Audit AI Coding Sessions for Cost, Latency & Failures

This skill inspects local AI coding-agent sessions using agenttrace, a local-first observability tool that reads session logs from Claude Code, Cursor, Gemini CLI, Aider, Codex CLI, OpenCode, Qwen Code, Kimi, and generic JSON/JSONL traces. It surfaces token and cost spikes, tool failures, retry loops, latency gaps, anomalies, health scores, and session-to-session diffs so you can understand why a run was slow, expensive, shallow, or unreliable.

Use this skill when a user asks why an AI coding run was slow, expensive, shallow, or unreliable. Use it when reviewing local agent logs before retrying a failed or suspicious task, when building a lightweight CI health gate for AI-assisted coding sessions, or when comparing two attempts to find changed tool paths, retries, or cost patterns.

Do NOT use this skill when session logs are unavailable locally or when you need production LLM application tracing (use @langfuse instead). Healthy trace metrics do not prove the final code is correct - still run tests and review diffs. Token and latency metrics can look healthy even when an agent confidently takes the wrong implementation path; pair trace audits with diffs against known-good attempts when semantic drift is the risk.

You provide local session logs or exported session files from supported AI coding tools. The skill discovers available sessions, produces human-readable Markdown reports or JSON output for automation, inspects individual sessions, compares attempts to detect semantic drift, and optionally enforces CI health thresholds.

Start with session discovery:

agenttrace --doctor
agenttrace --overview

For a concise human-readable audit:

agenttrace --overview -f markdown -o agenttrace-overview.md

Inspect the latest session or a specific export:

agenttrace --latest
agenttrace --latest -f json
agenttrace path/to/session-or-export.json
agenttrace --overview -d path/to/session-dir

For CI automation gates:

agenttrace --overview -f json -o agenttrace-overview.json
agenttrace --overview --fail-under-health 80 --fail-on-critical --max-tool-fail-rate 15

You receive Markdown or JSON reports analyzing sessions for anomalies, tool failures, token and cost patterns, latency gaps, health scores, and shallow sessions. JSON output enables programmatic CI gates; Markdown output supports human review and sharing.

Compatible with Claude Code, Cursor (exports), Gemini CLI, Codex CLI, Aider, OpenCode, Qwen Code, Kimi, and generic JSON or JSONL trace formats. Reads session logs locally without requiring external services.

Developers using AI coding assistants who need to understand why a run failed, cost too much, or produced unexpected results. Teams building CI health checks for agent-assisted workflows. Engineers comparing session attempts to detect when an agent took the wrong implementation path despite healthy cost and latency metrics. Use @langfuse for production LLM application tracing and @observability-engineer for broader service monitoring and SLOs.