feat: title each Decision evaluation row by what it judged

[zfy0701]

Summary

A Recent evaluations row showed only a time, an answer, and an outcome, so a page of one routing's rows looked alike and gave no reason to open any of them. This adds a one-line title to every gate, shared-bot routing, and repository routing summary row, so the console can show what each evaluation judged. The console change comes in a follow-up.

• protocol: DecisionEvaluationRecord and DecisionRoutingEvaluationRecord gain title: string | null (max 256).
• daemon: summaryRow derives the title from the frozen input it already parses. A code-host subject reads as its session title does (PR #42: Fix the parser, MR !7, Issue #3: …); anything else uses the current message's first non-blank line. Both reuse the session title clamp (80 characters). Once retention strips the input the title is null, so it never outlives the bodies. Agent model selection keeps its target and gets no title.
• control plane: a title quotes the input, so each list route returns it only to a caller the detail would admit; other summary readers get the row with title: null. For gates and shared-bot routing that is the conversation audience with bodies: true (edit access to the consumer agent before any session names an audience); for repository routing it is edit access to every member agent.
• docs: decisions.md §9.5 and the three list routes' OpenAPI descriptions say what the title is and who sees it.

Performance

The daemon already reads and parses each row's input to fill messageId, so deriving the title adds no query and no parse. The list read was measured against a test daemon's store with a code-host routing of 228 evaluations (about 17 KB of stored bodies per row): 3 ms for a 10-row page and 5 ms for a 50-row page, query plus parsing, so reading slim columns instead would save only 2–4 ms against a round trip through the CP. The title adds at most about 240 bytes per row; the existing 32 KiB page cap and cursor absorb it.

Testing

• daemon: titles from a pull request, a merge request without a title, a push (first non-blank line), an over-long message, a blank message, a reserved row, an expired row, and a router row
• control plane integration: a viewer reads summaries without titles before any session names an audience and with them once one does; a collaborator with edit access reads them titled; repository routing withholds titles from a member who cannot open details
• protocol, web decision tests, typecheck, lint

🤖 Generated with Claude Code . Claude Opus 5.5

[agentconnect-md-test]

The title derivation and retention behavior look sound, but the conversation list can expose a title after the caller loses edit access during the daemon read. The post-read summary check reloads the conversation; the new title check then uses the pre-read snapshot. Please run the body-access check against the refreshed conversation, as the detail route does.

_{sent by review-bot (Codex · gpt-6-sol) · open in session}

[agentconnect-md-test]

Approved. The conversation evaluation list now checks title access against the conversation reloaded after the daemon reply, so a caller who loses edit access during that read receives title: null. The added regression test covers that case. The title derivation, retention behavior, and other two list routes remain consistent with their detail-access rules.

_{sent by review-bot (Codex · gpt-6-sol) · open in session}

[zfy0701]

@aaa please review this pull request.

[agentconnect-md-test]

Found one compatibility issue: required title fields make the updated control plane reject non-empty evaluation pages from older daemons. Accept missing titles as null or negotiate the new response shape.

The earlier access issue is fixed. Tests could not run because pnpm’s store database failed to open.

Formal review was not submitted because the PR is already merged.

_{sent by aaa (Codex · gpt-6-astra) · open in session}