Larkspur disruption agent: the pod repo

A disruption-care agent for Larkspur Airlines, built on nine tools that already work (support/tools.py, running against a frozen copy of real airline data in data/americas/). You build agent.py, which holds what Claude is told about each tool and the loop that drives them.

Open guide/index.html for the steps. It starts on Welcome to your build, which is the case study, then the three questions your pod answers out loud, then Get set up. After that it is the build page by page. What you are building, what to do, what it looks like when it worked, and where to look when it did not. guide/Participant-Guide.pdf is the same thing on paper.

The same steps run on the build site at https://anthropicpartnerbasecamp.bts.com/, which your pod opens together in the room. It opens on the welcome too, and the three Start pages sit under the first tab. The site adds the build clock at the top of the page and the box where you paste the evidence code a gate prints. Nothing uploads either way. Pick your surface once on Get set up and it stays picked.

Nobody in the pod has an assigned job. You decide in the moment who does what. One rule stands. One person pushes the canon at the end of the build and everyone else takes it. Agree who before the clock runs out.

The five commands

These five run the whole build.

git clone <your pod repo URL>      # once
python3 setup.py                   # until it says READY
python3 run.py K7PQ2M --trace      # every build
python3 verify.py 1.2              # every gate
python3 pod_sync.py --take-canon   # at the end of a build

setup.py checks python, the SDK, git, a real reach to the pod repo, and one live call to Claude. Every failure it prints names the fix. Run it at your own desk. There is no offline path in this pack, so a credential that does not work stops the build.

run.py shows the wire, every request and every reply exactly as it went. verify.py checks behavior on the wire and prints your evidence code. claude in this folder gives you /setup, /build, /check and /readout. Start there if you would rather not lead with a terminal.

One repo, several people

One person in the pod makes this repo from the template the room lead posts in chat. Use this template, then Create a new repository, owner their own account, Private. Then Settings, Collaborators, Add people for every podmate by GitHub username, plus the room lead's handle, which is in the chat. The overnight review needs that handle to reach your repo, and a repo nobody shared gets no file back. Then they post the repo URL in the pod thread, and nobody else does anything until it is there.

Everyone else has an invitation email waiting. Accept it before you try to clone, because a private repo refuses you until you do. Clone it rather than downloading a zip. The checks read git history.

Nobody commits agent.py mid-build. One person pushes the canon at the end of the build with python3 pod_sync.py --push-canon, and everyone else takes it with python3 pod_sync.py --take-canon. Agree who before the clock runs out.

Everyone builds their own agent.py on their own laptop. Two people editing one file in one repo gives you a merge conflict inside a loop you are all still learning to read.

--take-canon saves your version to .workshop/mine/ first and prints the path. Nothing you wrote is lost, so diff it against the canon. If you are behind or something broke, the pod repo is your checkpoint:

python3 pod_sync.py --take-canon --force

It restores what the pod pushed. Anything gitignored stays gone. Your .env (re-create it from .env.example), .venv/ (rebuild with python3 setup.py --fix), and .workshop/, which holds your banked codes and your bench numbers. Copy .workshop/ out before you re-clone into a fresh folder, and copy it back in after. If it is already gone, say so rather than re-running bench.py --label before against the current agent.

The scripts

Script	What it does
setup.py	Checks this machine. --fix builds the venv and installs. Run it until READY.
run.py <PNR> --trace	Runs the agent on one ticket and shows every turn on the wire. --all runs the five shapes and writes the totals.
verify.py <step>	A gate: 1.2, 1.3, 1.4, 2.1, 2.2, 3.1, 4.1. Run it with no step and it prints the status board.
pod_sync.py	The pod's canon: --push-canon, --take-canon, --status.
eval_harness.py	Runs evals/cases.json, your pod's own cases, against your agent.
bench.py --label <name>	Measures a run: latency, tokens, cache, cost per contact. The before/after pair around your lever.
readout.py	Writes the one page that says what your agent is and what it just did. The canon push publishes it.

TEAM.md is the pod's name and a typed roster, written once by whoever created the repo. The review reads the names off it. PITCH.md and evals/cases.json are the pod's shared record, and they are normal commits. The bench pair never leaves your laptop.

Ground rule

If you cannot explain a turn on your own trace (python3 run.py <PNR> --trace), you have not finished the step, whatever the gate says.