fix: ship the codemode sandbox in pie-pi-process

[oxwen11]

Requirement

Every codemode call in Pie fails, so MCP servers with the default (codemode) exposure are unusable. Found while checking whether MCP, codemode and tool_search load in Pie. This already happens on main (Pi 0.99.1 bundle), so it is not a regression from #433. Stacked on #436 because it extends the same Pi 1.0.2 patch.

The bundle shows two separate failures:

1. Bun's .wasm file import yields ./quickjs-<hash>.wasm, relative to the bundle. Pi's bun/runtime-setup.js passes that string through as is, so it is read relative to the Session's cwd: Failed to load QuickJS: ENOENT … './quickjs-r1pq2107.wasm'.
2. Pi resolves the codemode worker as codemode-worker.js beside the bundle (PI_BUNDLED_NODE), but build-pi-process.ts never built it: Cannot find module …/codemode-worker.js.

Expected behavior

• codemode scripts run in the sandbox whatever the Session cwd is.
• Scripts can call MCP tools and built-in tools.
• Compiled Bun binaries (absolute $bunfs paths) are unaffected.

Changes and risks

• patches/@earendil-works__pi-coding-agent@1.0.2.patch: in dist/bun/runtime-setup.js, a relative wasm path is resolved against import.meta.url. It uses fileURLToPath, so install paths with spaces (macOS Application Support) work.
• build-pi-process.ts: builds Pi's dist/extensions/codemode/worker.js into dist/pi-process/codemode-worker.js, the same way as image-resize-worker.js.
• pnpm-lock.yaml: only the patch hash changes.
• The upstream fix belongs in Pi's runtime-setup; the patch hunk can be removed once Pi ships one.

Verification

Tested revision: c66cb336.

• New rpc-bundle.test.ts case:
  • copies the bundle into runtime/ and runs it with a different cwd;
  • activates codemode, and the provider calls it with return 6 * 7;
  • asserts isError:false and output 42.
  • Removing codemode-worker.js makes it fail with Cannot find module. Restoring the old wasm line makes it fail with ENOENT './quickjs-…wasm'. With the fix it passes.
• pnpm build, pnpm check and the node tests pass: 1342 passed, 1 skipped.
• Web runtime through the Pie daemon (pie-verify web, isolated PIE_HOME and PI_CODING_AGENT_DIR, fake e2e provider, three stdio MCP fixtures with direct, codemode and deferred exposure):
  • Declared tools: codemode and tool_search, plus the direct MCP tools.
  • codemode calling tools.mcp__code_tools__say_hello returns hello from code-tools.
  • tool_search loads 2 deferred tools, and mcp__deferred_tools__weather_lookup then returns sunny per deferred-tools.
  • Direct mcp__dev_tools__say_hello returns hello from dev-tools.
• Also checked with a script that sends RPC commands straight to the bundle: a codemode script calls tools.bash({command:"pwd"}) and returns the Session cwd.
• Not verified:
  • a real model writing codemode scripts;
  • the packaged Desktop app (the build output directory is copied whole, but I did not rerun electron-builder for this change).

[oxwen11]

Web runtime evidence at c66cb33: codemode→MCP call, tool_search→deferred tool call, and a direct MCP call through the Pie daemon (screenshots + recording-002).

recording-002.webm

[pkg-pr-new]

npx https://pkg.pr.new/oxwen11/pie/@getpie/cli@442

commit: 499d016

[github-actions]

React Doctor skipped this pull request — it changed no React files.

_{Reviewed by React Doctor for commit 499d016.}

[oxwen11]

Rebased onto #436 (aff3ff7) on top of main db2acfa. New head: b4065d3. The runtime-setup hunk is reapplied to the regenerated Pi patch, and the lockfile still changes only the patch hash.
Re-verified at this head:

• pnpm build and pnpm check pass.
• Node tests: 1371 passed, 1 skipped.
• The rebuilt bundle contains codemode-worker.js and quickjs-*.wasm.
• A bundle codemode script returns 42, and codemode calling an MCP tool returns hello from code-tools.
• The MCP harness passes all 18 checks.

