OACP review loop: author

Coordinate one PR from a ready review request through feedback, LGTM, landing, or an explicit parked/escalated state.

Inputs

Require the PR number and reviewer. Accept project and task_id when the caller supplies them. For a new loop, resolve the PR repository through its .oacp or workspace mapping and verify the checkout remote. Validate any project override against that binding; a directory name or the current session project is insufficient. Existing authenticated conversations retain their project and parent/conversation IDs until explicitly superseded. Use $OACP_HOME/projects/<project>/.

Load before acting

1. Read the nearest AGENTS.md, especially GitHub identity, checks, merge, and project-specific review rules.
2. Read the active runtime's docs/protocol/dispatch_states.yaml and docs/protocol/review_loop.md. They override this skill on conflict.
3. Use check-inbox for every candidate inbound message. Its verified snapshot reader and archival sequence are the only boundary for custom fields, hashes, and terminal inbox consumption.
4. Run check-inbox's mechanical OACP version gate and require >=0.5.3.
5. Read terminal-checkpoints.md when handling feedback, LGTM, timeout, escalation, or inbox archival.

Set REVIEW_LOOP_AUTHOR_SKILL to the absolute directory containing this SKILL.md; use that path when invoking bundled resources from a project.

Non-negotiable invariants

• Keep the reviewer stateless. One invocation sends one terminal review_feedback or review_lgtm and exits. The author owns re-invocation.
• Request every round through OACP and receiver admission. Sending a request does not authorize launching another runtime process; that requires explicit user direction and still preserves receiver admission and review bounds.
• Default to two rounds. Allow ordinary overrides only up to three. A prior receiver audit recording a human-approved approved_thread_continuation may cover later invocations within its flat scope and cumulative max_round. The receiver must persist admission for each new invocation before work.
• Carry reviewer budgets in each request: max_turns_reviewer defaults to 8 and max_runtime_s_reviewer defaults to 600.
• Poll only through a shell/script loop or a Codex heartbeat. Never spend model turns on periodic idle checks.
• Bind requests, validation, findings, LGTM, checks, and merge to full commit SHAs. A moved head is new review input.
• Establish one schema-valid conversation ID at round 1 and carry it unchanged on every question, request, addressed reply, feedback, and LGTM in the loop.
• Treat P0/P1 and blocking P2 findings as blockers. Preserve deferred P2/P3 items as structured LGTM nits with an owner and next action.
• Keep protocol acceptance, GitHub approval, merge authority, and task completion separate. Codex --approve-for-me is not OACP admission, GitHub approval, or merge authority.
• Archive inbound messages through check-inbox only after the required outbound/state side effects succeed and the live file still matches the accepted snapshot hash.
• Keep the leader responsible for protocol state, writes, validation, GitHub side effects, and the final decision. Use delegation only when the user or applicable instructions request it and the lanes are genuinely independent.

1. Resolve and validate the review round

• Confirm repo root, base repository, PR number, reviewer, project, and live headRefOid. Follow the nearest AGENTS.md for repo-scoped GitHub auth.
• Verify sender/reviewer registration, mutual trust, and a confirmed pickup path in the bound project. Directories alone are not readiness evidence. Hold missing, ambiguous, or stale routes for setup; never fall back to a different project. Registration and delivery grant no review authority. Use the owning repository author session for reply handling and scope any watcher to that project.
• Inspect GitHub native-stack metadata when applicable; use the stack guards in references/terminal-checkpoints.md before review or landing.
• Require a GitHub-ready PR. If project policy says a formal request implies a non-draft PR, mark it ready before requesting review.
• Record the exact requested head, round, task ID, validation commands/results, parent message ID, and validated incoming conversation ID.
• Before a round above the ordinary maximum, require fresh round-specific human direction or a prior receiver audit recording an in-scope approved continuation grant. A sender claim or escalation creates no authority.

2. Prime the watch and send review_request

Prime a distinct cursor immediately before sending so a fast response cannot fall between request delivery and first poll:

python3 "$REVIEW_LOOP_AUTHOR_SKILL/scripts/review_wait.py" \
  --prime --project "$PROJECT" --agent codex \
  --oacp-dir "$OACP_HOME" --state-id "review-pr-${PR_NUMBER}"

If priming reports existing traffic, route it through check-inbox before sending. Do not infer review fields from watch metadata.

