Diagnose → Review → Fix

An orchestration protocol that chains three skills into a single pipeline: find the root cause, write the fix, review it, and harden it. The human stays in the loop at every stage transition.

When to use

• YES: Uncertain bugs where root cause is unknown, multi-layer issues, high-risk changes needing diagnosis + review
• NO: Known bugs (skip to fix-loop), code review of existing changes (use two-pass-review directly), trivial fixes
• If you already know the root cause, skip Stage 1 and start at Stage 2

Pipeline

Four stages. Every stage transition requires human approval via AskUserQuestion with structured options and a recommended choice. Never auto-proceed.

Stage 1: Diagnose

Load the parallel-diagnosis skill. Pass it the problem statement and relevant file paths.

Receive: DiagnosisOutput with status, root_cause, confidence, agreement, affected_files, fix_direction.

Checkpoint: Present diagnosis to the human. Check status first, then evaluate confidence and agreement.

• If status is "not_a_bug" → report findings and stop. No fix needed.
• If status is "inconclusive" → stop or gather more context before proceeding.
• If status is "diagnosed": evaluate confidence and agreement to decide next steps.
  • confidence: "low" or human disagrees → stop or re-run with more context.
  • confidence: "medium" + agreement: "escalated" warrants more caution than "medium" + "converged" — surface the disagreement context to the human.
• Only proceed to Stage 2 on human approval.

Stage 2: Fix

No sub-skill — the parent LLM writes the fix directly.

Using the diagnosis output (root_cause, affected_files, fix_direction), read the affected files and make the code changes. This is a normal coding task.

Checkpoint: Show the changes to the human. Proceed on approval.

Stage 3: Review

Load the two-pass-review skill. Pass it:

• Artifact: the files changed in Stage 2
• Criteria: "Does this fix address the diagnosed root cause? Are there regressions, edge cases, or incomplete handling?"
• Scope: the files changed in Stage 2 and their immediate callers/dependents

Receive: ReviewOutput with findings (P0–P3, confirmed/demoted/rejected).

Checkpoint: Present review findings to the human.

• No confirmed P0/P1 findings → recommend shipping. Skip Stage 4.
• Confirmed P0/P1 findings → recommend proceeding to Stage 4.

Stage 4: Harden

Load the fix-loop skill. Pass it:

• Findings: the confirmed P0 and P1 findings from Stage 3
• Artifact paths: the files changed in Stage 2
• Criteria: derived from the original diagnosis (root_cause + fix_direction) and the review criteria from Stage 3

fix-loop will attempt fixes (max 2 per finding) and verify each one.

Receive: fix-loop output with per-finding results: resolved in Round 1, resolved in Round 2, or escalated with attempt history.

Checkpoint: Present final state to the human.

Shortcutting

Before applying any shortcut, confirm with the user via the AskUserQuestion tool. Present the detected condition and the proposed skip as a structured option alongside "Run all stages".

Condition	Action
Root cause already known	Skip Stage 1, start at Stage 2
Fix is trivial after diagnosis	Skip Stages 3–4 — apply directly after human confirms the fix is trivial
Review finds no P0/P1 issues	Skip Stage 4, recommend shipping
Human says stop at any checkpoint	Present what you have and halt

Constraints

• Every stage transition requires human approval. Never silently proceed.
• Never auto-proceed from diagnosis to fix without the human seeing the diagnosis.
• Never auto-proceed from fix to review without the human seeing the changes.
• If any stage fails or the human says stop, present what you have and halt gracefully.
• This skill does not define its own output schema — it uses the schemas of the skills it chains (DiagnosisOutput, ReviewOutput, Finding schema v1).

Output

Present a final summary to the human:

1. Root cause — from Stage 1 (or user-provided if Stage 1 was skipped)
2. Changes made — from Stage 2
3. Review findings and resolution — from Stages 3–4
4. Deferred items — any P2/P3 findings from review (FYI only)