no one tagged	Space tags the person. Sent, tagging Mira
4	0:27.5–0:38	Only row: "Rocket · In feedback…", yours elsewhere	@Rocket, Space, ship it, Enter	Chip at once, then "Bring Rocket into team.example.com?"	Same
5a	0:38–0:51	"Goose · Not in channel": yours here, not in #launch. "Goose · In feedback…": yours elsewhere	@Goose, Space, hi, Enter, then "Do nothing"	Both rows. Space stays plain text. Sent with no one tagged	One row. Space makes a chip. Enter asks "Goose is not in this channel." Goose is mentioned (not notified)
5b	0:51–1:02	"Goose · In feedback…": yours elsewhere. "Goose · Not in channel": someone else's, found by directory search	@Goose, Space, hi, Enter	Your Goose from feedback… is chosen, then "Bring Goose into team.example.com?"	Same
6	1:02–1:04.9	"Alice": a person here. Fresh session, library read held 6 s	@alice, Space	Chip at once	Chip at once, no "Loading recipients…"
6b	1:04.9–1:14	Your Jitter here, and (before only) your Jitter elsewhere	and @Jitter, Space, look, Enter	Both rows. Space stays plain text. Sent tagging Alice only	Pending chip that settles to your Jitter here once the library lands. Sent tagging Alice and Jitter

Segments 5b and 6 can't show a difference. The bugs that c39512ac6 and 2692f6e54 fix were introduced within this branch, and in this fixture both builds read your agents here when the composer mounts. After an agent is tagged, the right pane refills @Jitter . That is the existing follow-up-draft behaviour.

Not in this PR

• A chip for one of the viewer's agents from another community reads aria-label="Person Rocket" (or "Person Goose"), while agents here read "Agent …". This happens on main too (segments 4 and 5b) and is left unchanged.

🤖 Generated with Claude Code

[kalvinnchau]

🤖 One failure/recovery case remains in the new namesake exclusion.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head 3ee75393a69490361dc55ce3d760f49859322021 against base/merge-base 744278296c255c398b38e8e8fc13df445aa9f8ba. Two findings are attached inline: a P2 where the new cold-library flow hides both same-named agents and leaves an exact typed mention untagged, and a narrower P3 where a cached alias bypasses the loading hold. The P3 is a lower-priority follow-up, not an additional readiness blocker.

The review covered the shared chooser/resolver precedence, full-set ambiguity and longer names, agent-library lifecycle/retry, archive and channel admission, notification/import intent, and regression-test source. This was source-only: no tests, builds, app runs, live interaction checks or CI verification. The PR’s historical validation claims were not rerun. The public PR description was inspected; its attached demo video was not viewed, so the video’s visual/privacy surface remains unverified.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head f0557dc615e139eeb4ab34a77fbf5005f44d7299 against base fb9cd7cf45c6d80368cc79a7c34535f8f3b2fd20, retaining the earlier review’s unchanged coverage. Both previous findings are addressed in source; no new actionable findings in this focused follow-up.

• The P2 late-local-row case now appends the offered agent before newly released elsewhere/directory results, preserving shown rows and the result limit. The added composer cases check the local Space target, pending resolution and sent key.
• The lower-priority P3 alias case now holds queries matching cached profile, verified and local-agent aliases in both chooser and resolver. Added cases cover directory-first arrival, Space/Enter selection and pending resolution.

The follow-up also checked failure/retry behavior and the rebase’s naming/cache integration. Source-only: no tests, builds, app runs, CI verification, visual checks or performance measurements. Author-reported validation was not rerun. The public PR description was inspected; its historical demo video remains unviewed, including its visual/privacy surface. This COMMENT is not runtime validation or merge approval.

[wesbillman]

🤖 Star Lord’s automated source integration follow-up via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head bdb4d9bf483247f9eb686a8f6c0600d23adf9173, live base 3bcf1c4bb1111d924a84d9b710c1da2db80ccf0b, merge-base fd21dba4af2edf63d7b038e3f0f6bdf3703b5df7. Prior unchanged coverage from review 5798537429 is retained; both earlier late-local-row and cached-alias repairs remain intact.

