* fix(review): message worker on changes_requested instead of relying on SCM poll (#337)
review.Engine.Submit previously only persisted the verdict/body and left the
worker to learn about requested changes via the SCM poll loop, which is gated on
GitHub's reviewDecision and never reaches CHANGES_REQUESTED for self-reviews or
COMMENT-state reviews. Submit now nudges the worker's live pane directly via
ports.AgentMessenger (the same mechanism lifecycle uses) whenever the verdict is
changes_requested.
Extended flow: the reviewer reads back the GitHub review id it posted and passes
it through `ao review submit --review-id`; the id is stored on the review_run row
(new column + migration 0016) and included in the worker message so the worker
knows exactly which review to address and reply to.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): mark worker nudge as AO internal review, ask to reply + resolve
Distinguish the AO internal review nudge from the external GitHub-reviewer
feedback the lifecycle SCM loop relays. For an AO review the worker is now asked,
once it has pushed its fix, to reply on the review referencing its id with what
it changed and resolve the inline review comment threads it addressed (the
reviewer posts inline comments, so the per-finding threads are resolvable via
resolveReviewThread; the top-level review object is not, hence the reply).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): generalise the changes-requested worker nudge wording
Drop the "not an external GitHub PR reviewer" aside and the assumption that the
worker pushes a fix — it may resolve the feedback without code changes. The nudge
now reads "Review the feedback below and address it" and asks the worker to reply
with how it addressed the review and resolve the threads it addressed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): harden the review-id read-back against array order and empty results
The reviewer read the just-posted review id with `--jq '.[-1].id'`, which trusts
the REST API to return reviews in ascending submission order and errors when no
review exists. Review ids are monotonic, so select the highest id instead and
emit nothing when the list is empty: `--jq 'map(.id) | max // empty'`. Update the
matching `--review-id` flag help.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): post the review via gh api and capture its id from that response
The reviewer must use `gh api --method POST .../reviews` to attach inline
comments anyway (`gh pr review` cannot), and that response already contains the
created review's id. Capture `.id` from that single call instead of a second
read-back, dropping the array-ordering/pagination heuristics entirely — the id is
the exact review just created.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): send the review as a JSON body so inline comments are a real array
gh api -f/-F cannot build an array of objects: comments[][path] is sent as a
literal key, so the inline comments are dropped — defeating the reason for using
gh api over gh pr review. Post the review via --input JSON instead, keeping the
.id capture and the approve/COMMENT fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore: format with prettier [skip ci]
* chore(review): drop accidentally committed reviewer scratch, write review out of tree
review.md was the reviewer agent's own writeup, swept onto the worker branch by a
stray `git add -A` in 5df20c9. Remove it, gitignore `/review.md` as a backstop,
and change the reviewer prompt to write its review to a temp file outside the
checkout instead of into the worktree (where it could be committed onto the
worker's branch).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): message the worker before marking the run complete
If messenger.Send failed after UpdateReviewRunResult had already flipped the run
to complete, a retried `ao review submit` tripped the status='running' guard and
could never record the result. Send first; only mark the run complete once the
worker has been notified, so a failed send leaves the run retryable. A landed
message followed by a failed DB write degrades to one extra nudge on retry — the
same trade lifecycle's sendOnce makes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* feat(review): accept the review body on stdin so the reviewer writes no file
`ao review submit --body -` now reads the review from stdin, and the reviewer
prompt pipes its writeup via a heredoc instead of writing a file. Previously the
reviewer wrote review.md into its checkout to pass as --body, which could be
committed onto the worker's branch (as it just was). A file path is still
accepted for backward compatibility.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): always post approvals as COMMENT, drop the APPROVE attempt
The reviewer posts from the PR author's own GitHub account, so event=APPROVE
always 422s. Drop it: request changes with REQUEST_CHANGES, approve with a
COMMENT-event review whose body states it is an approval.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(review): post every review as event=COMMENT (author can't APPROVE or REQUEST_CHANGES own PR)
The reviewer posts from the PR author's own account, where GitHub rejects both
APPROVE and REQUEST_CHANGES. Always post a COMMENT-event review and state the
verdict in the body; the machine-readable verdict still reaches AO via
`ao review submit --verdict`.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(cli): accept underscore flag names on `ao review submit`
Reviewer agents routinely invoke the submit command with --review_id
instead of --review-id, which cobra rejected as an unknown flag and
dropped the GitHub review id from the worker notification. Normalize
underscores to hyphens on the command's flags so both spellings resolve
to the same flag.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix: sanitize review id in worker notifications
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Vaibhaav <user@example.com>
* fix(review): serialize concurrent triggers per worker to stop double-spawn
Engine.Trigger was a read-then-write (idempotency check -> reviewer spawn ->
InsertReviewRun) with no serialization and no backing constraint. Two near-
simultaneous triggers for the same worker at the same head SHA both passed the
GetReviewRunBySessionAndSHA check, both spawned a reviewer against the same
deterministic review-<id> handle, and both inserted a running run for one commit.
