agent-trajectory

English | 中文

See what your AI agent actually did — turn by turn, tool by tool, token by token.

Interface and interaction model based on the trajectory view in DeepSeek Harness.

---

`agent-trajectory` turns a Claude Code or Codex session log (or normalized `agent-trajectory/v1` JSON from another host) into a trajectory viewer: a per-turn ledger of every tool call with its arguments, result, duration and status; a zoomable timeline with a context-window usage curve; and an optional **learn tier** where an AI annotates every step — what the turn set out to do, why each pivotal call was made, and a short storyline of what actually happened, failures included.

Built for people learning how agents work under the hood, and for anyone who wants to review what an agent did on their machine.

Highlights

• Two tiers. Raw is pure log parsing — no model calls, instant. Learn adds AI-written annotations (per-turn goal, per-step why, per-turn storyline), grounded strictly in the parsed events so it can't invent steps.
• Live mode. scripts/serve.py watches the session log and updates the page while the agent works. Real-time, local, zero dependencies.
• Ask AI. Hover any step, press 问AI — the question lands in a notebook drawer. In live mode it's answered in-page by a local CLI (claude -p / codex exec / anything via $AGENT_TRAJECTORY_ASK_CMD); in a static page it's copied for pasting into any chat.
• Self-contained output. One HTML file, no network, no build step, light and dark, print-to-PDF ready. Send it to anyone.

Quick start

# live view of your current session (run inside your project directory)
python3 scr