PR delivery post-mortem routine

Use this routine after a long, costly, or repeatedly red PR iteration. The goal is to convert delivery-loop waste into a platform improvement, not to retell the whole PR.

Trigger

Run the routine when any of these are true:

• Local-CI or PR checks consumed a full expensive run that did not produce usable evidence.
• A platform state such as quiescence, sandbox drift, stale lease contention, or main freshness invalidated otherwise-good work.
• Agents had to scrape noisy truncated logs to identify the next action.
• The PR required three or more mechanical iterations after the code change was already understood.

Capture

Record the case study with these fields:

• PR number and head SHA.
• Local-CI lease id, evidence id, or pending evidence file.
• Quiescence status from get_quiescence_status, including level, runId, trigger, retryAfterSeconds, and drain blockers.
• The compact failureSummary from local-CI evidence before quoting raw logs.
• Whether the impacted-test question belongs to BI-A4EC0EA6 rather than this PR’s ad hoc analysis.
• The wasted loop: what was rerun, why it was avoidable, and the cheapest platform signal that would have prevented it.

Route Learnings

Create or update backlog/docs/commons in this order:

1. Backlog: file the smallest platform improvement with a concrete acceptance test. Link the PR and evidence ids.
2. Docs/runbooks: update the operator or agent routine that would have changed the next agent’s first action.
3. Commons: if the lesson changes standing agent doctrine, update the relevant founder-kernel principle, AGENTS.md rule, or skill instruction.
4. Tests: add a contract test for the failure mode before editing the substrate.

PR #3480 Reference Pattern

PR #3480 is the post-mortem case study for BI-C22152E7. It showed that:

• A self-upgrade drain can refuse MCP evidence writes after local-CI already passed.
• The lease cleanup path must remain available during quiescence.
• A passed gate with pending evidence needs finalize-only recovery, not a full rerun.
• Failure evidence must carry compact summaries so agents do not burn context on truncated log tails.
• Code graph impacted-test recommendations are owned by BI-A4EC0EA6; local-CI should point there instead of duplicating the recommender.