Add a per-worker keyed mutex (lockWorker) held across the whole Trigger body, so
the loser re-reads the freshly-recorded run and short-circuits to Created:false
instead of spawning. Back it with a partial unique index on
review_run(session_id, target_sha) (migration 0013) as a cross-restart safety
net; rows with an empty target_sha (head not yet observed) are excluded so they
are not blocked.
Adds a concurrency test asserting N simultaneous triggers spawn once and record
one run.
Closes#242
* fix(review): make migration 0013 dedup-safe and handle the unique conflict in Trigger
Pre-#242 daemons can already hold duplicate (session_id, target_sha)
review_run rows, on which CREATE UNIQUE INDEX fails and wedges startup.
Migration 0013 now collapses each duplicate group to a single survivor
(a completed pass over a still-running one, then newest by created_at)
before building the index.
Trigger now treats a unique-constraint hit as a fallback rather than an
error: InsertReviewRun maps it to the new domain.ErrDuplicateReviewRun
sentinel, and Trigger re-reads GetReviewRunBySessionAndSHA and returns
that run with Created:false instead of surfacing a raw error after the
reviewer may already have launched.
* feat(review): configurable AO code review backend (V1)
Add per-project configurable code review of a worker's PR. A reviewer
agent runs one-shot over the worker's own worktree and posts its result
to the PR; the worker picks the feedback up through the existing SCM
observer review-nudge path.
- domain: ProjectConfig.reviewers (+ default reviewer harness), Review /
ReviewRun types and verdict/status vocab.
- storage: review + review_run tables (0011), sqlc queries, store methods.
- service/review: rewrite the in-memory stub as a persisted ReviewService
(Trigger/Submit/List) with a reviewer Runner over agent resolver +
runtime; ports.PRReviewPoster implemented on the GitHub adapter.
- http: session-scoped routes POST /sessions/{id}/reviews/trigger,
POST .../submit, GET .../reviews; regenerated OpenAPI + TS types.
- cli: ao review trigger|submit|list.
- frontend: adapt ReviewDashboard to the per-worker reviews API.
Closes#192
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): address review — drop submit/poster/CLI, default reviewer to worker harness
Per PR #197 review feedback:
- Reviewer agent posts its review to the PR itself, so remove the
ports.PRReviewPoster port, the GitHub review poster, the submit HTTP
route + DTO, and the service Submit method (#1, #4, #7).
- Trigger spawns the reviewer agent over the worker's worktree with its
own review prompt, mirroring the session launch flow (resolve agent by
harness -> argv -> runtime.Create) (#8, #9).
- Default reviewer harness reuses the worker's harness when supported,
falling back to claude-code; reviewer config stays independent of the
worker override (#5, #6).
- Drop the `ao review` CLI for this PR's scope (#2, #3).
Regenerated OpenAPI + TS types.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* feat(review): restore ao review submit (records verdict+body in AO)
Per maintainer request, bring back `ao review submit`. AO records the
reviewer's verdict and body on the review_run and marks the pass complete;
it does not post to GitHub — the reviewer agent posts its review to the PR
itself.
- storage: add review_run.body (0011), persist via Insert/UpdateReviewRunResult.
- service: restore Submit (no SCM poster) storing verdict + body.
- http: restore POST /sessions/{id}/reviews/submit + SubmitReviewInput.
- cli: ao review submit [worker] --verdict --body (worker from arg/--session/$AO_REVIEW_WORKER).
- runner: reviewer prompt instructs posting to GitHub and recording via ao review submit.
Regenerated OpenAPI + TS types.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): move reviewer runner to its own package; sharpen prompt
Per PR #197 review:
- Move the concrete reviewer runner out of the service layer into a new
internal/review_runner package (package reviewrunner), beside other
orchestration packages like session_manager. The service keeps only the
Runner interface + RunSpec it depends on; the agent-resolver + runtime
launch flow lives in review_runner.
- Sharpen the reviewer prompt: tell the agent to diff against the PR base,
focus on high-confidence findings, post via `gh pr review`, and record
the result with `ao review submit`; review-only (no commits/edits).
