claude-pty-web-harness 0.5.0

Python backend parity for claude-pty-web-harness: drive Claude Code in a pupptyeer pty and stream its JSONL transcript over the same HTTP/WS protocol as the TS server.

claude-pty-web-harness (Python backend)

A Python parity of the backend: drive Claude Code in a pupptyeer pty and stream its JSONL transcript over the same HTTP/WS protocol as the TS server, so the existing React frontend (packages/app) works against it unchanged.

Mirrors the TS packages:

Python module	TS equivalent	Role
claude_pty_web_harness.protocol	@…/protocol	wire types (ChatEvent, summaries, messages)
claude_pty_web_harness.harness (ClaudeHarness)	@…/core	transport-agnostic logic (pupptyeer + JSONL + daemon-rendered modal/readiness detection)
claude_pty_web_harness.{detect,jsonl}	core/src/*	supporting modules
claude_pty_web_harness.server	@…/server	reference FastAPI + WebSocket adapter

It wraps the stdlib-only pupptyeer Python client, installed from PyPI as pupptyeer-client (a declared dependency). For local development against a pupptyeer checkout, set PUPPTYEER_PY_CLIENT to its clients/python directory to import that copy instead.

The pupptyeer client is synchronous (background reader thread); the harness is asyncio and runs its request/reply calls in a thread executor so the FastAPI event loop never blocks.

Run

cd packages/python
uv venv && uv pip install -e .
PORT=4318 uv run uvicorn claude_pty_web_harness.server:app
# or: PORT=4318 uv run python -m claude_pty_web_harness.server

It serves the identical API on :4318, so npm run dev:app (the React app, Vite proxy to :4318) talks to it with no changes. Run either the TS server or this one, not both (same port).

Env: PORT, HOST, PUPPTYEER_SOCK (non-default daemon socket), READINESS=delay (fallback for daemons without working capture), PUPPTYEER_PY_CLIENT (dev only: import a pupptyeer client checkout instead of the installed pupptyeer-client).

Reusing the core (any transport)

from claude_pty_web_harness import ClaudeHarness

harness = await ClaudeHarness.create(
    allowed_roots=["/home/me/dev"],  # optional: reject create_session outside these
    # socket_path=..., readiness="screen" | "delay"
)
remove = harness.add_listener(lambda kind, sid, payload: ...)  # "chat" | "status"
summary = await harness.create_session(cwd="/repo", model="sonnet")
await harness.send_prompt(summary["id"], "first line\nsecond line")  # multi-line, one paste
# also: harness.list(), harness.get(id), harness.transcript(id),
#       await harness.interrupt(id), await harness.kill(id)

create_session also takes command, permission_mode, extra_args, and env (merged over the daemon's own environment for the spawned process; see USAGE.md), plus cols/rows. send_prompt delivers the text as a bracketed paste so multi-line input lands in the TUI intact, then submits with one Enter (pass submit=False to stage it).

send_prompt captures the screen once before writing anything (in readiness="screen" mode only) and raises PickerOpenError (code = "picker_open", exported from this package) if a numbered picker - an AskUserQuestion prompt, a tool-permission prompt, and the trust modal all render identically, so it cannot say which one - is open: the trailing Enter that submits a prompt would otherwise confirm whichever option is highlighted. It fails open on a capture timeout, and it still leaves a narrow race between the capture and the Enter, so treat it as reducing collisions, not eliminating them. A caller that already knows what is on screen can pass force=True to skip the check; it is not available over the reference HTTP/WS server.

Permission modes (read this before exposing it)

Sessions default to --permission-mode bypassPermissions, which approves every tool call with no checks: the agent can run arbitrary commands and edit any file inside the session's cwd without prompting. The default favors low friction and is not safe for untrusted input. Contain it with allowed_roots and the server auth, and only point it at directories you trust.

Because the harness drives the real claude TUI in a pty, it can use claude's interactive auto mode (a classifier vets each action), which is the best overall balance for a web-driven agent and works here precisely because we own the pty:

await harness.create_session(cwd="/repo", permission_mode="auto")

permission_mode is passed straight through to claude --permission-mode, so default, plan, acceptEdits, auto, and bypassPermissions all work. Auto mode needs a recent claude CLI and an eligible model; if it is unavailable claude falls back, so confirm it actually engaged.

Mounting into your own FastAPI app (with auth)

include_harness_routes mounts the same REST + WS endpoints onto an existing app, so the harness can sit behind your app's auth instead of running open on localhost. REST routes accept FastAPI dependencies; the WebSocket is guarded separately (browsers can't set an Authorization header on a WS).

from fastapi import Depends, WebSocket
from claude_pty_web_harness import ClaudeHarness
from claude_pty_web_harness.server import include_harness_routes

harness = await ClaudeHarness.create(allowed_roots=["/home/me/dev"])

async def authenticate_ws(ws: WebSocket) -> bool:
    return await verify_ticket(ws.query_params.get("ticket"))  # your check

include_harness_routes(
    app,
    harness,                                 # or a zero-arg callable returning one
    prefix="/api/chat",
    dependencies=[Depends(validate_token)],  # guards the REST routes
    authenticate_ws=authenticate_ws,         # guards the WS upgrade
)

/health is left unauthenticated for liveness probes. Keep the harness itself bound to localhost / a private interface and let your app be the only authenticated front door; never expose it directly. create_router(...) returns the APIRouter if you'd rather include it yourself.

Requires

A running pupptyeer daemon (install the @petersr/pupptyeer binary and pupptyeer daemon install). The harness connects to it and fails loud if it is unreachable; it never spawns one. Restart the daemon after upgrading pupptyeer (pupptyeer daemon restart), else captureScreen can return an empty grid and readiness never fires (chat still works via JSONL); or run with READINESS=delay.