When the master starts (first deploy or after a long outage), it
needs to learn about existing open PRs/issues that weren't created
via discovery-tick-during-uptime. Backfill = discovery + a one-time
marker so subsequent restarts know "this isn't the first time."
tools/controller/master/backfill.py:
- run_startup_backfill(engine, owner, repo, list_prs, list_issues):
- Calls run_discovery (already idempotent — skips existing entities)
- Records a 'controller-backfill-complete' marker in controller_events
associated with the first new workflow OR an existing workflow OR
skipped if Forgejo is truly empty (no FK target)
- Returns BackfillReport{first_time, discovery}.
- has_backfill_run(engine, owner, repo): existence-check on the marker
by parsing controller_events.payload. Multi-tenant isolated — a
marker for (owner_a, repo_a) doesn't satisfy a check for
(owner_b, repo_b).
- The marker is informational; the dedup is provided by discovery's
unique-constraint skip. The marker exists so operators can answer
"has backfill ever run for this repo?" in one SQL query.
Wired into master __main__:
- Runs AFTER engine/create_all + Forgejo callback wiring, BEFORE
master_main_loop.
- Try/except wrapped so Forgejo flake at startup doesn't prevent
the main loop from running (discovery tick will retry).
- New --skip-backfill flag for tests + warm restarts.
8 new tests in test_master_backfill.py:
- has_backfill_run: no marker → False; multi-tenant isolation
(different owner OR different repo → False).
- First-time backfill creates workflows + marker; empty Forgejo
skips marker (no FK target).
- Second run reports first_time=False; picks up newly-appeared PRs
+ emits a second marker.
- Multi-tenant (owner-a, repo-a) and (owner-b, repo-b) both get
their own marker.
- New workflows are in DISCOVERED state.
Total: 415 controller tests; full auto_agents suite 2777 pass.
drew
ba2e9472bc