AgentDiff

Two AI coding-agent sessions. One side-by-side failure diff — plus a plain-English “why they forked.”

Point it at two Claude Code or Codex runs of the same task. AgentDiff aligns their tool-call timelines, jumps to the first divergence, writes a narrative summary you can paste into Discord/PRs/X, and ships a single offline HTML report.

Install

npx @barood-labs/agentdiff compare run-a.jsonl run-b.jsonl --out report.html
# or two session folders:
npx @barood-labs/agentdiff compare ~/.claude/projects/.../a ~/.claude/projects/.../b
npx @barood-labs/agentdiff demo --open

# Markdown summary (Discord / PR / X-ready):
npx @barood-labs/agentdiff compare a.jsonl b.jsonl --md

Open report.html. No account. No upload. No cloud deps in the viewer. Hit Copy summary in the report banner to grab a shareable one-liner + bullets.

Live gallery: barood-labs.github.io/agentdiff

Why

AgentTown replays one session as a city. WhyDiff scores one diff. AgentDiff answers the question you actually ask in Discord:

“These two agents got the same prompt — where did the bad one go off the rails?”

Typical divergences it highlights:

• different files edited
• different shell commands
• one side errored / failed tests
• steps only present on A or B

What you get

| Piece | Detail | |-------|--------| | CLI | compare, demo, --json, --md, --out, --open, folder or .jsonl | | Parsers | Claude Code + Codex JSONL (adapted from AgentTown, MIT); home paths redacted to ~ | | Align | LCS on actionable steps (read/edit/write/bash/test/…) | | Explain | Plain-English headline + bullets (explainDiff / renderMarkdown) | | Report | Dark single-file HTML: story banner, dual timeline, scrubber, “first divergence” jump, Copy summary, tool histograms, drift score | | Keys | j/k next/prev ≠ · ←/→ step · f jump to first | | Fixtures | examples/login-bug, flaky-test, wrong-file |

Drift score

Rough 0–100 “how far apart”: share of diverging aligned rows, plus file asymmetry and error asymmetry. Not a model quality ranking — a debug compass.

Library

import {
  loadSession,
  compareSessions,
  renderReport,
  explainDiff,
  renderMarkdown,
} from "@barood-labs/agentdiff";

const a = loadSession("a.jsonl");
const b = loadSession("b.jsonl");
const diff = compareSessions(a.session, b.session);
console.log(explainDiff(diff).headline);
console.log(diff.firstDivergence, diff.drift, diff.stats);

Gallery

After npm run build, open docs/index.html for the three fixture reports — or use the live gallery. Push to main also deploys Pages via .github/workflows/pages.yml.

Develop

npm test          # node:test
npm run build     # write docs/*.html gallery
npm run demo      # compare login-bug → /tmp/agentdiff-demo.html

Not in v0.2

Live webhooks, cloud agent APIs, Cursor log format, GIF export, SaaS.

License

MIT © barood-labs

Session parsers adapted from @barood-labs/agenttown (MIT). Divergence framing inspired by @barood-labs/whydiff (MIT).