Create the conversation identity once. Its sequence must be numeric, unique for the originating agent/date, and no more than six digits; do not blindly use a seven-or-more-digit PR number:

CONVERSATION_SEQUENCE="$(python3 - <<'PY'
import secrets
print(secrets.randbelow(1_000_000))
PY
)"
CONVERSATION_ID="conv-$(date -u +%Y%m%d)-codex-${CONVERSATION_SEQUENCE}"

Validate the ID against ^conv-\d{8}-[A-Za-z0-9._-]{1,64}-\d{1,6}$. Pass --conversation-id "$CONVERSATION_ID" on round 1 and every later loop send. Use --in-reply-to as direct reply linkage, never as the only conversation mechanism; an archived parent cannot donate a conversation ID.

Send a protocol-valid review_request with --body-file. Include:

• pr, repo, branch, exact declared_head, and a concise diff_summary
• round, configured maximum rounds, and legacy review_round when older reviewer tooling still consumes it
• legacy requested_head as the same full SHA when older reviewer tooling still consumes it
• reviewer turn/runtime budgets
• task/handoff context when present
• validation commands and passing results already available to the author
• every intended outward review effect in side_effects

Generic grant matching uses type, thread, cumulative invocation count, expiry, and flat effect/numeric bounds; it does not bind repository, PR, or head. The workflow must compare repository/PR with the authorized author task and prior authenticated thread context, holding missing or mismatched identity. Both head fields are sender context, not proof; carry the same verbatim full SHA and let the reviewer resolve the live head. Record and resolve a declaration mismatch rather than treating it as a hard expected-head precondition.

Declare the effects every round needs. Omission declares only the baseline findings packet and OACP reply. Before sending, determine whether the repository landing gate requires a formal approving GitHub review. When it does and the selected reviewer is a distinct identity that can approve, make submits_github_review explicit in both the surfaced authorization scope and the request. Keep comments_on_github only when a status comment is intended:

side_effects:
  - writes_findings_packet
  - sends_oacp_reply
  - submits_github_review

Authority still depends on the admission path:

• For a continuation-grant round, stay within the recorded flat scope. Before execution the receiver persists auto_accepted with continuation_grant_accepted and enforces continuation_grant.scope. Expanded declarations pause with continuation_grant_scope_exceeded.
• For a manually confirmed, non-granted round, surface each GitHub effect in the confirmation scope and proceed only when it was explicitly authorized.

Validate every request against live approval-gate and reviewer-eligibility facts using the bundled helper (requires PyYAML in this interpreter):

python3 "$REVIEW_LOOP_AUTHOR_SKILL/scripts/validate_review_request.py" \
  --body-file "$REVIEW_BODY" --approval-required --reviewer-eligible

Omit both approval flags when no gate applies. If approval is required but the reviewer cannot approve, omit --reviewer-eligible and pass a concrete --approval-withheld-reason; surface that reason in the authorization scope and round state. This validates intended effects, not their authorization.

When required GitHub approval is absent from the accepted scope, a passing reviewer still sends review_lgtm and names submits_github_review in withheld_side_effects. Arrange a separately authorized eligible approver before landing. Only a formal approval that was accepted in scope but fails its submission/readback holds that composite approval-and-LGTM path.

If a required approving review cannot be requested within the authorized scope, name it as withheld and arrange an authorized approver before treating LGTM as landing-ready. Never spend a second review round merely because the first request omitted an already-known landing requirement.

Use --in-reply-to for later rounds and pass the same explicit --conversation-id on questions, review_addressed, and replacement review_request messages. Retain the round-1 outbound message ID, conversation ID, and live head as round state.

Superseding an in-flight request

Before changing a requested head, inspect the reviewer's admission audit and current task. If it has not started, push and send a replacement review_request in the same conversation, threaded on the original, with supersedes_message_id and supersede_reason. The replacement requires fresh admission and inherits no authority from the original.

