agent-runlog

Wrap any command and save a concise, safety-aware run log for agents.

AI agents do a lot of work through shell commands, test runs, linters, build scripts, and sub-agent CLIs. When something goes wrong, the transcript is usually messy and the useful evidence is buried. agent-runlog gives every run a small audit trail: command, duration, exit code, output tails, git state, and simple risk/loop/failure signals.

npx agent-runlog -- npm test

What it writes

By default, logs go to .agent-runs/<timestamp>/:

• report.md — human/agent readable summary
• handoff.md — compact paste-ready summary for another agent, a PR comment, or CI step summary
• run.json — structured report for automation
• stdout.log — stdout, redacted by default
• stderr.log — stderr, redacted by default

Install

npm install -g agent-runlog

Or run without installing:

npx agent-runlog -- npm test

Usage

agent-runlog [options] -- <command> [args...]

Options:

• --out DIR / -o — output directory. Default: .agent-runs/<timestamp>.
• --cwd DIR — working directory for the command.
• --quiet / -q — do not mirror child output to terminal.
• --json — print the final report JSON to stdout.
• --timeout N — stop hung commands after a duration like 30s, 5m, or 1h.
• --handoff FILE — also write the compact handoff summary to a specific file.
• --github-summary — append the compact handoff to $GITHUB_STEP_SUMMARY.
• --no-redact — store raw stdout/stderr instead of redacted logs.

Examples:

agent-runlog -- npm test
agent-runlog -o .agent-runs/lint -- npm run lint
agent-runlog --json -- node scripts/migrate.js > run.json
agent-runlog --timeout 5m -- npm test
agent-runlog --handoff /tmp/test-handoff.md -- npm test
GITHUB_STEP_SUMMARY="$GITHUB_STEP_SUMMARY" agent-runlog --github-summary -- npm test
agent-runlog --no-redact -- node scripts/debug-local.js

Why this exists

Agent observability tools are getting powerful, but sometimes you need the boring local primitive: wrap a command, capture the evidence, and make it easy for the next agent or human to see what happened.

This is useful for:

• coding agents running tests/builds
• CI repros
• long-running CLI tasks
• sub-agent harnesses
• debugging repeated failure loops

Timeout guard

Agents sometimes get stuck behind watch modes, stalled network calls, or tests that never settle. Add --timeout to turn those hangs into useful evidence instead of an endless wait:

agent-runlog --timeout 10m -- npm test

When the timeout is hit, agent-runlog sends SIGTERM, records timedOut: true and timeoutMs in run.json, marks the run as failed, and adds a timeout finding to report.md.

Agent handoffs

Every run now includes handoff.md, a shorter summary designed for the next agent or human who needs the result without reading a full transcript. It includes:

• the command, status, duration, and working directory
• links to the generated evidence files
• the top findings
• git state before/after the run
• a suggested next step

Use --handoff FILE when you want a stable summary path outside .agent-runs/, for example:

agent-runlog --handoff .agent-runs/latest-handoff.md -- npm test

In GitHub Actions, --github-summary appends the same compact summary to the job summary:

- name: Test with run evidence
  run: npx agent-runlog --github-summary -- npm test

Redaction by default

Agent logs often accidentally capture API keys, bearer tokens, private keys, or .env-style assignments. agent-runlog now redacts obvious secret-looking values before writing stdout.log, stderr.log, run.json, and report.md. The original byte counts are still recorded, and the report includes redaction counts by type.

If you explicitly need raw logs for a local-only debugging session, pass --no-redact. Be careful before sharing those logs with another agent or human.

What it detects

agent-runlog currently flags:

• repeated output lines that may indicate loops
• non-zero exits
• error/failure language
• obvious secret-looking output patterns
• very long runs

It is not a full secret scanner or observability platform. It is a tiny portable run ledger.

Library API

import { runLogged } from 'agent-runlog';

const { report, outDir } = await runLogged('npm', ['test'], { timeoutMs: 5 * 60 * 1000 });

// Preserve raw logs only when you really need them:
await runLogged('node', ['debug.js'], { redact: false });
console.log(report.analysis.summary, outDir);

License

MIT

Agent Skill

This package includes an OpenClaw/Claude-style skill at skills/agent-runlog that teaches agents when and how to wrap commands with agent-runlog for redacted run evidence and handoffs.