test(mcp_server): add smoke test for tools/call deadlock regression

[Teerapat-Vatpitak]

Summary

• Add a regression test for #85 (the MCP tools/call deadlock fix).
• Spawns vibe-trading-mcp over stdio and verifies the JSON-RPC happy path: initialize + tools/list + tools/call analyze_options.
• No network dependency; total runtime ~4s with the fix in place.

Why

#85 fixes a deadlock where tools/call hangs indefinitely if the registry pre-warm in mcp_server.main() is removed. Without a regression test, a future contributor could revert that line during a refactor and reintroduce the bug — initialize and tools/list would still pass, so the breakage would be silent until users hit tools/call.

This test fails deterministically (15s timeout on tools/call) when the pre-warm is removed, with an assertion message that points back at #85.

What it covers

1. initialize — JSON-RPC handshake completes within 30s (covers cold imports).
2. tools/list — at least 20 tools registered, including analyze_options, get_market_data, and list_skills. Catches tool-registration regressions and missing package data.
3. tools/call analyze_options — pure CPU Black-Scholes calculation; must return within 15s. This is the call that deadlocks without the fix.
4. Result shape — price > 0, delta ∈ [0, 1]. Catches Black-Scholes formula regressions.

Differential validation

Verified locally on Windows + Python 3.12 + fastmcp from main:

Run	mcp_server.main() state	Result	Wall time
1	with pre-warm (current main)	PASS	3.91s
2	pre-warm commented out	FAIL — timeout 15s on tools/call	17.45s
3	pre-warm restored	PASS	3.81s

tools/list succeeds in all three runs — only tools/call exposes the deadlock, which is the exact symptom users hit (e.g. #42, #32).

Why no network in the test

analyze_options is a Black-Scholes calculation with no external service. Picking it over get_market_data keeps the test:

• Deterministic (no API rate-limits / outages)
• CI-friendly (no outbound network needed)
• Fast (sub-second after init)

Test plan

• pytest agent/tests/test_mcp_server_smoke.py -v passes locally (~4s)
• Test fails with a timeout when the pre-warm in mcp_server.main() is removed
• Restoring the pre-warm makes the test pass again
• Test only spawns one subprocess and cleans it up in finally; no shared state with other tests

Notes for reviewer

Tested locally on Windows + Python 3.12 with the differential validation table above. CI will be the first run on Linux + Python 3.11. The test uses only portable APIs (subprocess.Popen, pathlib.Path, os.pathsep, sys.executable, threading), so cross-platform issues are unlikely — but I'm happy to follow up with fixes if anything surfaces in CI.

This is also the first subprocess-spawning test in this repo (existing tests use direct imports / FastAPI TestClient). I picked subprocess because the deadlock only manifests across the FastMCP server thread boundary, which is hard to mock faithfully. Happy to refactor toward a fixture-based / mocked approach if that fits the repo's testing philosophy better.

Checklist

• No changes to protected areas (agent/src/agent/, agent/src/session/, agent/src/providers/) — only adds a test file under agent/tests/.
• No hardcoded secrets / paths; uses sys.executable and resolves repo root from __file__.
• Code follows CONTRIBUTING.md (Conventional Commit title, @pytest.mark.integration, Google-style docstrings).
• Documentation updated — N/A (test file is self-documenting).

[warren618]

Thanks for adding this MCP smoke test. I verified the test locally as well, and it is useful for guarding against the tools/call deadlock regression.

One small note for future PRs: since this is an integration-style subprocess test, we may later tune the default pytest selection if the suite gets slower. For now this looks good.

[Teerapat-Vatpitak]

Thanks for the local verify and the heads-up on integration-style test categorization — noted for future PRs (likely will add a @pytest.mark.integration marker if I introduce more subprocess tests, so default pytest runs stay fast).