Debug

Start with what should happen, what actually happened, and the best evidence already available. Use Adapt if relevant facts are missing. Bring in other skills only for a distinct question, such as Design for unclear interaction intent or Test when a reliable check needs specialist work.

Open the matching Debug tool guide before interpreting tool-specific behavior or choosing a tool-specific diagnostic, even when working from saved traces. Skip guides when the evidence already establishes the cause and no tool-specific check is needed.

Find the cause

• Fix directly when the evidence is clear. If a focused check confirms the failure and the code directly explains it, make the smallest authorized fix and rerun the check. Skip a separate orientation or investigation pass.
• Trace an unclear cause. Use the smallest safe reproduction. Compare a working run with a failing one and trace the relevant path to the first point where their behavior diverges. For each plausible cause, choose a check that can rule it out. After two plausible hypotheses fail, pause the investigation. Report what you ruled out, where the evidence stops, and the next check you recommend.
• Keep intermittent reports separate. If the reported failure will not reproduce, record the conditions and what remains unknown. Fix an independently proven, in-scope defect if you find one. Leave the reported failure unverified until evidence connects it to that cause.

Fix and verify

• A clearly reported bug authorizes an in-scope fix. Stop at diagnosis when requested, and ask before changing agreed behavior. An orientation finding alone does not authorize a repair.
• Reuse a focused check that catches the failure. If a regression case needs coverage, extend an existing test when it fits naturally. Add a separate test if it needs different setup or would make the existing test unclear. Run the required project checks and verify affected behavior.
• Report the cause, fix, and actual results. If the original check cannot run, deliver the in-scope fix with the evidence you have, name the exact check still needed, and leave that behavior unverified. Honor explicit completion gates. Use Preflight for setup blockers and its rules for weaker fallbacks.

During an incident, Deploy’s operations branch owns mitigation and recovery. Give it causal evidence and any authorized fix. Use Audit for a separate security or privacy review and Migrate for persistent state transitions.