Skip to main content
On this page

P9 · Fully automated review mode (adk auto) — progress note (2026-05-21)

Status: SHIPPED as MVP (locally). Core orchestrator + dry-run + report; safety guards deferred.

Commits

Text
<latest> v4/P9: adk auto — headless review orchestrator (MVP)

Test pass count

364 passed (+8 new in test_auto_run.py).

Exit-criteria check (plan §8 P9)

Criterion Evidence Status
adk auto --dry-run prints what it would do auto_run.main([..., "--dry-run"]) → JSON summary listing would_review URLs. ✓
adk auto runs end-to-end with no input Implemented. Spawns <agent> -p "/adk-pr-review <url>" per eligible PR; captures logs; writes report.md. ✓
Exit-code 0/1/2 semantics 0 = all succeeded, 1 = some failed, 2 = aborted by guard (today: invalid args). ✓

What got done

  • skills/adk-cli/scripts/auto_run.py — orchestrator.
  • bin/adk — auto wired into dispatch table + USAGE block.
  • _eligible_rows(queue_path, exclude=...) — uses ready_for_review (P5’s §6.u predicate) + FIFO sort.
  • _spawn_review(pr_url, agent, run_dir, log) — captures per-PR stdout to a log file; returns structured summary; handles missing-binary gracefully.
  • _write_report(run_dir, results, ...) — writes report.md with per-PR bullets + run metadata.
  • 8 tests covering eligibility filtering + dry-run + max-reviews cap + report generation.

What got skipped (deferred to a follow-up)

  1. --quiet-hours WW-XX — refuse to spawn during the window. Simple to add (parse args.quiet_hours, check current hour against the local clock, exit 2 if inside).
  2. --max-cost-usd N — pre-flight estimate × ready-PR-count. Needs per-agent cost table (could hard-code claude=$X in v1).
  3. --report-to-slack #channel — one summary post to that channel at the end. Reuses slack_helpers (P6 deferred Slack work). Once §6.y lands, the post fits naturally.
  4. Per-agent registry / picker — today’s --agent accepts any binary name. The plan §8.5 talks about an agent registry (registry.launch(slash_cmd, args)) so the TUI + auto can swap agents (codex / cursor / opencode). Deferred to P8 (TUI) since that’s where the abstraction pays off.
  5. launchd plist example in SETUP.md — a 10-line example to fire adk auto every 2h.

How to use

Shell
# Preview what would happen (no agent spawned):adk auto --dry-run# Run end-to-end (default: 1 agent at a time, up to 20 PRs, default agent claude):adk auto# Custom parallelism + cap:adk auto --parallel 4 --max-reviews 10# Exclude a specific PR:adk auto --exclude https://github.com/acme/foo/pull/123# Skip the pr-sync pre-flight (e.g. if it just ran):adk auto --no-sync# Use a different agent:adk auto --agent codexadk auto --agent cursor

Decisions made this run

  • Default --parallel 1 — serial execution by default. Parallelism is opt-in to give predictable runtime for cron schedules.
  • Default --max-reviews 20 — matches plan §7.4. Conservative cap so a runaway sync doesn’t burn budget.
  • --dry-run is side-effect-free — even pr-sync is skipped under —dry-run.
  • Missing-binary handling — _spawn_review catches FileNotFoundError and emits a structured failure dict instead of crashing the run. Other PRs in the same run still get a chance.
  • Log location — ~/.agents-devkit/skill-setup/auto-runs/<ts>/ per plan §7.3. Each PR gets its own .log file; the run-level summary lands in report.md.

Open questions for the user

  1. Worth ironing out --quiet-hours + --max-cost-usd + --report-to-slack in a focused session — they’re the user-facing knobs that make adk auto schedulable safely.
  2. The agent-registry abstraction (so --agent codex works equally to --agent claude) needs a sketch — today both work via the same spawn pattern, but a clean registry pattern (per-agent argv builder + per-agent cost coefficient) would tighten the contract.

Exact commands to inspect

Shell
git show HEAD --statpython3 -m pytest skills/adk-cli/scripts/tests/test_auto_run.py -vadk auto --dry-run