AgentTrace
Audit your AI coding-agent sessions locally — cost, tokens, latency, failures, and health — without any data leaving your machine.
The skill mandates local-only processing, no log upload, no overwriting user reports, and no invented metrics; the repo has SECURITY.md and dependency review, and CI uses least-privilege permissions. Deducted for unverified publisher identity and a curl|sh install path whose checksum logic is not visible within the skill's evidence.
The workflow is self-consistent with --doctor fallback and explicit no-session guidance; CI runs tests, output-contract and docs-command smoke checks. Deducted because this is a static read: actual test files are not in evidence, malformed-input behavior is only inferable, so the static ceiling applies.
Trigger description is precise across many agent log sources with capability-level non-fit boundaries (Detailed/Aggregate/Limited); local-first design suits mainland-China networks and a Chinese README exists. Deducted because binary acquisition via cargo/Homebrew/npm has unverified domestic reachability and boundaries live mainly in README, not the skill file.
MIT license, layered docs index, CI governance and contribution flow are present, and SKILL.md is well structured. Deducted for the 0.0.0-dev workspace version, no visible changelog or versioning commitment, unverified maintainer identity, and missing install prerequisites in the skill file itself.
The skill supplies directly copyable command sequences and report-focus requirements covering real cost/token/latency/failure audit needs. Deducted because static review cannot confirm outputs are directly usable, costs are estimates, and comparative benefit over manual analysis lacks verified evidence, so the effectiveness ceiling applies.
SHA-pinned CI workflows plus output-contract, determinism, semantic-consistency and npm package-contract checks provide multiple evidence types. Deducted because all are self-reported repo material; static review cannot independently reproduce results, and awesome-list mentions do not verify function, so the score stays at or below 5.
- This assessment is a purely static source review; nothing was executed and confidence is low.
- One recommended install path is curl | sh remote script execution; review install.sh and verify checksums before use.
- Publisher is unverified and treated as unknown; workspace version is 0.0.0-dev with no changelog evidence.
- Cost and delivery signals are explicitly estimates/heuristics, not provider billing or proof of merge.
- npm install requires Node.js 18+; reachability of Homebrew/npm/crates.io from mainland-China networks is unverified.
What does this skill do, and when should you use it?
AgentTrace is a local-first Rust TUI/CLI for auditing AI coding-agent session history. It parses logs from Claude Code, Codex CLI, Gemini CLI, Qwen Code, Cline, Aider, Cursor exports, OpenCode, Kimi CLI, and a dozen-plus other sources plus generic JSON/JSONL traces, aggregating spend, token usage, wall-clock time, and slow-run causes. Each session gets an honest Detailed/Aggregate/Limited capability level so missing event-level evidence is never presented as a complete trace. Output options include an interactive TUI, Markdown/JSON/HTML reports, and CI health gates with configurable thresholds. Everything runs locally; prompts and code never leave your machine.
Reads existing local AI agent session logs (a dozen-plus sources plus generic JSON/JSONL); runs agenttrace to open the TUI or produces reports via --overview, --latest, --audit, --recommend, --mcp-governance, and related commands; computes cost, input/output/cache tokens, latency stats, long gaps, retry loops, tool failure rates, and health scores; compares against a local baseline for regression evidence and offers read-only heuristic Git commit correlation; enforces CI gates via flags like --fail-under-health and --max-tool-fail-rate; accepts pricing-overrides. for model aliases and per-million-token price overrides.
- A heavy multi-agent user wants a month-end rollup of token spend by model and project
- A developer debugging why a task ran slowly — checking for long gaps, retry loops, or slow tool calls
- A team lead adding CI health gates, e.g. blocking when tool failure rate exceeds 15% or critical sessions exist
- A privacy-conscious engineer auditing agent behavior without uploading logs anywhere
- Someone who needs to share audit results via self-contained HTML or Markdown reports
What are this skill's strengths and limitations?
- Local-first: logs never leave your machine, privacy-safe
- One binary audits a dozen-plus agent log formats
- Honest capability levels per session; missing evidence is never shown as zero
- JSON/Markdown/HTML output integrates directly into CI gates
- Broad distribution: Homebrew, npm, WinGet, cargo, install scripts
- Costs are estimates, not provider billing; pricing inference may need manual overrides
- When source logs lack event-level data, latency and tool-failure evidence is unavailable (marked as such, but functionality is limited)
- Git delivery correlation is a read-only heuristic only; it cannot prove a commit reached main
- Install notes indicate some package names only become available after their corresponding release is published
How do you install this skill?
Install the agenttrace binary via any of these methods:
- macOS/Linux: brew install luoyuctl/tap/agenttrace
- Cross-platform (Node.js 18+ required): npm install -g @zack78/agenttrace
- Windows: winget install --id Luoyuctl.AgentTrace --exact
- Manual: curl -fsSL https://raw.githubusercontent.com/luoyuctl/agenttrace/master/install.sh | sh or cargo install --git https://github.com/luoyuctl/agenttrace agenttrace
Verify with agenttrace --version. Place the skill folder (skills/agenttrace-session-audit/) in your client's Skills directory.
How do you use this skill?
Run agenttrace bare to open the TUI; or ask your agent to run an audit, e.g. the prompt: "Use agenttrace to audit my recent Claude Code sessions and show the costliest and most failure-prone ones." Common commands:
- agenttrace --doctor (diagnose detected agent directories)
- agenttrace --overview -f markdown -o agenttrace-overview.md (human-readable report)
- agenttrace --overview -f (for automation)
- agenttrace --overview --fail-under-health 80 --fail-on-critical --max-tool-fail-rate 15 (CI gate)
- agenttrace --latest (most recent session)
- agenttrace path/to/session. (audit a specific file)