Recovery Playbook

Eval below threshold

HITL prompt fires automatically. Reject + add reason -> run terminates, recovery row on Monday.

Buffer publish failure

Automatic retry 3x (exponential). If still failing, alert in #recovery. Manual options:

Reschedule for later
Post a correction
Mark as cancelled

Orphan run

runner.py detects stale state.json lock (>1h old) on next tick -> logs orphan, alerts.

Image generation failure

Run marked failed with image_unavailable. Re-run from Step 03 manually.

Wrong-country copy

Triggers country_copy_check failure -> HITL -> reject -> re-draft.

TODO: add SOP for each failure mode after W6 simulation pass.

Grants Pipeline Failure Modes

Budget blocking gate fires -- salary above 40 pct

The 40 pct salary check is a hard gate. When fired:

Run is blocked and cannot reach pre-submission
Telegram alert in hitl-approvals with restructuring guidance
Options: restructure budget and re-run, or waive with explicit override

Partner unresponsive -- more than 14 days

Automatically flagged during scan. Partner status downgraded to unresponsive. Alert in agent-ops. Consider replacing partner or dropping consortium component.

Capacity guard -- 2 or more active drafts

New evaluating opportunities are deferred until an active draft completes. Logged as capacity_deferred in audit. No intervention needed; next tick retries.

HITL gate timeout

If no response within timeout: run terminates cleanly. Opportunity stays at current stage. Next daily tick retries.

Eval score below threshold

Routes through pre-submission HITL with the score and failing checks visible. Reviewer can approve despite low score.

Subagent verify unavailable

Graceful degradation: check returns passed=True with weight=0 so it does not count. Other checks still score normally.