[PR #1461] [MERGED] fix(mcp): validate doctor probe against its discovered output contract #2902

New issue

Closed

opened 2026-10-17 21:11:36 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-17 21:11:36 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/morluto/rea/pull/1461
Author: @N0zoM1z0
Created: 10/10/2026
Status: ✅ Merged
Merged: 10/10/2026
Merged by: @N0zoM1z0

Base: main ← Head: fix/mcp-doctor-discovered-output-validation

📝 Commits (1)

946bde7 fix(mcp): validate doctor probe against its discovered output contract

📊 Changes

3 files changed (+144 additions, -7 deletions)

View changed files

📝 src/mcpDoctor.ts (+20 -6)
📝 tests/boundary/mcp/mcpDoctor.test.ts (+76 -1)
➕ tests/fixtures/mcpDoctorReply.mjs (+48 -0)

📄 Description

The MCP doctor could report healthy: true for a binary_session reply that violated the output schema advertised by the same production server. Discover the complete tool/prompt inventory first, select the probe definition by name, and pass that definition to the SDK call. Keep the absolute deadline and owned-child cleanup.

A rejected probe now reports the request/schema failure while retaining the already observed server and inventory, rather than replacing them with an undifferentiated transport failure. The successful report and default Evidence contract remain unchanged.

Validation:

A committed stdio replay regression captures fresh metadata and a valid reply from the actual production server, then checks both valid and schema-invalid replies. The invalid case fails on the original main implementation and passes with this fix.
All nine doctor/CLI-help tests, full serial typecheck, lint, duplication gate and changed-file formatting pass under the documented two-CPU and process-family budget.
The actual public CLI doctor report is byte-identical (1,593 bytes). A 74-byte JavaScript target produces identical default Evidence to main through CLI and SDK stdio MCP; text/structured parity, source digest and subsequent ping pass.
Independent fault injection at the actual production stdio producer confirms main accepts an extra output property and the fix rejects it. This exercises the production client/server boundary; it does not certify a deep-analysis provider.

This is a diagnostic correctness fix. Interleaved cold CLI timings did not establish a speedup, and no model-token saving is claimed. The SDK internal cause of the original default-path behavior remains unverified.

Closes #1454. Refs #1065, #723, #1433.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/morluto/rea/pull/1461 **Author:** [@N0zoM1z0](https://github.com/N0zoM1z0) **Created:** 10/10/2026 **Status:** ✅ Merged **Merged:** 10/10/2026 **Merged by:** [@N0zoM1z0](https://github.com/N0zoM1z0) **Base:** `main` ← **Head:** `fix/mcp-doctor-discovered-output-validation` --- ### 📝 Commits (1) - [`946bde7`](https://github.com/morluto/rea/commit/946bde7dd8091da53cd645d7e80dd83d42d0e87f) fix(mcp): validate doctor probe against its discovered output contract ### 📊 Changes **3 files changed** (+144 additions, -7 deletions) <details> <summary>View changed files</summary> 📝 `src/mcpDoctor.ts` (+20 -6) 📝 `tests/boundary/mcp/mcpDoctor.test.ts` (+76 -1) ➕ `tests/fixtures/mcpDoctorReply.mjs` (+48 -0) </details> ### 📄 Description The MCP doctor could report `healthy: true` for a `binary_session` reply that violated the output schema advertised by the same production server. Discover the complete tool/prompt inventory first, select the probe definition by name, and pass that definition to the SDK call. Keep the absolute deadline and owned-child cleanup. A rejected probe now reports the request/schema failure while retaining the already observed server and inventory, rather than replacing them with an undifferentiated transport failure. The successful report and default Evidence contract remain unchanged. Validation: - A committed stdio replay regression captures fresh metadata and a valid reply from the actual production server, then checks both valid and schema-invalid replies. The invalid case fails on the original main implementation and passes with this fix. - All nine doctor/CLI-help tests, full serial typecheck, lint, duplication gate and changed-file formatting pass under the documented two-CPU and process-family budget. - The actual public CLI doctor report is byte-identical (1,593 bytes). A 74-byte JavaScript target produces identical default Evidence to main through CLI and SDK stdio MCP; text/structured parity, source digest and subsequent ping pass. - Independent fault injection at the actual production stdio producer confirms main accepts an extra output property and the fix rejects it. This exercises the production client/server boundary; it does not certify a deep-analysis provider. This is a diagnostic correctness fix. Interleaved cold CLI timings did not establish a speedup, and no model-token saving is claimed. The SDK internal cause of the original default-path behavior remains unverified. Closes #1454. Refs #1065, #723, #1433. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-17 21:11:36 +00:00

closed this issue
added the
pull-request
label

bsakaguchi referenced this issue

2026-10-31 01:07:00 +00:00

[PR #2902] [MERGED] test(windows): verify retained inspection ownership after rename #5482

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

morluto/rea#2902

No description provided.

Rows
Columns