[oxwen11]

Independent review: ready

• Reviewed head: 499d01681c4b0a8272c2fa3389cd8eda10d96a2c
• Base: main @ e8ad3fa89218ea658408ebd8b73d5de8ccc29714
• Trusted rules: origin/main (.agents/rules last changed in 7b4d86b1), pull-requests.md, review.md, acceptance.md, toolchain.md

1. CI

Every required check finished successfully on 499d0168: Check (run 37439928664), react-doctor (37439928711), and Publish @getpie/cli preview (37439928621 and 37439884956). No check was pending, skipped or stale.

2. Review

• Patch hunk (dist/bun/runtime-setup.js): only a relative path (./quickjs-<hash>.wasm, which Bun emits for a non-compiled --outdir bundle) is resolved against import.meta.url. In the bundle, that URL is pi-process.js, and the wasm sits next to it. Absolute $bunfs paths from compiled binaries pass through unchanged. fileURLToPath decodes %20, so install paths with spaces work. The worker gets the wasm from the main thread through workerData; it does not resolve the file again.
• Worker build: Pi's resolveCodemodeWorkerSpecifier("bundled-node") returns new URL("./codemode-worker.js", import.meta.url). The new loop emits that file with the same flags as image-resize-worker.js, and a build failure still stops the build.
• Lockfile: only the patch hash changes. I checked this by diffing with the hashes normalised.
• Distribution: the CLI copies the whole dist/pi-process directory. npm pack --dry-run for @getpie/cli lists codemode-worker.js, quickjs-r1pq2107.wasm and pi-process.js. Desktop asarUnpack covers @getpie/server/dist/pi-process/**, so both new files ship as real files.
• Security: no new trust boundary. The path being resolved is a constant that Bun writes at build time, never user input. Codemode still runs in the existing QuickJS worker sandbox.

3. Independent verification

Run in a clean reviewer worktree at 499d0168, with pnpm install --frozen-lockfile and pnpm build through Turbo. dist/pi-process was freshly built and contains codemode-worker.js and the patched setEmbeddedQuickJSWasmPath(...startsWith(".") ? fileURLToPath(new URL(...)) ...).

• vitest run rpc-bundle.test.ts fff-island.test.ts: 7/7 pass.
• pnpm --filter @getpie/server test: 648 passed, 1 skipped.
• Negative controls on the built bundle:
  • Removing codemode-worker.js makes the new test fail with Cannot find module …/runtime/codemode-worker.js.
  • Restoring setEmbeddedQuickJSWasmPath(quickjs_default) makes it fail with Failed to load QuickJS: ENOENT … './quickjs-r1pq2107.wasm'.
  • Restoring both makes it pass again. The test therefore catches both regressions.
• Runtime check against the bundle:

  • Setup: bundle copied to a path containing a space (…/App Support/srv/pi-process/, with the sibling fff/). The Session cwd was a separate project directory.

  • Isolation: separate HOME and PI_CODING_AGENT_DIR, fake e2e provider, no real model. Three stdio MCP fixtures with direct, codemode (default) and deferred exposure.

  • Declared tools include codemode and tool_search.

  • Codemode results:

    Script	Result
    return 6*7	42
    tools.mcp__code_tools__say_hello({})	hello from code-tools
    tools.bash({command:"pwd"})	the Session cwd
    throw new Error("boom")	isError with Error: boom (failure path)

• Installed CLI layout: the same check against packages/pie/dist/pi-process/pi-process.js returned sunny per code-tools from tools.mcp__code_tools__weather_lookup.

Gaps (not blocking)

• No real model wrote codemode scripts. Calls were driven by the fake provider.
• I did not rebuild or launch the packaged Desktop app with electron-builder. Packaging coverage is based on the asarUnpack glob above and on CI's Desktop e2e baseline.
• I did not repeat the run through the Pie daemon or web UI. The change only affects code inside the pie-pi-process bundle, which I exercised directly with Pie's spawn arguments.

Re-review needed if

The head or base changes, the Pi version or patch changes, or build-pi-process.ts output layout changes.

[oxwen11]

Merged the independently reviewed and verified head 499d016 (record: #442 (comment)) onto base e8ad3fa by squash. Merge commit: 91247ad.