feat(review): per-PR review targeting and a Run review action on PR cards - #6136

A requested review now fetches the worker's open PRs fresh from the provider
before deciding what is due, reusing the session claim path, so the pass
reviews the commit really on the PR instead of whatever the SCM observer last
saw. Refresh is bounded (10s) and best effort for tracked PRs: on failure the
trigger falls back to stored facts and REVIEW_HEAD_NOT_OBSERVED. Auto-review
does not refresh.

ao review trigger --pr <url> (request prUrl, matching #6136) reviews only that
PR. If AO does not track it yet it is attached first, never taken over from
another active session (409 REVIEW_PR_OWNED_BY_OTHER_SESSION). With no tracked
PR, the error now points at --pr. Workers are told to pass --pr right after
opening a PR, and to re-trigger when ao review ls shows failed or cancelled.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…6262)

* feat(review): let agents request AO native review of their own PRs

Workers and orchestrators now start AO's native reviewer directly instead of
spawning a generic worker to review:

- ao review trigger/ls/cancel default to the calling AO session; trigger takes
  --agent/--model/--effort per pass, refuses inside reviewer panes, records
  the pass as source "agent", and turns on the session's review auto-inject
  unless --no-inject.
- A head that is already being reviewed or already has a review is rejected
  with 409 REVIEW_ALREADY_RUNNING / REVIEW_HEAD_ALREADY_REVIEWED; --rerun
  reviews it again (migration 0175 narrows the unique key to running passes)
  or adds a reviewer with another agent alongside a running one.
- Reviewers with different agents run concurrently: trigger no longer tears
  down a working reviewer, stale-pane cleanup and cancel cover every reviewer,
  and the list API exposes activeReviewers.
- Approved verdicts are delivered to the worker when auto-inject is on, with
  wording that an AO approval is not a GitHub approval and never authorizes
  merging.
- New project setting workersRequestReview chooses whether workers request a
  review when ready or ask first; the orchestrator prompt no longer routes
  code review to spawned workers.
- With no reviewer configured, the reviewer defaults to the project's default
  worker agent and its model/effort, never its permissions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(review): keep parallel reviewers alive across worker restore

Restoring a worker restored its selected reviewer by tearing down every
other reviewer and cancelling its passes as "reviewer agent was switched".
With concurrent reviewers that cancelled a reviewer still running alongside
the selected one. Restore now releases only idle reviewer panes, and fails
another reviewer's running passes only when its pane did not survive.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(review): show agent-started reviews live and keep Chat reviewers on restore

Found testing #6262 with real agents in a dev build:

- The inspector never refreshed when a review started or finished outside
  the window (an agent's ao review trigger, auto-review, a reviewer's
  submit): local review_run CDC only refreshed workspaces. Refresh that
  session's reviews on review_run_created/updated.
- When an agent picked a non-default reviewer (--agent codex) the selected
  reviewer stayed current although idle, so the working reviewer had no tab.
  List now reports the working reviewer as current when the selected one is
  idle.
- Worker restore cancelled a running Chat reviewer as having no pane; Chat
  reviewers have none by design and are recovered by RecoverChatReviewers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(review): deliver AO reviews to workers only through PR comments

One channel per review source. The worker now hears AO's reviewer the same
way it hears any reviewer: the SCM observer forwards the reviewer's inline
GitHub comments. AO no longer sends a separate verdict message, which
delivered every review twice and turned optional inline suggestions into
required work.

- Remove post-submit verdict delivery (lifecycle ApplyReviewBatch and the
  review service's delivery step); results are recorded and stay complete.
- The reviewer prompt requires an inline comment for every finding that needs
  a change, including design-level ones, and keeps optional suggestions in
  the summary, because the worker receives only inline comments.
- Worker instructions, CLI output, skill and docs say findings arrive as PR
  review comments and the verdict is in ao review ls.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(review): explain a pushed head AO has not observed yet

Right after a push, ao review trigger could refuse with 'already reviewed,
push new commits' because the SCM observer (30s cadence) had not seen the new
commit, so the worker was told to push what it had just pushed. When the head
AO knows is already reviewed, compare it with the workspace's remote-tracking
ref for the PR branch; if the worker has pushed past it, return 409
REVIEW_HEAD_NOT_OBSERVED with both commits and a retry hint. No automatic
retry: the worker retries once AO has seen the push.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(review): refresh PRs from the provider and accept --pr on trigger

A requested review now fetches the worker's open PRs fresh from the provider
before deciding what is due, reusing the session claim path, so the pass
reviews the commit really on the PR instead of whatever the SCM observer last
saw. Refresh is bounded (10s) and best effort for tracked PRs: on failure the
trigger falls back to stored facts and REVIEW_HEAD_NOT_OBSERVED. Auto-review
does not refresh.

ao review trigger --pr <url> (request prUrl, matching #6136) reviews only that
PR. If AO does not track it yet it is attached first, never taken over from
another active session (409 REVIEW_PR_OWNED_BY_OTHER_SESSION). With no tracked
PR, the error now points at --pr. Workers are told to pass --pr right after
opening a PR, and to re-trigger when ao review ls shows failed or cancelled.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(review): say a not-yet-visible push clears in seconds

A trigger now refreshes the PR from the provider itself, so the only
remaining window is the provider's own lag right after a push.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(review): address AO review of worker-requested review

- Drop the local-git pushed-head probe. The trigger already re-reads the
  PR from the provider, and the probe misfired when the provider was ahead
  of the workspace's tracking ref. The already-reviewed message now says to
  retry after a just-made push.
- Keep the selected-reviewer fields meaning the selected reviewer. Clients
  open a working non-selected reviewer from activeReviewers instead.
- Remove trigger entry points with no production callers and the unused
  delivered-run writer; the delivered status stays for existing rows.
- Cancel legacy harness-less runs once, not once per run.
- Fix a stale ApplyReviewBatch mention.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Prateek K <prateek@Prateeks-Mac-mini.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>