oh-my-pi/docs/tools/eval.md

# eval

> Execute Python or JavaScript code in persistent cell-based runtimes.

> **Notice:** Do not shell out to `python -c`/`python -e`, `bun -e`, or `node -e` via the `bash` tool for ad-hoc code execution. Use this tool instead — it gives you persistent state across cells, structured `display()` output, image/JSON capture, and proper cancellation/timeout handling that one-shot `-e`/`-c` invocations cannot provide.

## Source
- Entry: `packages/coding-agent/src/tools/eval.ts`
- Model-facing prompt: `packages/coding-agent/src/prompts/tools/eval.md`
- Key collaborators:
  - `packages/coding-agent/src/eval/backend.ts` — backend execution contract
  - `packages/coding-agent/src/eval/js/index.ts` — JS backend adapter
  - `packages/coding-agent/src/eval/js/executor.ts` — JS execution + output sink
  - `packages/coding-agent/src/eval/js/context-manager.ts` — persistent VM contexts, prelude, tool bridge
  - `packages/coding-agent/src/eval/js/prelude.txt` — JS global helpers
  - `packages/coding-agent/src/eval/py/index.ts` — Python backend adapter
  - `packages/coding-agent/src/eval/py/executor.ts` — kernel session retention, reset, cleanup
  - `packages/coding-agent/src/eval/py/kernel.ts` — Jupyter gateway/kernel protocol, display capture
  - `packages/coding-agent/src/eval/py/prelude.py` — Python helper functions and status events
  - `packages/coding-agent/src/session/streaming-output.ts` — truncation, artifacts, streamed chunks
  - `docs/python-repl.md` — Python kernel/gateway internals

## Inputs

Tool parameters are a JSON object with a single `cells` field — an ordered array of cell objects. Each cell is a structured record; there is no `*** Cell` header parsing, no language sniffing, and no implicit single-cell fallback. Cells run in array order; state persists within each language across cells and across tool calls.

| Field | Type | Required | Description |
| --- | --- | --- | --- |
| `cells` | `EvalCellInput[]` | Yes | Cells executed in order. At least one cell is required (`.min(1)`). |

Each `EvalCellInput` (from `evalCellSchema` in `packages/coding-agent/src/tools/eval.ts`):

| Field | Type | Required | Description |
| --- | --- | --- | --- |
| `language` | `"py" \| "js"` | Yes | Backend selector. `"py"` maps to the IPython/Jupyter kernel (`python` backend); `"js"` maps to the persistent JavaScript VM. |
| `code` | `string` | Yes | Cell body, verbatim. JSON-encoded — embed newlines, quotes, and indentation directly; no fences, no headers. |
| `title` | `string` | No | Short label rendered in the transcript (e.g. `"imports"`, `"load config"`). |
| `timeout` | `integer` | No | Per-cell timeout in seconds, clamped to `1..600`. Defaults to 30 when omitted. |
| `reset` | `boolean` | No | Wipe this cell's language kernel before running. Reset is per-language: a `py` cell's reset does not touch the JS VM and vice versa. Defaults to `false`. |

Minimal example matching the live schema:

```json
{
  "cells": [
    { "language": "py", "title": "imports", "timeout": 10, "code": "import json\nfrom pathlib import Path" },
    { "language": "py", "title": "load config", "code": "data = json.loads(read('package.json'))\ndisplay(data)" },
    { "language": "js", "title": "summary", "reset": true, "code": "const data = JSON.parse(await read('package.json'));\ndisplay(data);\nreturn data.name;" }
  ]
}
```

## Outputs

Final result from `EvalTool.execute()` is single-shot, but `onUpdate` streams partial text and `details` while cells run.

Returned shape:

- `content`: one text block containing combined cell output, or `(no text output)` / `(no output)` when only rich outputs exist.
- `details` (`EvalToolDetails` from `packages/coding-agent/src/eval/types.ts`):
  - `cells`: per-cell code, status (`pending`/`running`/`complete`/`error`), output, duration, exit code, status events, markdown flag
  - `language`: first backend used
  - `languages`: distinct backends used, in first-use order
  - `jsonOutputs`: structured values emitted via `display(...)`
  - `images`: image payloads emitted by Python rich display or JS `display({ type: "image", ... })`
  - `statusEvents`: aggregated helper/tool status events
  - `notice`: backend fallback notice (currently unused; reserved for future per-cell notices)
  - `meta`: truncation metadata
  - `isError`: set on cell failure or cancellation

Renderer behavior in `packages/coding-agent/src/tools/eval.ts`:

- call preview renders each cell's `code` with syntax highlighting based on its declared `language`
- result view renders each cell separately, including status, duration, and output
- markdown outputs are rendered with the Markdown component instead of plain text
- `jsonOutputs` render as a tree, collapsed or expanded depending on UI state
- timeout / truncation notices render as dim metadata lines
- images are carried in `details.images`; generic tool UI image handling renders them outside the text block

Side-channel artifacts:

- `session.allocateOutputArtifact?.("eval")` may allocate an `artifact://...` backing store for spilled output.
- Truncated output metadata points at that artifact when available.

