[dodeja]

PR #384's source inventory recognizes server.registerTool, while its offline protocol smoke accepts any 14 tools containing search_docs and whoami. A registration through an alias can escape the source check, and a same-count substitution can still pass that smoke.

Compare the actual tools/list names with the existing feature map in the smoke and the existing modern/legacy protocol tests. This removes another hand-maintained count and checks the public inventory regardless of registration syntax. The static implementation-source check remains in place.

Stacked on #384. The diff is limited to the protocol smoke and its existing tests.

Validation:

• npm run agent-verify passes all 7 steps, including builds, lint, TypeScript, 52 guard tests, 47 policy tests, 153 CLI tests, 389 SDK tests, 470 MCP tests, generation checks, and offline protocol discovery. The existing live SDK smoke suite remains skipped.
• A controlled same-count substitution retains an unreachable original registration and adds an aliased replacement. The original static guard and original smoke pass; the updated smoke and all six protocol cases reject the changed inventory. Updating the map to match passes the new smoke. Original source/map were then restored and verification passed.
• The sandbox rejected localhost listeners with EPERM. The unchanged Sentry tests and full gate pass with loopback access. No test settings or credentials were changed.

This verifies the exercised tool inventory. It does not establish live API behavior or every runtime configuration.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

Final-head follow-through

Head a201b1f11389891cc5dbf6d92bb02b89d2412d42 was verified again after review inspection. The full agent-verify gate, MCP coverage test, eval type-check, generated-filter comparison, MCP/API lint, coverage generators, and the six protocol smoke versions all passed locally in the isolated checkout with loopback access. The checkout remains clean.

Normal hosted CI did not run because its branch filter excludes this stacked base. The relevant offline CI commands were run locally without changing the workflow or its security settings. This is local verification, not a hosted-CI pass. Authenticated preview/live checks and the Node 20/22 consumer matrix were not run.

The hosted Claude review wrapper reports success, but its log records an authentication failure: 401 OAuth access token is invalid. The actual review is blocked, and no review findings were produced. No credentials were changed and that failed action was not retried. The existing Vercel preview integration and Preview Comments completed successfully; no manual deployment was requested.

The PR remains draft and stacked on #384.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

The PR checks advertised MCP tool names against skills/agent-trust/feature-map.json instead of a fixed count.

• Protocol checks compare listed tool names with the feature map.

_{Reviews (1) · Last reviewed commit: "fix(ci): compare advertised MCP tools wi..." · Reviewed by Greptile}

[vercel]

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

Project	Deployment	Actions	Updated
api	Ready	Preview	Oct 17, 2026 8:32am UTC

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-19T02:05:03.881399Z	a201b1f	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.