Aller au contenu principal

Acceptance Criteria: Instant Data-Repo → Main-Repo Sync

Feature ID: data-repo-instant-sync Spec: ./spec.md Status: Draft Last updated: 2026-05-16

Core behaviour​

AC-1: Webhook-driven sync (App installed)​

  • Editing categories.yml on the data repo and committing to default branch triggers a sync within 30 seconds (debounce window).
  • The main repo's README.md reflects the new category after the sync completes.
  • An activity row of type data-sync.success appears with source: webhook, correct beforeSha / afterSha, and filesChanged >= 1.
  • Five commits to the data repo within 10 seconds coalesce into a single sync run (one .success row, afterSha = HEAD of last commit).

AC-2: Poller-driven sync (App not installed)​

  • On a Work with githubAppInstalled = false and syncIntervalMinutes = 5, editing a markdown file in the data repo via a manual git push results in a sync within ≤ 5 minutes.
  • Activity row data-sync.success records source: poll.
  • If no commits are made for 24h, only 24 data-sync.skipped reason=no-changes rows appear (rate-limited to 1 per hour), not 288.

AC-3: Mutex with generation pipeline​

  • When a generation pipeline run is RUNNING, a concurrent sync attempt emits data-sync.skipped reason=generation-in-progress and does NOT call MarkdownGeneratorService.syncFromDataRepo().
  • After the generation pipeline completes, the next webhook delivery (Path A) or next poll tick (Path B) detects the SHA delta and successfully runs the deferred sync.
  • When two sync runs are triggered within the same second, only one acquires the lock; the other emits .skipped reason=sync-in-progress.

AC-4: Lock TTL recovery​

  • If dataRepoSyncTask crashes with the lock held (simulated via kill -9), a new sync attempt 5 minutes later acquires the lock cleanly (no stuck Work).
  • The dispatcher (WorkScheduleDispatcherService.dispatchDue) skips the locked Work and resumes scheduling for it once the lock expires.

Render correctness​

AC-5: Render parity with full pipeline​

  • For a Work with no item changes (only category/tag changes), the resulting README.md produced by syncFromDataRepo() is byte-identical to the README produced by a full WorkScheduleDispatcherTask run.
  • details/<slug>.md files are written exactly as they would be by the full pipeline.
  • No items/ regeneration occurs — verified by absence of AI provider calls in test logs.

AC-6: Idempotency​

  • Running syncFromDataRepo twice in a row against the same data-repo SHA produces filesChanged: 0 on the second run.
  • The second run still emits an activity row (.success with filesChanged: 0) so users can see the sync attempt happened.

Failure modes​

AC-7: Data repo unreachable​

  • If the data repo returns 404 / 403 from GitHub, the sync emits data-sync.failed with errorClass: 'data-repo-unreachable' and errorTail containing the last 200 chars of stderr.
  • The lock is released; subsequent runs are not blocked.
  • The Work's lastSyncedDataRepoSha is not updated.

AC-8: Main repo push rejected​

  • If the main repo push is rejected (branch protection / force-push refused / merge conflict from a parallel manual edit), the sync emits data-sync.failed with errorClass: 'main-repo-push-rejected'.
  • On the next tick the sync retries; if the underlying conflict resolves naturally, the retry succeeds and emits .success.

AC-9: Webhook signature invalid​

  • A push payload with an invalid x-hub-signature-256 returns 401 from the webhook controller, does not emit any activity row, and does not touch the lock or queue.

Force-sync endpoint​

AC-10: POST /api/works/:id/sync​

  • Returns 202 with { activityRowId, status: 'enqueued' } for a healthy Work.
  • Returns 202 with { status: 'skipped', reason: 'generation-in-progress' } when the pipeline is RUNNING (no error, by design).
  • Returns 403 for an unauthenticated caller; 404 for an unknown Work id.
  • Auth check rejects users without write access to the Work.

Activity feed UI​

AC-11: Sync chip + rows​

  • The Sync filter chip appears on the activity page next to Generate.
  • Clicking the chip filters to only data-sync.* rows.
  • data-sync.success rows show short SHAs (7-char prefix), files-changed count, and a clock duration.
  • data-sync.failed rows render with an alert variant and an expandable disclosure showing errorTail.
  • data-sync.skipped rows render with a muted variant and the reason verbatim.

Telemetry​

AC-12: Counters and histogram​

  • On a successful webhook-triggered sync, data_sync_success_total{source="webhook"} increments by 1 and data_sync_duration_ms records a sample.
  • A skipped sync increments data_sync_skipped_total{reason,source} with the correct labels.
  • A failed sync increments data_sync_failed_total{errorClass,source}.
  • PostHog dashboard Data Sync shows per-day totals, p50/p95 duration, and top error classes.

Feature-flag gating​

AC-13: Both flags off (default at first deploy)​

  • With subscriptions.dataSync.webhookEnabled = false: push events are accepted (200 OK) but do not enqueue. No activity rows.
  • With subscriptions.dataSync.pollerEnabled = false: the poller task either does not run or short-circuits and emits no rows.
  • Toggling either flag to true enables the corresponding path without a redeploy (env var read on each task run).

Rollout / regression​

AC-14: No regression to the scheduled pipeline​

  • The existing scheduled generation pipeline (WorkScheduleDispatcherTask) continues to run end-to-end on its existing per-Work cadence.
  • Items, comparisons, taxonomy, and markdown rendering are byte-identical to the pre-feature output on a control Work that does not receive any data-repo edits during the test window.

AC-15: Migration safety​

  • Migration up + down + up cycle on a snapshot of staging data completes in < 60 seconds with no row-level errors.
  • Backfill of github_app_installed = true covers every Work that has a non-null github_app_installation_id and only those Works.

AC-16: Telemetry sampling sanity​

  • Over a 24-hour soak on develop with 50+ active Works, total counter values are within ±2% of the activity-row counts (no drops, no double-counts).