legal-agent-workshop

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

Legal Agent Workshop

Two working sessions. Participants clone this repository and launch Muse Code from agent/ only; cases/, world/, and solutions/ stay outside Muse Code’s workspace. First you build a general legal agent; then you point it at a case.

• Session 1 — Build the agent: the agentic loop, in plain OpenAI Python SDK code, on a pre-built kit (one file per tool, Files API document reading, a case-management system with quote-checked facts). The agent drafts a checklist, plans, investigates, and ends with an unverified memo.
• Session 2 — Make it trustworthy: first verify the Session 1 memo and checklist and see what they miss (./helper verify). Then build the two checks of the verification layer (the checklist against the case’s ground truth, and the memo rules), which the kit runs outside the agent’s control, and continue the Session 1 run from its checklist with verification on (./helper branch): refusals, escalation, release of new evidence, resume, and attorney sign-off.

The Agent Console in ui/ (pre-built) shows any run live: python3 ui/ui.py, then open http://127.0.0.1:8765.

Start

git clone https://github.com/meta-models/legal-agent-workshop.git
cd legal-agent-workshop/agent
cp .env.example .env      # add MODEL_API_KEY; CASE_DIR points at ../cases/northbridge-v-caldera
./helper setup            # creates .venv with the OpenAI Python SDK, copies the case's documents, checks the key
muse

Start Muse Code

From agent/, run muse. Your first message, and again after any restart or new session, is the ground-rules prompt in HANDBOOK.md, section 1.7.

Structure

• HANDBOOK.md — the participant handbook: setup, Session 1, then Session 2 step by step. It holds the ground-rules prompt for Muse Code (1.7).
• agent/ — starter code and the only Muse Code workspace: a general legal agent, no case inside. See agent/README.md.
• cases/northbridge-v-caldera/ — the case you hand the agent: objective.md and documents/. A new case is just a folder with an objective and its documents.
• ground_truth/ — what the checklist check compares against: the ground-truth checklist per case (northbridge-v-caldera.json). Never part of the case, never seen by the agent or Muse Code.
• world/ — the simulated outside world: withheld follow-up documents and release.py.
• solutions/ — 01_session1/ and 02_session2/ snapshots.
• ui/ — the Agent Console: start runs, watch each turn, see the agent’s checklist and how each item is proven, release evidence, sign off. Pre-built; outside Muse Code’s workspace.

Facilitator commands

python3 world/release.py                 # after a run reaches needs_input
cd agent && ./helper verify RUN_ID --memo        # Session 2: what the Session 1 memo misses
cd agent && ./helper verify RUN_ID --checklist   # Session 2: what the Session 1 checklist misses
cd agent && ./helper branch RUN_ID               # Session 2: continue from the checklist, verification on