PR Watch

Overview

Poll GitHub PR review status. Auto-fix review comments (Critical/Major/Minor by default) using pr-review-dev. Exit when PR reaches a terminal state: approved, changes_requested_fixed (all blocking issues resolved), timeout, or capped (max fix cycles reached).

Announce at start: "I'm using the pr-watch skill to monitor reviews for PR #{pr_number}."

Implements: OMN-2524

Quick Start

/pr-watch 123 org/repo
/pr-watch 123 org/repo --timeout-hours 48
/pr-watch 123 org/repo --max-review-cycles 5
/pr-watch 123 org/repo --fix-nits
/pr-watch 123 org/repo --no-auto-fix

Poll Loop

Fetch PR review state via gh pr view {pr_number} --repo {repo} --json reviews,reviewDecision
If reviewDecision == APPROVED: exit with status: approved
If reviewDecision == CHANGES_REQUESTED and auto_fix=true and cycles remaining:
- Dispatch pr-review-dev agent with review comments
- Increment fix cycle count
- Wait for CI (delegate to ci-watch if needed), then re-request review
If fix cycles exhausted: exit with status: capped
If elapsed > timeout_hours: exit with status: timeout
If all blocking issues resolved but no explicit approval: exit with status: changes_requested_fixed

Fix Dispatch Contract

Task(
  subagent_type="onex:polymorphic-agent",
  description="pr-watch: fix review comments for PR #{pr_number} (cycle {N})",
  prompt="Invoke: Skill(skill=\"pr-review-dev\", args=\"{pr_number}\")

    Review comments:
    {review_comments}

    Fix all Critical, Major, and Minor issues.{nit_instruction}
    Push fixes to branch: {branch_name}

    Report: issues fixed, files changed, any issues skipped."
)

Skill Result Output

Write ModelSkillResult to ~/.claude/skill-results/{context_id}/pr-watch.json on exit.

{
  "skill": "pr-watch",
  "status": "approved",
  "pr_number": 123,
  "repo": "org/repo",
  "fix_cycles_used": 2,
  "elapsed_hours": 3.5,
  "context_id": "{context_id}"
}

Status values: approved | changes_requested_fixed | timeout | capped | error

Push-Based Notification Support (OMN-2826)

pr-watch supports two notification modes, selected automatically based on infrastructure availability. The push-based mode eliminates polling for CI status checks during the review cycle.

EVENT_BUS+ Mode (preferred)

When Kafka and Valkey are available (ENABLE_REAL_TIME_EVENTS=true):

Register watch: agent registers interest in (repo, pr_number) via Valkey watch registry
Wait for inbox: block until a pr-status event arrives in the agent's inbox topic (onex.evt.omniclaude.agent-inbox.{agent_id}.v1)
Process result: on CI completion after a fix cycle, check review status and decide next step

from omniclaude.services.inbox_wait import register_watch, wait_for_pr_status

# Register watch for CI re-runs after pushing fixes
await register_watch(agent_id=agent_id, repo=repo, pr_number=pr_number)

# After pushing a fix, wait for CI to complete (replaces polling)
ci_result = wait_for_pr_status(
    repo=repo,
    pr_number=pr_number,
    run_id=latest_run_id,
    agent_id=agent_id,
    timeout_seconds=3600,
)

STANDALONE Mode (fallback)

When Kafka/Valkey are unavailable:

Spawn gh run watch {run_id} --exit-status as background process after pushing fixes
Wait for result in file-based inbox (~/.claude/pr-inbox/)

from omniclaude.services.inbox_wait import wait_for_pr_status

# Unified interface -- automatically falls back to STANDALONE
ci_result = wait_for_pr_status(
    repo=repo,
    pr_number=pr_number,
    run_id=latest_run_id,
    timeout_seconds=3600,
)

Migration Notes

The original review polling loop is unchanged. The inbox-wait pattern replaces only the CI status checks that happen between fix cycles (when pr-watch pushes a fix and needs to wait for CI before re-requesting review). The wait_for_pr_status() function provides a unified interface that works in both EVENT_BUS and STANDALONE modes.