[Chore] Unskip VS Code e2e replay for use_mcp_tool

[roomote]

Opened on behalf of Elliott de Launay. View the task or mention @roomote for follow-up asks.

What problem this solves

Fixes an issue where the use_mcp_tool VS Code e2e suite still stayed skipped, which left MCP filesystem delegation unverified in the deterministic replay flow.

Why this change was made

This slice rewrites the skipped MCP suite around stable fixture tags embedded in real prompts, dynamic fixture registration, and a real filesystem MCP server wired into the temporary test workspace. It also pins the filesystem MCP server package and keeps the error-path assertions focused on MCP behavior instead of brittle assistant phrasing.

User impact

No direct product behavior changes. The MCP e2e coverage becomes replayable and helps catch regressions in the MCP delegation path earlier.

Related PRs

• Read-only tools
• Mutating tools
• Subtasks

[roomote]

No actionable issues found. See task

• use-mcp-tool.test.ts no longer hard-codes requested file in the read_file completion assertion, so the suite no longer depends on one assistant phrase after the MCP read succeeds.
• The unknown-server case no longer assumes partial MCP approval messages keep serverName and toolName empty before validation fails.
• The suite pins @modelcontextprotocol/server-filesystem, so the deterministic replay path no longer floats with npm publishes.
• The prior pending Windows state is superseded by a fresh CI run on the latest head.
• use-mcp-tool.test.ts now keeps the stable USE_MCP_TOOL_*_SMOKE tags inside real natural-language prompts, so test:ci and test:record still exercise MCP workflows outside replay mode.
• The new GET_FILE_INFO_PROMPT now accepts either the relative fixture path or the absolute workspace path, so live-model runs no longer fail when get_file_info chooses the prompted relative path.
• runMcpTask() no longer keeps only the first use_mcp_server ask, so live MCP runs can wait for the later complete request instead of capturing partial payloads with blank toolName or incomplete arguments.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[roomote]

I see some merge conflicts here. Working on them now...

[roomote]

I see some merge conflicts here. Working on them now...

[roomote]

Resolved merge conflicts in:

• apps/vscode-e2e/src/fixtures/apply-diff.ts
• apps/vscode-e2e/src/fixtures/execute-command.ts
• apps/vscode-e2e/src/fixtures/write-to-file.ts
• apps/vscode-e2e/src/runTest.ts

Warnings:

• A direct pnpm --filter @roo-code/vscode-e2e test:run attempt still stops in unrelated existing e2e suite compilation errors around @roo-code/types module resolution before the touched mock suites execute.
• The resolved branch was still validated with git diff origin/main --check, pnpm --filter @roo-code/vscode-e2e check-types, the repo lint hook during git merge --continue, and the repo check-types hook during git push.

[roomote]

Addressed the remaining PR feedback on 77052fc22.

• Replaced the bare USE_MCP_TOOL_*_SMOKE task text with real MCP instructions while keeping the stable fixture tags embedded for replay-mode matching.
• Relaxed the read_file completion assertion so it no longer depends on the assistant literally saying requested file.
• The earlier fixes on this PR remain in place: the filesystem MCP server package is pinned, the unknown-server case no longer depends on empty partial fields, the get_file_info completion check is wording-agnostic, and the stray fixture-only churn is gone.

Validation:

• The push hook pnpm check-types passed.
• TEST_FILE=use-mcp-tool.test pnpm --dir apps/vscode-e2e test:ci:mock passed locally with all 6 MCP tests green.
• pnpm --dir apps/web-roo-code test passed.
• curl -I http://127.0.0.1:3000 returned 200 OK.

Browser proof is not applicable here because this is test-only e2e fixture work.

[edelauna]

Unskipping existing e2e tests to increase coverage.