Skip to content

CTX7-3637: Benchmark Context7 after normal MCP setup - #3688

Open
fahreddinozcan wants to merge 3 commits into
CTX7-3596/direct-mcp-server-instructionsfrom
CTX7-3637/normal-install-benchmark
Open

fahreddinozcan wants to merge 3 commits into
CTX7-3596/direct-mcp-server-instructionsfrom
CTX7-3637/normal-install-benchmark

Conversation

@fahreddinozcan

@fahreddinozcan fahreddinozcan commented Nov 5, 2026 •

Copy link
Copy Markdown
Collaborator

Adds the Context7 MCP instruction benchmarks, including answer correctness and unnecessary lookups after the standard Claude Code install.

  • Keep the shared runner, case sets, tool-loading controls, and recorded results together.
  • Compare the old and new prompts with the installer-generated rule and skill present in both arms.
  • Record per-case lookup counts and both blinded correctness scores.
  • In the normal setup run, unnecessary lookups fell from 1/30 to 0/30. Neither judge found a gain in fully correct answers.
  • Stack on the server prompt change in CTX7-3596: Make the MCP server instructions direct so agents call Context7 more often #3665.

Closes CTX7-3637

@vercel

vercel Bot commented Nov 5, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment
Project Deployment Actions Updated
context7-admin-mcp Ignored Ignored Nov 10, 2026 9:15am UTC

Request Review

@linear-code

linear-code Bot commented Nov 5, 2026

Copy link
Copy Markdown

CTX7-3637

…n metadata

The runner archives a pinned commit as the fixture for each run and stops on a
fixture that lacks the files the cases name. Each case runs three times by
default, and the summary reports the rate per repeat and the unstable cases.
The report records the served instruction text hash, tool loading flags,
fixture hash, and eval file hashes.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant