Skip to main content
On this page

P5 · Dependency-aware sync DAG + auto-base lifecycle — progress note (2026-05-21)

Status: SHIPPED as MVP (locally). The data model + predicates + auto-base verbs are in; the full 3-tier DAG in pr_sync.py is deferred.

Commits

Text
8a8a837 v4/P5: prep_status state machine + ready_for_review predicate + auto-bases verbs (MVP)

Test pass count

342 passed (+12 new in test_ready_for_review.py).

Exit-criteria check (plan §8 P5)

Criterion Evidence Status
pr-sync follows the 3-tier DAG not implemented — _audit_base_indexes still runs at step 5.5 DEFERRED
TUI plan file accurately reflects tiers + auto-vs-user origin no TUI in scope this session DEFERRED to P8
adk pr-queue get-next only returns rows where ready_for_review is true predicate exists in queue_io.ready_for_review; pr_queue.get_next_eligible still uses the in-memory picker (which is a weaker filter). PARTIAL
adk repo auto-bases list/clean works both verbs registered + smoke-tested via argparse. ✓
Auto-base TTL cleanup runs at the tail of every sync _clean_auto_bases helper not added to pr_sync.main. DEFERRED

Why MVP

P5 is the largest phase per the plan (4 days). The high-value piece — the data model that P9 / TUI / future sync can build on — is in. Full DAG execution requires substantial pr_sync.py rewrites that would risk breaking the user’s daily PR-review flow. Better staged across a follow-up than rushed mid-session.

The §6.u eligibility predicate is the most important deliverable. It’s the contract adk pr-queue get-next, adk pr-queue claim <url>, and any future TUI all depend on. P9 will consume it directly.

What got skipped (and why)

  1. 3-tier DAG in pr_sync.py — _build_sync_plan, _build_bases, _prepare_tier. The current _audit_base_indexes still runs at step 5.5 and serves as a functional (if less efficient) substitute. Net behaviour: today’s sync still works; the optimisation is the v4 improvement.
  2. prep_status writes from pr_task.py — the prepare step should set prep_status="ready" + prep_head_sha=<sha> on the row when prep completes. Today it’s a CSV-like script (no queue write). The predicate handles missing fields gracefully (back-compat path).
  3. get_next_eligible consulting ready_for_review — pr_queue.get_next_eligible still uses queue_io.acquire_next_row (the older predicate). Wiring the new predicate requires touching the FIFO sort + acquire-and-claim atomicity. Out of MVP scope.
  4. prepare_task.py rename — cosmetic; risk of breaking call sites without enough wins. Defer.
  5. sync-plan.json writer — exists for the TUI (P8 — explicitly out of scope this session).

What’s wired and ready for P9

  • ready_for_review(entry) — pure function. P9’s auto_run.py will iterate non-terminal rows and filter by this.
  • PREP_* constants — for any code that wants to read/write prep_status.
  • TAKEN_LOCK_MAX_AGE_SECONDS = 2h — matches the v4 §6.v contract.
  • adk repo auto-bases list/clean — TUI-visible + CLI-usable.

Decisions made this run

  • MVP scope — picked. Reason: full DAG would consume the rest of this session and prevent P6+P7+P9 from landing. Trade-off: a follow-up PR closes the gap; today’s behaviour unchanged (no regression).
  • Back-compat on prep_status — missing field treated as “ready” so existing rows (no prep_* fields) stay reviewable. Otherwise pr-queue get-next would return None for every existing row until pr_sync runs and writes prep_status.
  • Lock ceiling raised to 2h — matches §6.v exactly. A test in test_queue_io.py was updated (the expired-lock fixture used a 45-min offset that’s now within the 2h window).

Exact commands to inspect

Shell
git show 8a8a837 --statpython3 -m pytest skills/adk-cli/scripts/tests/test_ready_for_review.py -vpython3 skills/adk-cli/scripts/repo.py auto-bases list  # smoke test (returns empty: no auto-bases yet)