Fix local Echo example MCP usage

[burtenshaw]

Updates the local Echo example to use the supported MCP client methods and sync wrapper against a running Echo server.

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[cursor]

Alignment Review Report

Reviewed the migration of examples/local_echo_env.py from the removed EchoAction API to the MCP tool-calling API. I verified everything end-to-end against a live server (python -m echo_env.server.app).

Automated Checks

• Lint: PASS (this PR). examples/local_echo_env.py passes ruff check, ruff format --check, and usort check, and byte-compiles. The lint.sh hook exits non-zero, but every file it flags is pre-existing formatting drift in unrelated envs (opencode_env, chat_env, repl_env, textarena_env, …) — none is this PR's file. Note examples/ is outside both CI (src/ tests/) and the hook's scope, so I checked the changed file directly.
• Debug code: CLEAN. check-debug.sh only scans src/; all hits are pre-existing there. The print() calls in this file are intentional demo output.

Open RFCs Context

• RFC 003 — MCP Support (In Review; @Darktex, @pankit-eng): directly relevant, and this change aligns with it — the example now drives the env purely through MCP tools (list_tools() / call_tool()), i.e. the "MCP as universal standard" direction. No conflict.
• RFC 010 ("ECHO") is env-token world-modeling, unrelated to echo_env.

Tier 1: Fixes Required

None. Verified against a running server:

• The old example was broken at import time — from echo_env import EchoAction raises ImportError (echo_env exports only EchoEnv, CallToolAction, ListToolsAction). This PR fixes a genuinely non-running example.
• The new example runs cleanly (exit 0). Output confirmed correct: reset → status='ready', message='Echo environment ready!', reward=0.0, done=False; both tools listed with descriptions; echoes correct; lengths 13 / 24 / 16.

Tier 2: Alignment Discussion

Principle Conflicts

None identified. Client-only example (imports EchoEnv, never server/) → client-server separation intact. Uses the Gym-style reset() plus MCP tools; no external reward computation introduced.

RFC Conflicts

None identified (aligns with RFC 003, above).

Summary

• 0 mechanical issues to fix
• 0 alignment points for human review
• 0 RFC conflicts

Nice cleanup. Beyond fixing the import, the migration also sidesteps two known echo_env gotchas: it reads results via call_tool (which unwraps the FastMCP envelope to .data) rather than the raw observation.result, and it drops the per-tool-call reward print (tool steps return reward=None, not 0.0 — only reset() is 0.0).

Optional, non-blocking (out of scope here): a few other spots still use the pre-MCP echo observation shape (echoed_message / message_length) — notably src/openenv/core/containers/test_local_docker_provider.py (already TODO-marked for removal and Docker-gated) and the env-generation templates under src/openenv/cli/templates/openenv_env/ and .claude/skills/generate-openenv-env/assets/. Flagging only for awareness.

  

_{Sent by Cursor Automation: Pre-review}

[cursor]

Verified correct against a running server: EchoEnvironment.reset() returns Observation(reward=0.0, metadata={'status': 'ready', 'message': 'Echo environment ready!'}), so both metadata.get(...) lookups resolve as intended (printed Status: ready / Message: Echo environment ready!).

[cursor]

Good use of the call_tool convenience method — it unwraps the FastMCP result envelope to .data, so echoed is the plain string and with_length (line 67) is the plain {'message': ..., 'length': N} dict (verified: lengths 13/24/16). This avoids the raw observation.result envelope gotcha, and dropping the per-call reward print is correct since MCP tool steps return reward=None (only reset() is 0.0).