Background jobs

These commands describe current main (3.0.0-rc.1). Use the Web UI for the equivalent inspection workflow.

Curation returns after its mutations, timeline events and enrichment intents commit together. It does not wait for embedding or consolidation. The service recovers pending work from PostgreSQL after restart; polling is authoritative.

codifix jobs --project example-project
codifix jobs --project example-project list --status blocked --limit 20
codifix jobs --project example-project inspect JOB_UUID
codifix jobs --project example-project --json inspect JOB_UUID
codifix jobs --project example-project retry JOB_UUID --request-id REQUEST_UUID --yes

Use the returned next_after UUID with list --after UUID for another bounded page. Inspect shows dependencies, sibling stages, attempts, next eligibility, sanitized failure codes and explicit skip reasons. Queue counts and oldest-ready age are project-scoped; the worker timestamp reports the serving process's poll. Raw prompts, source contents and provider responses are never included.

For timeline recovery, GET /v1/projects/{slug}/activities?after_seq=0&limit=100 returns ascending commit order and cursor.next_after, has_more, retention_gap and deleted_through. Save the cursor only after processing the page. A retention gap requires snapshot recovery, not silent cursor advancement. Cursor mode accepts 1–200 items and cannot be combined with time/kind filters; omitting after_seq preserves the legacy newest-first activity view.

Retry only blocked or failed jobs after addressing their cause. It needs LoopsRun, the operation's execution permissions, and the original execution identity's current authority. Read-only services reject retries. Completed stages and original inputs are retained; a new bounded retry generation is opened. Running and successful jobs cannot be restarted here. Retry cannot transfer authority from a revoked requester to an operator or bypass configuration policy.

Pre-cutover enrichment has no reconstructed historical jobs. Database operators can explicitly request fresh, version-pinned repair work using the R1 rollout and repair procedure.

Choose a request UUID once. If a reply is lost, repeat the same request UUID. Provider execution is not exactly once: an uncertain attempt can be billed again. Credential, quota, authorization and configuration failures need intervention; transient failures retry automatically until the generation's limit is reached.

Web and TUI

Web: More → jobs, filter or refresh, inspect a stage, Review retry, then confirm. Switching projects clears the selected details and pending confirmation. The same request ID is retained after an uncertain response.

TUI: open Jobs, press r to load, s to cycle filters, j/k to select, Enter to inspect and PgDn for the next page. R prepares a failed/blocked stage's retry; y confirms and n or Escape cancels. d/PgUp scroll details. Inspection replaces the list; Escape returns to it. The confirmation names the project, job and stable request UUID. Refresh to see subsequent status changes. Read-only services disable retry.