Your agent ignored a failed tool call: catch it in CI with tracelint

tracelint 0.3.0 shipped on PyPI on August 17 2026. The package bundles a JSON‑Schema validator, an OpenInference trace parser, and a CLI that exits 0 for clean runs, 2 for a provable structural defect, and 3 for input errors [DevTo].

── What shipped ──

• pip install tracelint installs a zero‑dependency library (jsonschema + stdlib) that runs on Python 3.10–3.12.
• tracelint check ./trace.json --tools ./tools.json scans a trace file and emits an HTML report with line‑level evidence.
• The CLI supports OpenInference, Langfuse, and OpenAI trace formats, so existing telemetry pipelines (Arize Phoenix, OTel collectors) can point at the linter without format conversion.
• Seven rule families are enforced: schema violation (R1), tool‑error reuse (R2), hallucinated argument (R3), loop detection (R4), redundant call (R5), malformed JSON (R6), unknown tool (R7) [DevTo].
• A demo suite ships with one planted instance of each defect and a clean control, guaranteeing deterministic outcomes.

── Why it matters ──

The linter returns a non‑zero exit code only when a rule is structurally proven, so CI can fail on concrete bugs instead of relying on probabilistic LLM judges. Running locally eliminates per‑trace API fees and the seconds‑long latency of cloud LLM calls, making it practical to gate every pull request. Each finding includes the exact trace line and the rule that triggered it, giving developers precise evidence to fix agent logic.

── Editor's take ──

Deterministic trace linting changes the security model for agent pipelines: LLM judges become optional heuristics, while a failed tool call now breaks the build, forcing teams to handle error paths explicitly and reducing silent failures.

Reader poll

Which method do you rely on to catch agent‑runtime bugs?

• Deterministic trace linter (tracelint)
• LLM‑based judge
• Manual code review
• No automated checks