Bug pipeline → production

A bug is anything that was broken, silently wrong, producing missing data, or producing duplicate data. Bugs are urgent and target production directly: the PR base is origin/main, not a staging branch. A human merges to main — the agent stops at PR-open.

Work the 10 steps in order. Do not skip ahead.

1. Investigate. Pull recent server/app logs for the failing path. Map the affected area with a knowledge-graph tool (god-nodes, communities, blast-radius) and index + RAG-search the relevant files so you understand the call graph before touching code. Cite what you find.
2. Root-cause, not symptom. Decide whether this is a local defect or an architectural problem. Write the one-sentence root cause. If it's architectural, say so in the plan — don't paper over it.
3. (Optional) Pull related context. If a recording, transcript, or ticket describes the repro, pull it in. Use it to confirm expected vs actual behavior.
4. PLAN — ASK THE USER. Enter plan mode. Present the root cause and the proposed fix. Do not implement until the user approves the plan.
5. Branch + code. Branch off origin/main (e.g. <your-name>-<slug>). Keep functions < 50 lines, wrap async in try/catch, validate input. Typecheck and build must be clean before you proceed.
6. FULL review gate. Run all three: your code-review bot, your diff-analysis tool, and your security-review tool against the branch diff. Paste every report inline in the PR body (not just links). Block the merge on any HIGH-severity finding, any auth / RLS / cross-tenant finding (regardless of severity), or any finding that touches user-facing data correctness (silently missing, miscounted, truncated, mislabelled, or wrong-record data). Show the user each blocking finding with file:line, propose a fix, push it to the same branch so the gate re-runs in place — never recreate the PR.
   • Skip the gate only for: docs-only diffs, pure config/CI YAML with no logic, or reverts of a commit on the same branch.
7. QA with evidence. Reproduce the original bug, apply the fix, and capture proof it's gone (logs, curl output, screenshot, or recording). Attach the evidence to the PR.
8. Open the PR (base main) then STOP. Target the default branch. Include a ## Test Plan section and the three inlined gate reports. A human merges to main — the agent does not.
9. Deploy + monitor. After merge, watch the deploy and the error stream for the affected path. Confirm the fix is live and nothing regressed.
10. Close out. Update TEAM-BOARD.md with the outcome, link the PR, and record the root cause as a finding. If the session was heavy, write a handoff for the next session.