Trim launch guidance to the plugin's skills, behind a setting

[selfcontained]

What

buildLaunchGuidance injects a set of startup rules into every agent, unconditionally, in every session. Most of that text is already stated somewhere the agent reads anyway. This PR does two separable things:

1. Unconditional — shortens the Autonomous Review rule for everyone, because Dispatch's own review injections already re-deliver most of it.
2. Behind a new setting (Settings → Agents → Launch guidance, off by default) — replaces the verbose rules with short pointers, so both variants can soak side by side instead of hard-cutting over.

Where the trimmed detail actually lives

Two different things carry it, and the distinction drives the design:

The MCP tool schemas — dispatch_pin's own description already lists every pin type, explains shortcut/confirm/disabled, and says to pair a blocking shortcut with waiting_user. dispatch_event's enumerates the status types. The guidance was restating a description the agent already has. This half does not depend on the plugin — every agent gets those schemas.

Plugin skills (#952) — ui-validation + sharing for the Playwright methodology, review-workflow for the create_pr routing line. This half genuinely needs the plugin installed, which is why the setting is worded as an assertion about it.

Nothing carries the no-task guardrail, so it stays verbatim in both states.

The trimmed ruleset, in full

1. No task, no work. …                          (verbatim)
2. Name the session with dispatch_rename_session once the topic is clear …
3. Report status with dispatch_event as you work and before your final response —
   blocked means genuinely stuck, not an error you're about to fix. Your reported
   status is verified against session activity and auto-corrected.
4. Surface important data to the user with dispatch_pin — anything they may need to
   read or copy — and use shortcut pins to offer a next step or ask them to pick
   between options.
5. Share artifacts with dispatch_share — screenshots, logs, reports. A file path
   pasted into chat is not a deliverable.
6. Autonomous Review … (only when enabled)

The two pin rules fold into one. Kept the two things no schema states: that blocked means genuinely stuck rather than an error you're about to fix, and that reported status is verified and auto-corrected. The dispatch_share nudge survives the trim deliberately — that habit already failed when stated in two always-on places, so it's the one tool-routing rule with a demonstrated failure history.

Autonomous Review, shortened for everyone

Two thirds of the old 830-char block was reactive — "after feedback arrives, call dispatch_review_list_feedback", "ask the reviewer to verify", "a clean zero-item approval needs no action" — and Dispatch already re-injects each at the moment it applies (buildLaunchPersonaResponseText, reviews/injection-prompts.ts). What survives is the proactive gate: nothing can deliver "before you emit done, do X" at the right moment, because the moment is the agent's own judgment that it's finished.

One durable pointer to dispatch_review_list_feedback stays: injection is best-effort (sendPromptBestEffort swallows the failure when the parent has no live session) and nothing replays it, so the agent needs one way to find a review submitted while it was down.

Scope guards

• Job runs are untouched, and the manager skips the settings read entirely for them — every rule there is a runtime protocol obligation. A test asserts both variants are byte-identical.
• Only Claude Code and Codex agents are ever trimmed. Opencode and Cursor have no plugin, so they keep the full guidance regardless.
• The setting is an assertion, not detection. Dispatch cannot read ~/.claude/settings.json or ~/.codex/config.toml. Default off; the copy says what turning it on without the plugin costs.
• Guidance is composed at launch, so a flip only affects agents started afterwards.

Verification

Live from the same dev API, autoReview on, read from the launched claude process's own argv — not a unit test:

2590 chars  setting off
1336 chars  setting on      (-48%)

And for the unconditional half, measured the same way before it was added: 2678 → 2270, −408 chars for every user regardless of the setting. With the setting on, a job-run agent still gets its untouched 620-char ruleset.

• Endpoint: GET default false, POST true → GET true, malformed body → 400.
• Playwright: toggle both directions with each persisting across a reload; checkbox disabled until the GET lands and after a failed one; aria-label/aria-describedby wiring; three rapid clicks leave UI and server on the same value; 390×844 mobile pass with no horizontal overflow.
• pnpm run check, finalize:web, test (3646), test:e2e (180) green locally.

Reviewed by backend-security-review, frontend-ux-review, architecture-review, and product-review — 8 findings, all resolved.

🤖 Generated with Claude Code