## Flow

1. `EvalTool.execute()` in `packages/coding-agent/src/tools/eval.ts` receives `params.cells` already validated by the Zod schema — no string parsing step.
2. For each cell, `execute()` maps `cell.language` to an `EvalLanguage` (`"py"` → `"python"`, `"js"` → `"js"`) and calls `resolveBackend(session, language)`:
   - `python` is gated on `eval.py !== false` and `pythonBackend.isAvailable(session)`.
   - `js` is gated on `eval.js !== false`.
   - A disabled or unavailable requested backend throws `ToolError`; there is no auto-fallback or sniffing.
3. The tool allocates an `OutputSink`, a `TailBuffer`, per-cell result objects, and a `sessionAbortController`. `session.trackEvalExecution?.(...)` can wrap the whole run for external cancellation tracking.
4. It resolves the executor session id from `session.getEvalSessionId?.()`, falling back to `defaultEvalSessionId(session)`. Subagents inherit the parent's id so both sides share the same JS VM and Python kernel for each backend.
5. Cells execute sequentially within one eval tool call. For each cell, `execute()`:
   - clamps `(cell.timeout ?? 30) * 1000` ms through `clampTimeout("eval", ...)`
   - builds a combined abort signal from the tool signal, the timeout, and the session abort controller
   - marks the cell `running` and emits an update
   - calls the backend’s `execute()` with `cwd`, `sessionId`, `sessionFile`, `kernelOwnerId`, `deadlineMs`, `reset` (defaults to `false`), artifact info, and chunk callback
6. JS cells dispatch through `packages/coding-agent/src/eval/js/index.ts` into `executeJs()`; Python cells dispatch through `packages/coding-agent/src/eval/py/index.ts` into `executePython()`.
7. Backend text chunks stream into the shared `OutputSink`; rich outputs are accumulated separately as JSON, images, markdown markers, and status events.
8. After each cell:
   - text output is trimmed and stored on that cell result
   - multi-cell runs prefix text with `[i/n]` and the optional title
   - cancellations return early with `isError: true` and a cell-specific abort message
   - non-zero exit codes return early with `isError: true` and a message naming the failed cell
   - later cells are skipped after the first error, but earlier cell state persists in the underlying runtime
9. On success, the tool joins all cell outputs, synthesizes `(no text output)` or `(no output)` when needed, and attaches truncation metadata from `summarizeFinal()`.
10. The renderer uses `details.cells`, `details.jsonOutputs`, and `details.statusEvents` to build notebook-style output. `mergeCallAndResult = true` and `inline = true`, so call and result render together in the transcript.

## Modes / Variants

### Backend selection

Backend choice is **explicit per cell** — there is no auto-detection.

- `language: "py"` → Python (IPython/Jupyter) backend
- `language: "js"` → JavaScript VM backend

If the requested backend is disabled or unavailable, the tool throws `ToolError` for that cell. The caller chooses; the tool does not silently substitute.

### JavaScript runtime

Implemented in `packages/coding-agent/src/eval/js/context-manager.ts` and `packages/coding-agent/src/eval/js/prelude.txt`.

- Persistent worker-backed VM sessions keyed by `js:${sessionId}`
- `reset: true` calls `resetVmContext(sessionKey)` before the cell executes; reset is destructive for all live runs on that JS session
- Top-level `await` and bare `return` are supported by wrapping code in an async IIFE when `wrapCode()` sees `await` or `return`
- Top-level static `import ... from ...` and dynamic `import(...)` calls are routed through `rewriteImports()`, which sends them via `__omp_import__` so the specifier resolves against the session cwd
- Module cache is busted for **local** imports between cells so edits to source files are picked up without restarting the runtime. `__omp_import__` deletes `require.cache[absPath]` before re-importing whenever the original specifier is a filesystem path: relative (`./x`, `../x`, `.`, `..`), POSIX-absolute (`/...`), home-prefixed (`~/...`), or Windows drive-letter (`C:\...` / `C:/...`). Bare specifiers (`react`, `lodash/x`) and URL/scheme specifiers (`node:fs`, `file://...`, `https://...`) are left in cache so package identity stays stable across cells. The cache-bust only fires when the resolved target is an absolute path — unresolved bare-package fallbacks (`resolveImportSpecifier()` returning the original specifier) skip it.
- The prelude installs globals:
  - `display`, `print`
  - `read`, `write`, `append`, `sort`, `uniq`, `counter`, `diff`, `tree`, `env`, `output`
  - `tool.<name>(args)` proxy for arbitrary session tool calls
- JS helpers are async because they cross the VM/tool boundary
- `display(value)` behavior:
  - plain objects/arrays become JSON outputs
  - `{ type: "image", data, mimeType }` becomes an image output
  - scalars become text
- The VM exposes a restricted `process` subset plus `Buffer`, `fetch`, `Blob`, `File`, `Headers`, `Request`, `Response`, `fs`, `require`, and browser-style globals
- Concurrent runs on the same VM are not queued end-to-end. Synchronous JS still runs on the single event loop; awaited regions can interleave with sibling runs.

