fix(agent): isolate MCP transport failures

[gloryfromca]

Summary

Prevent a failed MCP transport from cancelling the main agent turn during startup.

Keep each MCP transport and handshake in one lifecycle so SDK task-group failures unwind into ordinary per-server connection errors. Raven can then continue with the remaining MCP servers and enter the model loop.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

• Relevant tests pass locally

• Relevant lint / type checks pass locally

• User-facing docs or screenshots are updated when needed

• uv run pytest tests/test_sandbox_unit.py -x -q - 70 passed.

• uv run ruff check raven/agent/tools/mcp.py tests/test_sandbox_unit.py - passed.

• A real raven agent turn returned RAVEN_FIXED_OK after two configured MCP servers failed authentication.

• Docs and screenshots are not needed because no user-facing contract changed.

Risk

• Security impact considered
• Backward compatibility considered
• Rollback path is clear for risky changes

The change only alters failure isolation during MCP initialization. Roll back the single fix commit if transport lifecycle behavior regresses.

Related Issues

N/A

[gloryfromca]

Blocking: the required diff-coverage gate fails at 79.31%; see the inline note.

Reviewed current head 399a1dfd011d against github/main (b6f3818ae579). I covered AGENTS.md and the Runtime glossary, the complete diff, the AgentLoop startup and shutdown callers through close_mcp(), MCP 1.27.0 transport/session task-group behavior, relevant cancellation-fix history, backward compatibility, and whether tests were weakened. No other blocker or plain error survived refutation: the per-server nested stack unwinds a failed transport/session initialization before iteration continues, keeps successful sessions owned by the existing outer stack, and preserves the public function/config behavior.

Verification:

• uv run pytest tests/test_sandbox_unit.py -x -q - 70 passed.
• The new isolation regression fails against github/main and passes on this head.
• A stronger external-cancellation probe using an AnyIO task group propagated CancelledError correctly.
• uv run ruff check raven/agent/tools/mcp.py tests/test_sandbox_unit.py - passed.
• git diff --check and merge-tree checks - clean.
• GitHub CI executed 6901 tests successfully (33 skipped, 13 deselected), then failed the required changed-line coverage gate.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed current head 404c4cf4b755 against github/main and the previously reviewed 399a1dfd011d. The new commit is test-only and directly resolves the prior coverage blocker with focused assertions for the SSE client/header path and unknown-transport skip; it does not weaken, skip, or replace the transport-failure tests. I rechecked the repository rules and Runtime terms, the complete diff, the existing AgentLoop ownership/lifecycle callers, relevant history, and backward compatibility. There is no production-code delta from the prior review and no new plain error survived refutation.

Verification:

• uv run pytest tests/test_sandbox_unit.py -x -q --cov=raven --cov-branch --cov-report=json:/tmp/pr365-404c4cf4-coverage.json - 72 passed.
• Project diff-coverage gate - 100.00% (29/29), required threshold 90%.
• uv run ruff check raven/agent/tools/mcp.py tests/test_sandbox_unit.py - passed.
• git diff --check and merge-tree checks - clean.
• All GitHub checks are green, including the full unit and coverage job.