Pyodide Phase 3: edit the agent brain in the browser and re-run

[rsasaki0109]

What

Follow-up item ② Phase 3 from docs/pyodide_playground_strategy.md — the "edit the robot's brain in the browser" moment.

The pick_and_retry scenario now shows an "Agent brain" code editor pre-filled with the real PickAndRetryAgent source (fetched via inspect.getsource, so it can never drift from the file). Edit it, click Run edited agent, and the playground execs your class in Pyodide and runs it against the real Tabletop2D via the example's own run_agent(...) — the same loop the CLI uses, so there is no second loop to drift. Syntax/runtime errors surface inline.

Changes

• examples/manipulation/01_pick_and_retry.py — extract run_agent(agent, ...); run() delegates to it. run_agent reads belief attributes defensively, so a custom agent that changes or drops them still runs and serializes.
• docs/playground.{html,js,css} — editor panel (shown for pick_and_retry only), "Run edited agent" / "Reset code" buttons, inline error status. USER_SRC is injected via pyodide.globals (no string-escaping); the source-fetch driver registers the module in sys.modules so inspect.getsource works; the editable header re-adds from __future__ import annotations + typing.Any so the class execs standalone.
• tests/test_playground_trace.py — run_agent tolerates a custom (belief-less) agent and still produces valid, plain-JSON config.

Safety

The edited code runs entirely client-side — Pyodide is a WASM sandbox in the visitor's own browser, so exec'ing it is no more privileged than a local REPL. Nothing touches a server.

Verification

• Full suite green (122 tests).
• Verified the exact drivers in an unpacked-bundle sim: the default source reproduces seed=3 (4 steps, retries=2); removing the deliberate first offset makes the agent grab the belief mean immediately (2 steps, retries=0) — a live lesson in why the retry schedule exists; malformed code raises a caught error.

Still pending (browser, your side)

Pick "Pick and retry (real Python)", confirm the editor populates with the real source, edit offset_schedule / act(), and confirm Run edited agent re-runs and inline errors render.

🤖 Generated with Claude Code