GitHub
Docs With Pi

coding.sh with Pi

Record Pi coding agent sessions with a Pi extension: history next to your other agents, Pi's own cost per model call, failure flags and evals.

coding.sh records your Pi sessions through a Pi extension and puts them next to the ones from your other agents, with Pi's own cost per model call, flags for runs that went wrong, and evals with Pi as the agent. It runs on your machine. Nothing is sent anywhere unless you add a sink.

Set it up

shell 
npx coding.sh                       # or: coding.sh init --agent pi

init finds Pi when pi is on your PATH or ~/.pi/agent exists. It writes one file, the extension ~/.pi/agent/extensions/coding-sh.ts (under $PI_CODING_AGENT_DIR when set), then imports the sessions in ~/.pi/agent/sessions ($PI_CODING_AGENT_SESSION_DIR). Start a new Pi session to record live: Pi loads extensions when a session starts.

The extension only listens, so it can't block or change a tool call or a prompt. coding.sh uninstall --agent pi removes it. With --scope project it goes to .pi/extensions/coding-sh.ts, which Pi loads headless only when the project is trusted or with --approve.

Look at your Pi sessions

shell 
coding.sh sessions --where agent:pi
coding.sh show latest
coding.sh usage --by agent-model --since 30d
coding.sh ingest --backfill --agent pi       # import Pi's session files again

Pi sessions show up in sessions, search, show and the viewer next to your other agents. Pi writes a cost on every reply, and coding.sh keeps it on each model call, so the figure is not an estimate. Each session gets the same flags as every other agent, such as no-verification and repeated-tool-failure, and coding.sh improve ranks what to fix first.

What is recorded

The extension records, live:

  • your prompts and Pi's replies;
  • each model call: provider, model, input, output, cache-read and cache-write tokens, Pi's cost, stop reason and duration;
  • each tool call: arguments, output, failure, duration and the file changes it made;
  • compactions, model changes and the end of each run, including runs you stopped.

Importing the session files gives the same steps, plus model changes recorded in the file and the shell commands you ran yourself in Pi (as the user_bash tool). A step seen by both is stored once.

Run evals with Pi

shell 
coding.sh eval init --sessions --agent pi --verifier 'npm test'
coding.sh eval run failures --yes

eval init --sessions drafts a task from each failed or flagged session in this repo from the last 14 days, up to 10.

eval run runs pi --mode json --approve on each task in its own Pi agent directory, with a copy of your auth.json. Your provider logins work, and your own sessions and settings are left alone.

coding.sh harness snapshot --agent pi saves the files that shape a Pi run as a bundle you can compare against another: AGENTS.md, Claude Code's instructions file, .pi/settings.json, .pi/SYSTEM.md, skills, prompts and extensions, in the project and in ~/.pi/agent.

What it does not do

  • init does not register the coding.sh MCP server with Pi.
  • Pi's thinking blocks are not stored, only the text of each reply.
  • Pi has no subagents, so there is no session tree.

Pi capture is verified on a real install (Pi 1.5.0 against a local stub model), and CI repeats that run on Linux and Windows on every change.

Next