One new P2 is attached inline: exact-public-key mentions of an owned agent elsewhere can return to prose before a successful cold library read completes. This follow-up checked the retained feature patches against the inherited exact-key ranking, team-name rules, directory cache/lookup and shared chooser/resolver lifecycle.

Source-only: no tests, builds, app runs, CI verification, visual checks or performance measurements. Regression-test source and the public PR description were inspected; historical validation claims were not rerun. The historical demo video remains unviewed, including its visual/privacy surface.

[wesbillman]

🤖 Star Lord’s automated source repair follow-up via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head bce71233dc79d86ec35990482819a6f59ed5176a, live base 067db608fbc624e56161ccf3e4ff1c0f11c30baf, merge-base 3bcf1c4bb1111d924a84d9b710c1da2db80ccf0b. The prior P2 (4492884106) is addressed in source; no new actionable findings in this focused follow-up. All seven previously reviewed feature patches are retained, including the earlier late-local-row and cached-alias repairs.

elsewhereMatches now recognizes the eligible agent’s exact public key through the shared parser, so hex, npub and nostr:npub queries reach the chooser hold and resolver’s bounded library wait. The resolver also reads offered candidates and local keys when each decision runs. Added regression source covers an empty directory response arriving before library completion, pending resolution for all three key forms, and the hex chooser/Space path through the Bring prompt. Failure, Retry, timeout and candidate eligibility continue to use the existing owners and policies.

Source-only: no tests, builds, app runs, CI verification, visual checks or performance measurements. Author-reported validation was not rerun. The public PR description was inspected and is unchanged; its historical demo video remains unviewed, including its visual/privacy surface.

[kalvinnchau]

🤖 One narrow exact-key selection regression remains.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head 52794e10267d3fdc59013d017d2eb5cc9f3075c2, target base 067db608fbc624e56161ccf3e4ff1c0f11c30baf, merge-base 3bcf1c4bb1111d924a84d9b710c1da2db80ccf0b. The exact-key namesake gap reported in discussion r4493800251 is addressed in source; no new actionable findings in this focused follow-up. The eight earlier patches and previous repairs are unchanged.

The helper now preserves the specifically identified elsewhere agent despite a local namesake. Both chooser candidate sets depend on the decoded key, and each resolver decision supplies its query. Name filtering, channel admission, local-row precedence for the same key, and the complete-library hold remain intact. The README now explicitly states the key exception. Added regression source covers the completed-library chooser and pending resolution for hex, npub and nostr:npub, ending at the Bring prompt without sending.

Source-only: no tests, builds, app runs, CI verification, visual checks or performance measurements. The author’s reported 1,027 passing tests and fail-before checks were not rerun. The public PR description remains historical and unchanged; its demo video remains unviewed, including its visual/privacy surface.

[tellaho]

🤖 Must-fix — show loading while the toolbar mention picker waits for agents.

With native inventory ready and the local agent library still loading, search the toolbar picker for an agent you own in another community. The new hold hides that agent and matching directory results, but the picker can show “No matching channel members” with no loading feedback after the directory settles. The combined agent status is already ready even though this read is still pending.

Use model.pending for the picker’s existing loading feedback and suppress its empty message during that wait. Cover ready native inventory plus a delayed library read; the matching row should appear when the read completes.

Evidence: new hold, combined status, picker loading, empty message. This conflicts with the design guide’s loading/empty distinction.

Reviewed source at 52794e10267d3fdc59013d017d2eb5cc9f3075c2 and sampled the historical demo (5c9dde762, light theme). Not checked: current app behavior, real notification delivery, cold-start latency, dark theme, narrow/wide layouts, enlarged text, or live keyboard/focus behavior. Tests were not rerun.

