Red First — Test-Proven Bug Fixing

A fix without a proving test is a guess wearing a commit message. Order is fixed: reproduce → red → green → sweep. Never fix first.

Phase 1 — Reproduce & Root-Cause

1. Read the evidence: pasted output, stack trace, screenshot description. If the evidence is a log file, run the log-analyzer skill for diagnosis and come back with its root cause.
2. Read the full delegation chain around the failure — entry point → service → model → schema. Never diagnose from the entry point alone.
3. State the root cause in one sentence: which line, which wrong assumption, which input triggers it. If you cannot state it, keep reading — do not proceed on a hunch.

Phase 2 — Red

1. Write one regression test that fails because of this bug — named by behavior (it_marks_cycle_complete_when_single_email_sent), minimal setup, placed next to the feature's existing tests, matching their style.
2. Run it. Show the red output. A test that passes on broken code is testing the wrong thing — rewrite it until it fails for the stated root cause.
3. Test at the same level the bug lives: unit test for pure logic, feature test for flow/state bugs. Real database for anything touching persistence; fake only at the established seams (HTTP, job dispatch, event listeners, AI resolver).

Phase 3 — Green

1. Minimal fix, consistent with the file's existing patterns — same naming, same error-handling style. No refactoring, no drive-by cleanups, no "while I'm here."
2. Run the regression test: green. Run the full suite: green. A fix that breaks a neighbor is not done.
3. Formatter/lint clean.

Phase 4 — Sweep for Siblings

Same bug pattern elsewhere? Symmetric paths (the parallel driver, the cache path next to the API path), copy-pasted logic, the same wrong assumption in a neighboring class. Report sibling occurrences; fix them only on request — that's a scope decision, not yours.

Output Contract

1. Root cause — one sentence, file:line.
2. Red proof — test name + the failing assertion output.
3. Fix — what changed, why minimal.
4. Green proof — regression test + full-suite results.
5. Siblings — same pattern found elsewhere, or "none".

Hard Rules

• No fix before a red test exists — if the user explicitly waives the test, note the waiver in the report.
• If the bug can't be reproduced, say so with what was tried; never ship a speculative fix for an unreproduced bug.
• The regression test stays in the suite permanently; it is the bug's tombstone.