[NielsRogge]

Summary

• Vendors huggingface/pwc-cli mcp_server/SKILL.md at a2fc9d0 (Use the public-metadata API for papers outside the catalog pwc-cli#28, pwc-mcp 0.2.7) into backend/chat_sandbox_worker/PWC_MCP_SKILL.md, and updates the pin comment in mcp_research_prompt.ts.
• New skill text: papers not indexed on Papers With Code return in_catalog: false, a note, and available_at links (via Serve public metadata for papers outside the catalog #508's public-metadata endpoint), and the agent must cite those sources rather than a PwC paper page. It also brings the 0.2.1–0.2.7 tool docs (hybrid default search, paginated get_benchmark, search on task/method lists, etc.).
• Additional change: the new skill tells the agent to use get_paper_evaluations. The live MCP server already exposes that tool (confirmed with tools/list on paperswithcode.co/mcp), but the chat allowlist didn't include it, so the prompt would have named a tool the chat can't call. This PR adds it to the gateway TOOL_NAMES and Pi PWC_MCP_TOOLS (20 → 21 tools), maps pwc paper evaluations to it, and updates the catalog-usage prompt and docs/chat-pwc-mcp.md. Because discovery fails closed, this must not ship to an environment whose MCP server lacks get_paper_evaluations; production already has it.
• No sync workflow or hash check pins this file. sync-pwc-cli-skill.yml only syncs PWC_CLI_SKILL.md, and test_pi_prompt_assets.py compares the published file against the vendored one, not against a fixed hash.

Tests

• pytest backend/tests/chat_gateway/test_pwc_mcp.py backend/tests/test_pi_prompt_assets.py backend/tests/chat_sandbox/test_prompts_and_catalogs.py: 29 passed.
• On the wider set (chat_gateway/, external_pi/test_runtime.py, paper_edits/test_policy.py), 32 tests fail in test_turn_fence.py and external_pi/test_runtime.py. The same set fails on a clean origin/main checkout in this environment.
• pi_worker npm test: 199 pass. 2 tests fail ("native MCP …" web search), and they fail the same way on origin/main.
• I checked by hand that adaptMcpSkill renders pwc_get_paper_evaluations and the available_at guidance.

Chat-related: do not merge or deploy without explicit approval.

🤖 Generated with Claude Code

---

Note

Medium Risk
Expands the fixed MCP allowlist and prompt contract; discovery fails closed if an environment’s MCP server lacks get_paper_evaluations, and off-catalog citation rules change agent behavior for missing papers.

Overview
Updates the vendored PwC MCP skill to pwc-mcp v0.2.7 (new upstream pin in mcp_research_prompt.ts) so chat prompts match current tool behavior: hybrid default search, paginated get_benchmark, search on task/method lists, softer handling of unknown leaderboard metrics, and richer paper lookup (short titles, off-catalog in_catalog: false with available_at links instead of PwC paper URLs).

get_paper_evaluations is added to the gateway and Pi tool allowlists (20 → 21 tools) with CLI mapping and catalog-usage text, so agents are not instructed to call a tool discovery would reject. Prompt adaptation drops upstream guidance to pass "next" before the first read_paper read, since chat resolves cursors differently.

Docs and tests reflect the new tool count and prompt strings (available_at, pwc_get_paper_evaluations).

^{Reviewed by Cursor Bugbot for commit 60b5700. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 76df4c5. Configure here.}