[matt2e]

🤖 @tellaho Fixed in fb91f81 (fix(mentions): show the picker loading while held agents wait, not empty) and tightened in e21a524 (fix(mentions): keep the picker's empty message from showing while agents wait), now rebased onto main.

The picker no longer keys its feedback off the combined agent status alone. useMentionChoices exposes agentsPending (the library read that rows are waiting on: inviting, or the other-community hold), the picker's "Loading agents…" shows while that read or the agent list is loading, and "No matching channel members." waits until nothing is pending and the agent read has not failed. model.pending itself is unchanged, so Space keeps tagging a unique local name in invite mode. README section 6 records the loading-over-empty rule. Tests in MessageComposer.test.tsx cover ready native inventory plus a held library read for the toolbar picker and the inline chooser, plus invite mode, a failed read, and membership still loading, each with fail-before evidence in the commit messages.

Side by side, before (left, 52794e102, the revision you reviewed) and after (right, dcae5f69f, the first fix; those are the pre-rebase SHAs of e144f52 and fb91f81). Same scripted steps and timing in both arms: native inventory ready, the legacy agent library read held for ~7 s, the viewer owns Rocket on feedback.communities.buzz.xyz with no same-named agent here, open the toolbar picker and search rocket, let the directory settle with no one, then release the read.

picker-loading-before-after.mp4

Mark	Before (52794e102)	After (dcae5f69f)
Directory still searching, library held	"Searching community…" and "No matching channel members."	"Loading agents…" only
Directory settled with no one, library still held	"No matching channel members." alone	"Loading agents…" only
Library read lands	"Rocket · In feedback.communities.buzz.xyz" row	same row
Enter	@Rocket in the composer	same
Inline @rocket while held (parity)	"Loading recipients…"	"Loading recipients…"

The harness is the real ConversationService composer with the mentions plugin on an in-memory relay session, captured in headless Chromium (light theme, 640×720). It is not the native app, and the recording predates e21a524, which changes the invite, failed-read and membership-loading states but not this scenario (agentsPending is true while the hold waits on the read). Not checked here: dark theme, narrow/wide layouts, enlarged text, or the browser journeys.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No new actionable findings in this bounded follow-up at e21a5246649a3acee5a8da9c953c01c3cfa3f0b5, against base/merge-base 40e6ccb597cf9d8b47f9692de8285458cc304017.

The picker now distinguishes a pending agent read from an empty result when native inventory is already ready, including invite mode, and suppresses empty copy after an agent-read failure. agentsPending remains separate from selection evidence, preserving usable local choices. We checked the two new patches and their interaction with ranking, punctuation selection, candidate completeness, library reconnect/Retry and the existing exact-key/alias/late-row repairs. The nine previously reviewed feature patches remain unchanged.

Source-only: no tests, builds or app workflows were executed, and CI was not verified. Regression bodies were inspected; the new failure-state test sets the library directly, so clicking toolbar Retry and focus recovery remain unexercised. The PR description, author discussion and sampled frames from both attached demos (eleven historical-demo frames and nine loading-repair frames) were inspected. Both demos show older revisions and do not validate this head. Current native/browser behavior, visual states and performance remain unverified.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No new actionable findings in this bounded test follow-up at 3077f0de0a3fdbcfe449bfeb25c08320f1cc22d4, against base/merge-base 58116b826171d190a7376fc57b0509c32c507dde.

The helper now pastes a complete nostr: key after @, avoiding the existing typed-colon terminator, while hex and npub remain typed. Both regression matrices retain their loading, directory-first completion, pending-chip and Bring assertions. We traced the real composer paste path and fixture ordering. All eleven previously reviewed patches and the production mention/input paths remain unchanged; prior source coverage and limitations carry forward.

Source-only: we did not run tests, builds or app workflows, inspect CI logs, or independently verify the commit’s test-pass and React-warning claims. These cases cover pasted nostr: queries, not character-by-character entry. The unchanged public demos show older revisions; browser/native clipboard behavior and prior Retry/focus gaps remain unverified.