- Add unit tests for the runner.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): simplify review_run schema; provider-agnostic reviewer prompt
Per PR #197 review:
- review_run: status default 'running' (drop 'pending'), drop CHECK
constraints on status/verdict, drop the updated_at column and the
session/iteration index. Propagated through queries, domain, store,
service, and tests.
- Reviewer prompt no longer hardcodes GitHub/gh commands — it instructs the
agent to use whatever review tooling the provider offers, keeping the
flow extensible across SCM providers.
Regenerated sqlc + OpenAPI/TS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): launch reviewer before persisting the run
Trigger now spawns the reviewer agent first and then writes the review_run
with a status derived from the launch outcome (running on success, failed
if it never started), instead of inserting a running row and correcting it
to failed afterwards.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): pluggable reviewer registry distinct from worker harnesses
Reviewers are now their own pluggable adapter set, separate from the worker
agent registry — adding a reviewer (claude-code today, greptile tomorrow) is
a one-line registration that does not widen the worker harness vocabulary,
and a worker harness does not automatically become a valid reviewer.
- domain.ReviewerHarness: a distinct vocabulary (AllReviewerHarnesses) with
its own IsKnown; ReviewerConfig/Review/ReviewRun use it. ResolveReviewerHarness
reuses the worker harness only when it is itself a supported reviewer, else
falls back to claude-code.
- ports.Reviewer: a reviewer-specific contract (ReviewCommand → argv + env)
that models one-shot / non-prompt CLIs natively instead of forcing every
reviewer through the worker's interactive GetLaunchCommand(Prompt:...).
- internal/adapters/reviewer: a separate registry + resolver (mirrors the
worker agent registry) with the claude-code reviewer adapter, which owns the
review prompt and reuses the worker claude-code launch construction.
- review_runner resolves via the reviewer registry (not the worker
AgentResolver) and merges AO_REVIEW_WORKER into the adapter's env.
- daemon wires the reviewer resolver. Registry/domain parity is test-enforced.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* test(review): cover run-scoped reviewer submit
* fix(api): update generated review submit schema
* refactor(review): split core engine (internal/review) from API service
Move the review orchestration (Trigger/Submit/List, run-id generation,
deps, RunSpec/Runner, sentinels) into a transport-independent core package
internal/review (Engine). internal/service/review is now a thin API-flow
boundary: the controller-facing Manager interface + a Service that delegates
to the engine + error re-exports.
This keeps the service layer to API concerns and lets the same engine back a
future in-process CLI trigger without going through HTTP. review_runner now
depends on the core package; daemon builds the engine and wraps it in the
service. No API/schema changes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* feat(review): commit-aware trigger, reviewer handle for UI, no env vars
Reworks the review trigger lifecycle and drops env-based coupling:
- review_run gains target_sha (the reviewed commit) and drops iteration.
A repeat trigger for the same PR head short-circuits to the existing run.
- review gains reviewer_handle_id: the live reviewer pane's runtime handle,
reused across passes and exposed in the reviews API so the UI can attach
its terminal over /mux.
- Trigger flow: if a live reviewer pane exists and a new commit arrived,
message it to re-review; otherwise spawn a fresh reviewer. The run is
recorded only after the reviewer is launched.
- No environment variables: the reviewer adapter embeds the explicit
`ao review submit --session <w> --run <id>` command in the spawn prompt
and the re-review message. CLI submit requires --run/--session (no env
fallbacks).
- Merge review_runner into internal/review as a Launcher (spawn/notify/alive).
- Trigger returns 201 for a new pass, 200 when reusing an existing run.
Regenerated sqlc + OpenAPI/TS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): author the reviewer prompt centrally, not in the adapter
Mirror the worker model (session_manager builds the prompt; adapters just
place it via LaunchConfig.Prompt). The reviewer prompt now lives in
internal/review/prompt.go and is passed through ports.ReviewInvocation.Prompt;
the claude-code reviewer adapter just feeds inv.Prompt to its launch command
and returns it as the re-review message. One-shot CLI reviewers may ignore it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* refactor(review): split reviewer prompt into system+task, mirroring buildSpawnTexts
Mirror session_manager.buildSpawnTexts for the reviewer: a standing role goes
in the system prompt, the per-pass task (PR/commit + exact `ao review submit`
command) goes in the user prompt. internal/review/prompt.go now returns
(prompt, systemPrompt); both flow through ports.ReviewInvocation and the
claude-code adapter places them via LaunchConfig{Prompt, SystemPrompt}. The
re-review message reuses the per-pass prompt (role already established in the
running pane).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Vaibhaav <user@example.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
neversettle
Vaibhaav Tiwari
Khushi Diwan