Skip to content

diagnostics-repair-loop.md

Latest commit

 

History

History
61 lines (43 loc) · 1.69 KB

diagnostics-repair-loop.md

File metadata and controls

61 lines (43 loc) · 1.69 KB
Raw

Diagnostics And Repair Loop

Agent Maintainer keeps ordinary passing output quiet and writes detailed failure evidence to artifacts.

Summary-First Output

Verifier output should prioritize:

  • profile;
  • pass or fail;
  • run id;
  • duration;
  • failed checks;
  • smallest useful next command.

Raw stdout, stderr, stack traces, and long check details belong in run-scoped files under .verify-logs/runs/<run-id>/.

Failure Pointers

LAST_FAILURE.md is a convenience pointer to the latest failed run. It is not the authoritative history. When multiple agents work in one repository, another run may update that pointer.

Use run-scoped paths when repairing a specific failure:

python3 -m agent_maintainer context failures \
  --log-dir .verify-logs/runs/<run-id> \
  --limit 20

Retention

Run history retention is configured in pyproject.toml:

[tool.agent_maintainer.diagnostics]
run_history_limit = 10

The default keeps enough recent runs for overlapping agent work without letting diagnostics grow indefinitely. Use a larger value for heavily parallel repos. Use 0 only when run history must be disabled.

Repair Discipline

Read the failure note before changing code or configuration. Fix the root cause instead of lowering thresholds, adding broad suppressions, or bypassing hooks.

If a check is wrong, make the smallest correction to the check, config, or docs, and include the reasoning in the PR.

Read more: