TraceCI

TraceCI reads a failed GitHub Actions run the way a senior engineer would: it finds the step that actually failed, anchors on the real error instead of the exit code, and diffs the last green commit against the broken one to isolate what changed.

The agent is deliberately bounded — read-only, capped at six tool calls, and wired to say “not enough evidence” rather than guess. A dishonesty pass demotes confidence and strips suggested fixes whenever the diagnosis hedges.

What it does

• Anchors on the first genuinely failing step, not the generic exit-code noise
• Diffs the last successful commit against the failing one to isolate the change
• Agentic investigation that fetches only the files the evidence points to
• enforce_honesty() demotes confidence and drops fixes for hedged diagnoses
• Three-tier log windowing that strips noise but keeps traceback context
• Eval suite that tracks “confidently wrong” answers separately from low-confidence misses

Built with