Make mention choices consistent and stable

[loganj]

🤖
Stacked on #257. Review that one first. This PR's diff has only the chooser changes.

Summary

• The toolbar picker and inline @ completion now use the same rules, so the same search gives the same list in both places.
• The chooser opens with no wait. Before, the whole list showed "Loading recipients…" until a network search finished. This happened each time you opened it and each time you typed.
• Rows do not move while you look at them.
• Space can finish a mention when the name you typed is complete and unique.

The rules

Who is in the list

• Channel members, and the agents you can mention here.
• In channels and forums, also people from the community directory (the list of everyone in this community). DMs and sessions do not show outside people.
• Not shown: archived identities, but you always see yourself. The chooser is empty in archived or read-only channels.

Order

1. Members come first. People outside the channel come next.
2. In each group, better matches come first:
   1. The name is the same as your search.
   2. The name starts with your search.
   3. A word in the name is the same as your search.
   4. A word in the name starts with your search.
3. The name you see counts first. Other known names for the same identity count only if the visible name does not match.
4. If two agents match equally, agents you own come first.
5. Agents with the same name stay together as one group. The group goes where its first visible name sorts. Inside the group, these come first, in this order: agents you chose recently in this channel, managed agents, then agents that are online.
6. People never join an agent group. On an equal visible name, a person comes before the agent group.
7. Last, sort by the visible name, then by key. The order is the same for any input order.

What does not match

• Public keys.
• Text inside a word. For example, "oney" does not find "Honey".

When the list changes

• Members and agents show at once.
• People from the directory are added at the bottom when the search returns.
• A row never moves after it shows. The order resets only when you change the search or reopen the chooser.
• If someone becomes unavailable, their row stays in place but you cannot choose it. So Enter or Tab cannot select the wrong person.
• The list shows at most 50 rows. If there are more, it asks you to type more.

Directory search

• It starts after you stop typing for 200 ms.
• While a new search runs, people from the last search who still match stay in the list. A "Searching community…" line shows below the list.
• Searches that find people are kept for the session, so going back to an earlier search is instant. Failed searches are not kept, and Retry searches again.
• A search that finds no one is not kept. A new @ or a reopened picker searches again. Typing more after that search does not search, because a longer name cannot match.
• In the inline @ menu, text after an unknown name is treated as normal typing. The menu closes, and no search runs.

Space

• Space selects a person only if exactly one identity has the name you typed.
• Space does nothing if another name starts with your text plus a space. For example, "Honey" does not select when "Honey Bee" is also in the list.
• For other cases, use Tab, Enter, or a click.

Screenshots

Members come first. People outside the channel come after them.

A directory search is still running. Members show at once, and a "Searching community…" line shows below them until people from outside the channel arrive.

Two identities have the same name. Each row shows a label and a short key, so you can tell them apart.

A person leaves the channel while the list is open. Their row stays in place, and you cannot choose it.

Details

• A selected mention stores only the key and name. Selection checks again that the person is still available.
• The full rules are in docs/agents.md under "Chooser rules".

[wesbillman]

Star Lord’s automated source review via Wes’s account.

Reviewed head e0e2e7a6f6693ef6a850ea34014e851e7d7be9d5 against base 216a81909210ef006a4d0040e82c366f3cda222d (the #257 stack base), including chooser/ranking, directory lifecycle, completion-host integration, insertion/send callers and changed test sources. Two actionable P2 findings are recorded inline. This is non-blocking Comment feedback, not approval or a changes-requested review.

Validation limits: source analysis only, with an independent source cross-check. No PR code, tests, builds or app workflows were executed; CI was not assessed. Browser editing/focus, actual directory/relay behavior and exact-head test results remain unverified. The three added browser cases describe native editing/focus or representative integration boundaries, but the PR description does not record exact-head validation/deferred-check evidence; test source is not a passing result.

[wesbillman]

Automated source review — Star Lord (via Wes’s account)

No new actionable source defects found in this revision.

• The prior mixed human/agent comparator cycle is addressed by per-choice sort keys and stable same-name blocks. The new permutation coverage targets both the reported cycle and interleaved human/agent cases.
• The prior loss of directory fallback on every inline keystroke is addressed by session-scoped chooser state. I traced the real ComposerCompletions remount boundary and the added host-level regression test, along with picker selection, live eligibility checks, and send-entry archive handling.
• Ranking was assessed against the current docs/agents.md contract: on otherwise equal matches, viewer-owned agents precede humans as well as other agents. This is a documented behavior assessment, not a separate product-approval claim.

Reviewed head: 13d0b85921748f0f848c2a76cb10f89426a6804e
Base: 17de590f1d574a9a7f2d9af218edfe75fc6f8caa
Diff merge-base: 24fcb1ed31e48558d0b127850ced79fbc6c80a7b
Prior reviewed head: e0e2e7a6f6693ef6a850ea34014e851e7d7be9d5

Validation limits: source-only review of pinned Git objects, with independent ranking review. No PR code, tests, builds, or app workflows were executed; CI was not assessed. Regression tests were inspected, not run. Browser editing/focus, live directory behavior, and packaged-app acceptance remain unverified. This non-blocking COMMENT is not a GitHub approval or merge authorization.

[wesbillman]

Automated source review — Star Lord (via Wes’s account)

One P3 recovery-contract gap is recorded inline: the new empty-result refresh policy does not reach the persistent toolbar picker. This is an incomplete part of the new fix, not a claim that this behavior first appeared in the previously reviewed revision.

Follow-up scope: the 2adc990 archive/prose-search integration and its supported callers. The archived-viewer exemption is consistently applied to candidate selection, disabled-row explanations and send entry; the prior transitive-ranking and inline-remount fixes remain intact. The directory/prose lane received an independent source cross-check, reconciled against the picker’s actual lifetime. No unrelated earlier review areas were reopened.

• Head: 2adc990779f86cb707d5ba47486fbb78a981aec9
• Base: 734949a62229c56b2e3368701a4b4d5413ac555a
• Diff merge-base: 67fe94c750fc739e49caf3ee57e0f0940944cb04
• Earlier feedback-covered head: 13d0b85921748f0f848c2a76cb10f89426a6804e

Validation limits: source-only review of immutable Git objects; no PR code, tests, builds or app workflows were executed. Pinned diff whitespace checks passed. One exact-head CI snapshot showed JavaScript, browser measurements, DCO and security checks passing, with Rust and several browser shards still running; Windows was skipped. I did not wait or poll. Browser editing/focus, live directory behavior, complete CI and explicit human acceptance remain unverified. This non-blocking COMMENT is not approval, Request Changes, or merge authorization.

[wesbillman]

Automated source review — Star Lord (via Wes’s account)

No new actionable findings in this follow-up. The earlier P3 picker-reopen finding is addressed in source:

• MentionPicker.tsx:42–51,85–91 gives each opening its own directory lifetime. useMentionDirectory.ts:74–109,135–145 binds uncached empty/error state and request cleanup to that lifetime while retaining the intended non-empty session/query cache.
• The added persistent-picker regression (session-agents.test.tsx:1837–1883) closes and reopens the same mounted picker, expects another directory call for the same query, and checks that the newly available person appears.
• The composer-lab fixture now supplies archives.state/ensure/refresh; returning unknown for its unavailable archive snapshot matches the real read contract without weakening production archive filtering.

Scope: the four-file follow-up since 2adc990779f86cb707d5ba47486fbb78a981aec9 (63 additions / 5 deletions), its supported callers, cache/close/reopen lifecycle, and regression source. No unrelated earlier review areas were reopened. The target-branch changes since the merge-base do not overlap these four files.

• Head: 5075f7625ba6c46ce1be690d5e08e33d0f2b14be
• Base: 734949a62229c56b2e3368701a4b4d5413ac555a
• Diff merge-base: 67fe94c750fc739e49caf3ee57e0f0940944cb04

Validation limits: source-only review of immutable, blob-verified extracts; no PR code, tests, builds, or app workflows were executed. The four-file follow-up has no whitespace diagnostics. One exact-head CI snapshot showed browser measurements, DCO and security checks passing; JavaScript, Rust and all six browser-journey shards were still running, and Windows was skipped. No waiting or polling. Complete CI, actual browser/native interaction, live-directory recovery and explicit human acceptance remain unverified. This non-blocking COMMENT is not approval or merge authorization.