CRUCIBLE: AUTONOMOUS CI/CD & SELF-HEALING AGENT

Every CI/CD pipeline eventually breaks — a test fails, a build throws a runtime error, a dependency mismatch slips through. When it happens, a developer has to stop what they're doing, dig through logs, reproduce the failure locally, figure out the root cause, write a fix, and re-run the pipeline to confirm it worked. This context-switch is expensive and repetitive, especially for small, mechanical bugs (typos in config keys, missing flags, mismatched variable names) that don't need human judgment — just fast, careful triage. Crucible automates this entire failure-to-fix loop. It's an autonomous CI/CD triage and self-healing agent: it ingests a failing build or test log, isolates the root cause using pattern-based detection across both runtime errors and syntax errors, and hands the traceback to an AI patch-generation layer. The AI proposes a structured fix (broken into a triage summary, the actual patch, and a verification note), which Crucible then applies and validates by re-running the test suite. If the fix passes, the developer gets a ready-to-commit patch instead of a red pipeline and a stack trace. Target users are software teams and individual developers running CI/CD pipelines (GitHub Actions, pytest-based test suites) who want faster recovery from routine failures without waiting on a human to triage every red build. The interaction model is deliberately lightweight: point Crucible at a target repo/log with a single CLI flag, and it runs the full ingest → triage → patch → verify → report pipeline end to end, writing a JSON report of what it found and fixed. What makes it creative and unique is the closed-loop verification step: Crucible doesn't just suggest a fix and stop — it re-runs the actual test suite against its own patch before presenting it, so what reaches the developer has already been proven to work, not just proposed. It also handles two distinct failure shapes (runtime failures and syntax errors) with dedicated detection logic rather than a single brittle regex, and the AI prompt is hardened with explicit structural rules so the output is reliably parseable rather than free-form text. We validated this on two realistic demo scenarios: a billing calculation bug caused by a key mismatch (tax_rate vs tax_pct), and a CI/CD pipeline break caused by an invalid subprocess flag combined with a missing check=True. In both cases, Crucible correctly triaged the root cause, generated a working patch, and verified it via re-run — demonstrated live on our dashboard, which tracked 25 total healing runs with a 48% end-to-end success rate. The project is efficient by design: it only escalates a human when its own verification step fails, so its output is either a proven fix or nothing at all — no false confidence, no unverified guesses landing in your codebase.