### Python runtime

Implemented in `packages/coding-agent/src/eval/py/executor.ts`, `packages/coding-agent/src/eval/py/kernel.ts`, and `packages/coding-agent/src/eval/py/prelude.py`. See `docs/python-repl.md` for gateway and kernel details.

- Default mode is retained `session` kernels keyed by `python:${sessionId}`
- Optional `python.kernelMode = "per-call"` creates a fresh kernel for each cell and shuts it down afterward
- `reset: true` disposes the retained kernel for that session before the cell runs; later Python cells in the same tool call reuse the fresh kernel
- Startup path:
  - availability check
  - create/connect kernel
  - initialize cwd / env / `sys.path`
  - execute `PYTHON_PRELUDE`
- Python cells run in the runner's persistent asyncio event loop, so top-level `await` works; the prompt warns not to use `asyncio.run(...)`
- The Python prelude defines helpers with the same surface as JS where practical, including `tool.<name>(args)` through a per-run loopback bridge
- Synchronous statement blocks run in the default executor with ContextVar state copied in; the GIL still serializes bytecode execution, but awaited regions can interleave with sibling cells
- Kernel `display_data` / `execute_result` messages map to:
  - `application/x-omp-status` → status event
  - `image/png` → image output
  - `application/json` → JSON output
  - `text/markdown` → markdown output
  - `text/plain` → text output
  - `text/html` → HTML converted to markdown with `htmlToBasicMarkdown()`
- Interactive stdin is rejected: `input_request` sends an empty reply, marks `stdinRequested`, and the executor returns exit code `1`

### Multi-language call behavior

A single tool call can mix Python and JS cells. Persistence is per language runtime:

- `reset: true` on a Python cell does not touch JS state
- `reset: true` on a JS cell does not touch Python state
- each backend keeps its own retained session keyed from the same session-derived ID

## Side Effects

- Filesystem
  - JS/Python prelude helpers can read, write, append, diff, and traverse files under the session cwd or absolute paths.
  - Output may spill to an artifact file via `OutputSink`.
- Network
  - Python backend speaks NDJSON to a local `python3` subprocess over stdin/stdout (no network).
  - JS runtime exposes `fetch` and `tool.<name>()`; those tools may perform additional network I/O.
- Subprocesses / native bindings
  - Python availability check runs `<python> -c ...`.
  - Python backend spawns one `python -u runner.py` subprocess per kernel; cancellation sends `SIGINT`. Details in `docs/python-repl.md`.
- Session state
  - `session.assertEvalExecutionAllowed?.()` can block execution.
  - `session.trackEvalExecution?.(...)` can register cancellable eval work.
  - `session.getSessionFile?.()`, `session.getEvalSessionId?.()`, and `session.getEvalKernelOwnerId?.()` influence VM/kernel reuse and artifact lookup.
  - JS VM contexts persist across eval calls until reset/disposal.
  - Python retained kernels persist until reset, owner cleanup, or process exit.
- User-visible prompts / interactive UI
  - none; stdin requests are rejected programmatically
- Background work / cancellation
  - Python retained kernels have heartbeat and idle cleanup timers.
  - Cancellation hard-kills/resets the shared executor for that backend: JS terminates the worker, Python sends SIGINT and may escalate to subprocess shutdown.

