Debug and repair failed Agent QA test runs from evidence

Repairs a failed Agent QA run using the run's recorded evidence and the relevant local source, treating the automated failure classifier as a hypothesis rather than a verdict. It collects evidence through agent_qa_get_run, agent_qa_get_run_steps, agent_qa_get_run_artifact, agent_qa_get_run_logs, and agent_qa_get_run_execution_logs, calls agent_qa_classify_failure, identifies which surface actually failed (test definition, hook, application under test, runtime infrastructure, or agent behavior), inspects the relevant local files directly rather than inferring a patch from artifacts alone, and applies the smallest code or YAML change that the evidence actually accounts for. Before rerunning any changed Agent QA definition it validates it with agent_qa_validate_test, agent_qa_validate_suite, or agent_qa_validate_definition, then reruns only the narrowest affected test, suite, hook, or unit test within the approved environment, and reports root cause, changed files, the verification command or MCP action used, the result, and remaining risk.

Use it once a failed Agent QA run has already been triaged and now needs a code or YAML repair, when artifacts and logs point to a test, hook, product, runtime, or agent-behavior defect, when a proposed fix needs verification with the narrowest possible Agent QA or unit-test rerun, or when asked to self-heal or update a stale Agent QA definition from evidence.

It will not rewrite a test merely to make it pass when the artifact actually shows a product or runtime defect, will not invent selectors, screen states, screenshots, logs, or source files, and stops to report the blocker rather than guessing when evidence can't distinguish between materially different fixes. It doesn't authorize production changes, data mutation, dependency installation, or refactoring beyond the user's approved scope, and a passing narrow rerun is not treated as a replacement for the repository's normal test suite or human review.

Inputs: the repository, workspace, target environment, and files the user explicitly authorizes for modification, plus the failed run's evidence (steps, artifacts, logs, execution logs) and its classifier category. Output is a minimal, evidenced patch (with canonical Agent QA IDs preserved on any edited test, suite, hook, or memory file), a validation pass on any changed YAML definition, a rerun result for the narrowest affected test, and a report naming the root cause, the changed files, the verification action taken, the result, and any remaining risk or uncertainty.

Built around the Agent QA MCP toolset (agent_qa_get_run*, agent_qa_classify_failure, agent_qa_validate_*); when MCP is unavailable it falls back to dashboard REST APIs or local .agent-qa artifacts and explicitly states that MCP evidence was unavailable.

QA and engineering teams who need a failed automated test run fixed from actual evidence - not a classifier's first guess rubber-stamped, and not a test edited into passing over a real product defect - within a tightly scoped, user-approved blast radius.