Debugging

Read skills/blocks/engineering-method-selection.md from the installed plugin for review admission. Natural-language requests and --peer-review share that policy. Honor host-only requests; risk alone does not authorize paid usage.

Run the investigation on the current host. Routine debugging makes zero additional provider dispatches. Use a bounded external reviewer only for --peer-review, an explicit independent-review request, or an existing risk policy.

<HARD-GATE> DO NOT CHANGE PRODUCTION BEHAVIOR BEFORE REPRODUCING THE SYMPTOM AND TESTING A NAMED ROOT-CAUSE HYPOTHESIS. </HARD-GATE>

Read and apply skills/blocks/debug-feedback-loop.md from the installed plugin root.

Start with the user's observable symptom. Reproduce it, retain its failure signature while minimizing the scenario, test one named hypothesis at a time, and verify both the minimal reproduction and the original scenario after the fix. Do not treat a nearby passing helper test as proof.

For a race, use a synchronization barrier and a fixed run or time budget. For an unavailable production dependency, return inconclusive with the missing evidence. Remove temporary instrumentation before completion and preserve a stable reproduction as a regression test.

The final record is data, not an executable queue. Store commands as argument arrays and never evaluate provider-authored text.

Bounded recovery and strategy rotation

Use a 3-Strike Rule for failed fixes. After each failure, return to the evidence and test a materially different hypothesis. After two consecutive failures, a strategy rotation is mandatory: reconsider the root cause, the reproduction, and whether the test encodes the intended behavior. Do not attempt a 4th fix without explicit user approval.

Anti-rationalization check: “Should work now” means run the reproduction and the original scenario. Confidence is not verification.