Add configurable AO code review backend

[Vaibhaav-Tiwari]

Summary

Add configurable AO code review so a worker's PR can be reviewed by a configured reviewer agent after the PR is raised. The reviewer runs one-shot over the worker's own worktree, the result is posted to GitHub, and the worker picks the feedback up through AO's existing SCM observer → review-nudge path.

Source of truth: this body folds in the V1 design comment (2026-06-11), which supersedes the original HLD. Superseded items are listed under Out of V1 / later.

V1 behavior

• A single per-project config names the reviewer agent(s) by harness (Greptile CLI, Codex, Claude, or any adapter). There is no enabled on/off switch — review is always manual and runs only when triggered for a specific session. If no reviewer is configured, AO uses a default reviewer harness (claude-code).
• The reviewer is a one-shot run over the worker's own worktree: it reads what the worker already has checked out, reviews it, and reports a result. V1 assumes the worker adds no commits during a pass (no SHA-pinned copy, no separate reviewer worktree, no target_sha).
• A review happens only when triggered — no loop, no "until PR is ready" terminal condition. The worker fixes things and the user triggers another pass if and when they want one.
• The reviewer is not modeled as a child session. It is tracked by a separate review table (one row per worker) plus a per-pass review_run.
• The review result is posted to GitHub (PR review/comment); the worker receives it through AO's existing SCM observer → review-nudge path (no bespoke delivery message).

Backend shape

Domain & config

projects config gains a reviewers list (part of the existing per-project config JSON):

{
  "reviewers": [
    { "harness": "codex" }
  ]
}

• No enabled field.
• Empty/unset reviewers resolves to a default of one reviewer with harness claude-code.

Storage

review — one per worker:

Column	Purpose
id	identity
session_id (UNIQUE)	worker session under review; one review per worker in V1. A repeat trigger reuses this row.
project_id	owning project (for CDC)
harness	which reviewer agent reviews this PR
pr_url	the PR under review (recorded, not part of the key yet)
created_at, updated_at	when set up / last touched

review_run — per pass:

• id, review_id, session_id (worker), harness, pr_url
• status (pending | complete | failed)
• verdict ('' | approved | changes_requested)
• iteration
• created_at, updated_at

The review body is not persisted in V1 — it flows from ao review submit to GitHub.

Service layer

A ReviewService responsible for:

• loading reviewer config (with default resolution);
• creating/updating the review row and per-pass review_run;
• resolving the reviewer harness and running it one-shot over the worker's worktree;
• receiving submitted review results;
• posting the result to GitHub via a minimal SCM post capability.

CLI & HTTP

POST /sessions/{workerId}/reviews/trigger
GET  /sessions/{workerId}/reviews
POST /sessions/{workerId}/reviews/submit

ao review trigger <worker-session-id>
ao review submit --verdict changes_requested --body review.md
ao review list <worker-session-id>

CLI and HTTP stay thin; the daemon service owns behavior.

Out of V1 / later

These items from the original HLD are intentionally deferred:

• sessions.parent_session_id, sessions.kind = reviewer (reviewer is not a session).
• Disposable pinned reviewer worktree and target_sha (reviewer reuses the worker worktree).
• enabled config flag.
• Auto-trigger on PR head SHA advance.
• Derived in_review status.
• Greptile no-workspace / reviewer-only adapter; reviewer read-only permission mode.
• Relaxing the review key to one-per-(session_id, pr_url).

[neversettle17-101]

V1:
AO can run a code review of an agent's PR, by a review agent of your choice, after the PR is raised. A single per-project config names the review agent(s) (harness — Greptile CLI, Codex, Claude, or any adapter). There is no on/off switch: review is always manual and runs only when triggered for a specific session.

The review agent is a one-shot run over the worker's own worktree — it reads the code the worker already has checked out, reviews it, and reports a result. For V1 we assume the worker adds no commits during a pass (it reviews the worktree as-is; there is no SHA-pinned copy and no separate reviewer worktree to create or clean up). Every review it produces is fed back to the worker to act on. A review happens only when the trigger is selected — there is no loop and no "until PR is ready" terminal condition; the worker fixes things and the user triggers another pass if and when they want one.

Few comments based on the updated design for V1

• For the per-project config drop the enabled config.

• The reviewer will not be a child session instead it should be a separate table
  | Column | Purpose |
  | :---- | :---- |
  | id, session_id (worker, UNIQUE), project_id | identity + who/what is reviewed (project_id for CDC); one review per worker in V1 |
  | harness | which review agent reviews this PR (Greptile | Codex | Claude | …) |
  | pr_url | the PR under review |
  | created_at, updated_at | when review was first set up / last touched |
  session_id is UNIQUE in V1 — one review per worker. A repeat trigger reuses this row; the per-pass detail lives in review_run. (pr_url is recorded but not part of the key yet; relaxing to one-per-(session_id, pr_url) is a later change with no reshape.)

• Reviewer agent will work on the same worktree thus they don't need to have a target SHA.

• CLI & HTTP

POST /sessions/{workerId}/reviews/trigger 
GET  /sessions/{workerId}/reviews
ao review submit --verdict changes_requested --body review.md

• Once the review is generated -- reviewService will persist in the reviewRun and we want to post the review comments on github. If the review comments are embedded from github to worker session then resue that -- but if that is not possibel we would need to reuse lifecycle manager and embed it to the worker session using sendOnce method

[neversettle17-101]

Updated the issue body to fold in the V1 design comment end-to-end, so the body is now the single source of truth for implementation. The original HLD items that V1 supersedes (child-session reviewer, pinned reviewer worktree + target_sha, enabled flag, auto-trigger, derived in_review status) are preserved under the new Out of V1 / later section. Original V1 comment kept for history.