Pi Coding Trial

Can Pi read a small program, reproduce a bug, edit it, and run its tests while the terminal retains both the failed and successful results for later reading?

This extends the multi-round frontend with a generated JavaScript project. It reuses the Pi event adapter, TUI SDK, and experimental terminal host. One Pi conversation and one temporary project span two prompts, with a separate display Context per turn.

Run

Use the example setup to clone the repository and install dependencies on Windows with Node 24+ and pnpm 11.9.0. Stop any example server already using port 4178. From the repository root:

pnpm prototype:pi-coding

Open http://127.0.0.1:4178/, connect, wait for [ready 1/2], and send the prefilled repair prompt. After [ready 2/2], choose Prepare follow-up prompt and send it. Check retained failure, passing tests, search and copy inspects the terminal history without making a model call. The second turn exits the child; history remains available. Reload creates a fresh project. Run only one port-4178 fixture at a time and stop the server when finished.

The default provider scripts the tool choices. Pi executes actual tools, file edits, and child-process tests.

Use a real model

This path uses an OpenAI Codex subscription through Pi. If Pi is already signed in to that provider on this machine, skip the login step. Otherwise, start the repository’s pinned Pi CLI from the repository root (no global Pi install needed):

pnpm --dir prototypes/integration/pi-session exec pi

In Pi, enter /login, select OpenAI Codex, and complete the browser sign-in. Return to Pi to finish any prompt from the login flow, then exit Pi without sending a task. Pi stores the authorization in your user directory; do not copy tokens into this repository. See Pi’s provider guide for its login flow.

Start the model-driven example:

pnpm prototype:pi-coding-live

Live mode uses the existing Pi OAuth/model setup and SSE. Each submitted prompt sends the generated code, test output, and retained conversation to OpenAI and consumes subscription allowance. Enter only sample task instructions. Model requests start only when a prompt is submitted.

1. Open http://127.0.0.1:4178/ and confirm the page is labeled LIVE.

2. Choose Connect application, wait for [ready 1/2], and send the prefilled repair prompt. Expect failing tests first, a source edit, then passing tests.

3. At [ready 2/2], choose Prepare follow-up prompt, then Send prompt. Expect another passing test run using the same project.

4. After Child exited: 0, choose Check retained failure, passing tests, search and copy. It should report PASS; the earlier failure should still be findable beside the later success. This button makes no model request.

If the page reports a sign-in error, finish the Pi login step above and reload. Model access, account limits, and network errors may also stop a live run; read the page’s Local state and native rows before retrying. A live model may choose a different sequence, so the check reports missing evidence rather than assuming success. End a running session with End session, then stop the server with Ctrl+C.

Project and Tools

total.mjs initially sums prices without multiplying quantities. total.test.mjs contains six fixed assertions. Each connection copies both into an allocated temporary directory; the repository fixtures stay unchanged.

Tool	Access
read_coding_file	Read either sample file by its exact name.
replace_coding_text	Replace one exact occurrence in total.mjs; tests stay read-only.
run_coding_tests	Execute the fixed Node command and return its exit code and captured output.

Reads, edits, and test runs are serialized. A failed assertion returns a nonzero exit code for Pi to act on. Cancellation stops the running test child and leaves completed edits in place for the next turn. Session disposal waits for tool work and removes the temporary project.

Local limits are two prompts, 24 tool calls, eight test runs, 8,192 UTF-16 units of source, 16 KiB of test output, and five seconds per test process. Each turn has a two-minute cooperative deadline, with 2,048 adapter events and 64 Blocks. The existing ten-minute session and transport budgets apply.

The test child receives an explicit minimal environment and Node --permission with read access to the generated directory. Regression checks exercise denied outside reads, writes, and child spawning. Windows supplies additional standard environment keys. These guards reduce accidental access; Node’s permission model is not a hostile-code sandbox and this Node version does not restrict networking. Use this runner for the controlled sample; arbitrary repository execution needs a separate isolation and permissions design.

Observed Results — 2026-09-22

Both a local-provider browser run and a two-prompt live run used Windows bundled ConPTY and the pinned experimental xterm host. The live run used Pi 0.86.1, OpenAI Codex subscription, gpt-5.5, low thinking, and SSE.

Step	Expected	Observed in both runs
Read and test the original program	Reproduce quantity failures	Exit 1: three passed, three failed.
Edit and rerun	Correct quantity handling	One multiplication added; exit 0: all six passed.
Follow-up without edits	Reuse the same project	All six passed again; explanation included zero quantity.
Revisit earlier output	Keep the original failure alongside success	Failure and passing result searchable; old failure copied unchanged.
Finish	Close both Contexts and clean up	Both closed before EOF; child exited 0; temporary project removed.

The live file was also inspected between turns: only * item.quantity was added. Copy checks use a browser copy event with an in-memory sink. This run keeps the 40-by-8 viewport. Earlier multi-round and paired UX records cover additional reading and resize scenarios. The live calls used subscription allowance; token usage was not measured.

Verification and Files

On 2026-09-23, the fresh-install check also completed both coding prompts with the local provider: three baseline failures, six passing tests after the edit, six passing tests in the follow-up, child exit 0, and a passing history/search/copy check. The live entry reached [ready 1/2] using existing Pi authorization; no new model request was sent in this check. The model-driven results above are from 2026-09-22.

pnpm test:pi-session
pnpm typecheck
pnpm test
pnpm build:pi-coding

coding.test.ts checks real test execution, exact edits, path and size restrictions, cancellation, timeout/output limits, cleanup, and two-turn Pi-to-Endpoint state. Browser checks run separately through ConPTY.

At the recorded checkpoint, all seven coding tests and all 283 repository Node tests passed, along with type checking and the coding browser build. The build retains the existing xterm bundle-size warning. The original three multi-round browser scenarios were also rerun and passed after the shared application changes.

workspace.ts owns files and the test process; source.ts defines Pi tools and session setup; provider.ts supplies the deterministic model fixture. The main entries and browser page reuse the multi-round application and terminal host.