## Limits & Caps

- Per-cell timeout default: 30s (applied when `timeout` is omitted in `EvalTool.execute()`; clamped through `TOOL_TIMEOUTS.eval.default` in `packages/coding-agent/src/tools/tool-timeouts.ts`)
- Schema-level `timeout` range: integer `1..600` seconds (enforced by Zod on the cell schema)
- Timeout clamp at runtime: 1s minimum, 600s maximum (`TOOL_TIMEOUTS.eval` in `packages/coding-agent/src/tools/tool-timeouts.ts`)
- Transcript code/output preview: 10 lines by default (`EVAL_DEFAULT_PREVIEW_LINES` in `packages/coding-agent/src/tools/eval.ts`)
- Output truncation window: 50KB default (`DEFAULT_MAX_BYTES` in `packages/coding-agent/src/session/streaming-output.ts`)
- Output line cap inside truncation helpers: 3000 lines (`DEFAULT_MAX_LINES` in `packages/coding-agent/src/session/streaming-output.ts`)
- Streaming tail buffer for live updates: `DEFAULT_MAX_BYTES * 2` = 100KB (`packages/coding-agent/src/tools/eval.ts`)
- Python retained kernel idle timeout: 5 minutes (`IDLE_TIMEOUT_MS` in `packages/coding-agent/src/eval/py/executor.ts`)
- Python retained kernel cap: 4 sessions (`MAX_KERNEL_SESSIONS` in `packages/coding-agent/src/eval/py/executor.ts`)
- Python retained kernel cleanup sweep: every 30s (`CLEANUP_INTERVAL_MS` in `packages/coding-agent/src/eval/py/executor.ts`)
- Python owner-cleanup shutdown wait: 2000ms (`OWNER_CLEANUP_KERNEL_SHUTDOWN_TIMEOUT_MS` in `packages/coding-agent/src/eval/py/executor.ts`)
- Python heartbeat interval: 5s (`ensureKernelHeartbeat()` in `packages/coding-agent/src/eval/py/executor.ts`)
- Python external gateway availability check timeout: 5s (`AbortSignal.timeout(5000)` in `packages/coding-agent/src/eval/py/kernel.ts`)
- Python auto-restart budget: one restart per retained session before hard failure (`restartCount > 1` in `packages/coding-agent/src/eval/py/executor.ts`)

## Errors

- Zod validation rejects malformed `cells` arrays before `execute()` runs (missing `language`/`code`, out-of-range `timeout`, empty `cells`).
- Missing session without proxy executor throws `ToolError("Eval tool requires a session when not using proxy executor")`.
- Disabled/unavailable backends throw `ToolError` from `resolveBackend()`:
  - `eval.py = false` and a `py` cell is requested
  - `eval.js = false` and a `js` cell is requested
  - Python kernel unavailable and a `py` cell is requested
- JS runtime exceptions are converted into text output plus `exitCode: 1`; cancellations return `cancelled: true` and may append `Command timed out`.
- Python execution errors from the kernel become text output and `exitCode: 1`; later cells are skipped.
- Python stdin requests are treated as errors with the message `Kernel requested stdin; interactive input is not supported.`
- Cancellation is returned, not thrown, once backend execution has started. The tool formats it as a cell failure and sets `details.isError = true`.
- If output truncates, the tool still succeeds; truncation is surfaced through `details.meta` and artifact-backed full output when available.

## Shared executor trade-offs

- Parent agents and subagents share eval state bidirectionally when a subagent inherits the parent's executor id. Mutations in either direction are visible to the other participant.
- Async regions of concurrent runs can interleave. Synchronous JS still blocks the VM event loop; synchronous Python still contends on the GIL.
- Cancelling one run is destructive to the shared backend executor. This is intentional: JS worker termination and Python SIGINT/subprocess shutdown are the only reliable way to interrupt arbitrary user code.
- `reset: true` is destructive for every live run on that backend session id. New starts on that backend are rejected while reset is in flight.

## Notes

- Backend selection is now strictly explicit per cell: `language` must be `"py"` or `"js"`. The previous `*** Cell` header parser, the `eval.lark` constrained grammar, and the sniffer-based fallback have all been removed.
- `EvalTool.customFormat` no longer exists. Tool calls flow through the standard JSON schema; there is no Lark-constrained sampling path.
- `tool.<name>()` exists in both JS and Python. Python calls route through a per-run loopback bridge keyed by the current cell id.
- JS helper paths reject protocol URIs (`://`) in `resolvePath()`; the JS prelude is filesystem-only unless the code calls `tool.read(...)` or another tool explicitly.
- Python helper `output(...)` depends on `PI_SESSION_FILE`; it fails outside a session-backed run.
- `display()` can produce text and structured outputs from the same value; the renderer prefers markdown over `text/plain` when both exist.
- JS static imports are rewritten only at top level. Nested imports stay invalid and surface normal JS syntax/runtime errors.
- `EvalTool` is `concurrency = "exclusive"` within one agent session, but parent and subagent sessions can run eval concurrently when they share an inherited executor id.
- The tool description shown to the model is templated by backend availability (`getEvalToolDescription()`); if Python is unavailable, the prompt omits Python-specific instructions.