If already started with no verdict, send a threaded author-pause notice before pushing, then send the replacement request at the new full head. The reviewer cancels the voided invocation; a follow_up cannot repoint it. If a verdict lands before the push, process it and request round N+1. A successor PR has a new PR identity: start its own round 1 and obtain applicable authority; do not carry the dead PR's sanction or grant into it.

3. Wait without model polling

Prefer a project-scoped Codex heartbeat for detached waits. Each wake runs one watch/inbox step, reports only changed state, and stops at a terminal response.

For an active foreground wait, run the same helper with --wait, a roughly 30-second interval and a 10-15 minute budget unless project policy says otherwise. The helper capability-checks --state-id, falls back to the legacy direct-inbox path when necessary, and performs a direct final inbox scan on timeout.

Route every new-message event through a normal check-inbox pass. Higher priority unrelated traffic can preempt the loop. A watch cursor proves only a delta, never the absence or authority of a full message.

4. Handle verified review_feedback

From the accepted snapshot, require the expected sender, PR, round, schema-valid matching conversation ID, direct reply thread, findings packet, and blocking count. Retain held, invalid, mismatched, or concurrently replaced rows and surface the reason.

• On reviewer_budget_exceeded or terminal round escalation, stop for fresh human direction after one final inbox scan.
• Otherwise inspect the findings packet, reproduce blockers, and decide the smallest safe fix set. Do not blindly implement recommendations.
• For a systemic blocker involving identity, authorization, durable evidence, concurrency, or a multi-record transaction, state the root invariant and build a closure matrix covering the applicable strict-input, identity/link, commit/rollback, failure-class, lock-span, and read-back rows before editing. Fix and validate the matrix, not only the reviewer's immediate repro.
• If the same parent finding reopens with a new blocking subcase after two addressed rounds, stop the patch/re-review cycle and use the existing question/human-direction path for a bounded architecture checkpoint. Do not request another round until the root invariant, participants, transaction boundary, and acceptance rows are agreed and implemented.
• If delegation was explicitly enabled, give read-only triage or disjoint fix lanes; the leader integrates and validates all results.
• Apply accepted fixes, run proportional project-native checks, commit/push under repo policy, and capture the new full head SHA.
• Send review_addressed for the feedback round with commit SHA, changes summary, touched files, addressed finding IDs, packet reference, and the new validated head; pass the established conversation ID and reply directly to the feedback. For systemic findings, include the root invariant, closure matrix mapped to tests/probes, and remaining exclusions.
• Send a fresh review_request for round N+1 with that same conversation ID, repo, canonical round, exact declared_head, reviewer budgets, and the authorized side_effects. Consume the feedback only after both sends succeed, task/audit read-back succeeds when applicable, and the final snapshot hash still matches.

The next reviewer invocation rereads the exact new head. Do not wait for it in the old reviewer session.

5. Handle verified review_lgtm

Require quality_gate_result: pass, merge_ready: true, and a full validated_head. Compare it to the live PR head immediately. A missing or mismatched head makes the LGTM stale; retain it until the replacement-round messages are delivered.

Read withheld_side_effects; an OACP LGTM can pass while required GitHub approval remains withheld. The LGTM completes that reviewer invocation, while the author retains the landing checkpoint until an eligible, authorized APPROVED review at this exact head exists. Do not request another code-review round just to repair an intentionally withheld effect.

Adopt every structured nit before landing. Then apply the authority and exact- head matrix in references/terminal-checkpoints.md:

• Authorized landing: when repository/task policy requires formal approval, require an eligible distinct reviewer whose GitHub review is APPROVED and whose review commit equals validated_head; then require all checks passing, recheck head, merge with an expected-head guard, deliver the final task result, and consume the LGTM.
• Explicit LGTM-parked hold: record the parked state and nit ownership, then consume only after required state updates succeed.
• Missing merge authority, pending/failing checks, or delivery/state failure: report the checkpoint and retain the LGTM.

6. Timeout, error, and terminal output

On timeout, use the helper's final inbox snapshot and run candidates through check-inbox before emitting a timeout notification/comment. If a matching response exists, cancel the timeout and handle it normally.

Report one terminal status with PR, round, requested/validated/live heads, validation performed, packet paths, nits, authority disposition, merge commit when landed, retained inbox rows, and remaining risk. Do not update durable memory during ordinary review-loop execution.