Attention Engine

Attention Engine

Package: internal/attention. Severities are only:

• critical
• warning
• waiting

Legacy fingerprints (e.g. blanket open_pull_request) are resolved on evaluate. Periodic sweep runs about every 10 minutes. Long-running threshold defaults to 2 hours (attention.long_running_after). Bootstrap admins can override per-rule severity under Settings → Preferences (Attention Severity Overrides); overrides live in attention_rule_overrides.

Rule matrix

Type key	Typical severity	Trigger (summary)
failed_default_branch_workflow	critical	Failed/timed-out run on repo default branch
deployment_workflow_failure	critical	Failed/timed-out run whose name/path/event looks like deploy/release/prod/cd
pr_ci_failure	critical	Open PR with CI state failure
required_check_failed	critical	Open PR mergeable_state=unstable with CI failure or review pressure (best-effort)
approved_blocked_by_ci	warning	Approved-ish PR blocked by failing CI
awaiting_manual	warning	Run/job waiting or action_required
long_running_workflow	warning	Incomplete run older than threshold
awaiting_review	waiting	Open PR waiting for required review
approved_behind_target	warning	Approved PR behind target branch
merge_conflict	warning	Open PR with merge conflict
runner_unavailable_queued	warning (stub)	No-op — forge runners API may list runners, but job payloads do not expose a reliable “queued because runner offline” signal (see Status capability matrix)

Mutes / snooze

Table attention_mutes:

• Per-item mute by fingerprint from Attention UI: Mute → Snooze 24h / Snooze 7d / Mute Until Resolved
• Bootstrap-admin mutes are global (user_id NULL): evaluate skips opening matching fingerprints; current item is resolved immediately
• Non-admin mutes are personal: list hides the item for that user; evaluate still opens for everyone else
• until_at NULL = until resolved — cleared when the engine resolves that fingerprint
• Timed mutes expire; next evaluate reopens if the condition remains true

APIs: POST/DELETE /api/v1/attention/{id}/mute, GET/PUT /api/v1/attention/rule-overrides (PUT bootstrap-admin + CSRF).

Log Tail (on-demand)

GET /api/v1/attention/{id}/log-snippet fetches a capped log tail from the forge for a failed job linked to the item (entity_type job/run or job_id/run_id in metadata). Default ~4 KiB, max 16 KiB; not stored in GitSeer by default. UI: Log Tail on Attention cards when a job/run is resolvable.

Evaluation entry points

• EvaluateRun — default-branch failure, deploy failure, awaiting manual, long-running
• EvaluatePullRequest — CI failure, required check, approved-blocked, awaiting review, behind target, merge conflict
• EvaluateJob — awaiting manual (job), runner unavailable (stub)

Items are upserted by fingerprint and resolved when the condition clears. Global mutes short-circuit open().

Personal inbox

GET /api/v1/inbox surfaces ACL-scoped attention and open PRs for the current user with reason tags: author, requested_reviewer (when attention metadata_json includes requested_reviewers), failing_ci, blocked_on_me (changes requested / merge conflict / approved-blocked on authored PRs, or review request metadata). UI: /inbox with saved filter presets (saved_filters).