Add MCP provider feature discovery

[kongjiadongyuan]

Summary

Follow-up to #835 and #909: add provider feature discovery to the orchestration MCP surface.

• add list_provider_features for draft agent configurations
• reuse the existing AgentManager.listDraftFeatures plumbing already used by the app/WebSocket feature picker
• return provider-owned AgentFeature[] values so orchestrators can discover feature IDs like Codex fast_mode before setting them
• update the MCP docs and Paseo skill guidance to say agents should query features instead of guessing feature IDs

This complements #909: #909 lets orchestrators set provider features, and this PR lets them discover which features are valid for a draft agent config.

Design note

I kept discovery as a separate MCP tool because it mirrors the existing app/WebSocket feature picker path and avoids overloading provider/model listing. The caller passes a draft config (provider, cwd, optional model/modeId/thinkingOptionId/featureValues) and gets back the provider-owned AgentFeature[] shape.

Tests

• npx vitest run packages/server/src/server/agent/mcp-server.test.ts --bail=1
• npx vitest run packages/server/src/server/agent/mcp-parity.e2e.test.ts -t "list_provider_features" --bail=1
• npm run lint
• npm run typecheck
• npm run format:check

[kongjiadongyuan]

Added real MCP HTTP end-to-end coverage for provider feature discovery. The new test calls list_provider_features through the test daemon via StreamableHTTP MCP client and verifies the provider feature value comes back from a draft session; it is not a direct registered-tool handler test.\n\nWhile making the MCP parity e2e run cleanly, I also fixed/covered stale worktree list cache behavior after MCP create_worktree/archive_worktree.\n\nVerification run locally:\n- npx vitest run packages/server/src/server/agent/mcp-server.test.ts --bail=1\n- npx vitest run packages/server/src/server/agent/mcp-parity.e2e.test.ts --bail=1\n- npm run lint\n- npm run typecheck\n- npm run format:check\n\nCommit: 136a4f8

[kongjiadongyuan]

Follow-up cleanup: I moved the unrelated MCP worktree list-cache fix into the dedicated small PR #911. This PR now stays focused on provider feature discovery.\n\nRelevant verification after cleanup:\n- npx vitest run packages/server/src/server/agent/mcp-server.test.ts --bail=1\n- npx vitest run packages/server/src/server/agent/mcp-parity.e2e.test.ts -t "list_provider_features" --bail=1\n- npm run lint\n- npm run typecheck\n- npm run format:check

[kongjiadongyuan]

Small design note for review: I kept discovery as a separate list_provider_features MCP tool because it mirrors the existing app/WebSocket feature picker path (AgentManager.listDraftFeatures) and keeps provider/model listing stable. The caller provides a draft config (provider, cwd, optional model/modeId/thinkingOptionId/featureValues) and gets back the provider-owned AgentFeature[] shape, so orchestrators can discover IDs like Codex fast_mode instead of guessing them.

This is meant to complement #909: #909 sets provider features, this PR discovers which ones are valid for a draft agent config.