fix(mcp): preserve distinct structured tool results

[RealKai42]

Related Issue

Related to #2554 and follows up on #3234's handling of duplicate MCP output.

Problem

An MCP tool can return a short summary in content and its records in structuredContent. The current adapter drops the records whenever content contains any usable text or media: a response saying “Found 1 row.” reaches the model without the row itself. The dropped data never reaches the tool-result spill pipeline.

What changed

• Preserve structured results alongside summaries, human-readable reports, media, and conflicting text values. Omit the structured copy only when a complete text block parses to the same JSON value; whitespace and object-key order do not prevent deduplication.
• Compare numeric values conservatively. Noncanonical number spellings retain both representations instead of allowing JSON parsing to round distinct values into an apparent match.
• Escape < in serialized extras instead of deleting closing-tag strings, so decoded field values remain unchanged.
• Update existing unit and real stdio-server tests, add an executor/spill/Read regression that retrieves all 1,200 records without repeating the MCP call, and document the behavior in both locales.

Human-readable renderings can overlap with structured data; they are retained because the adapter cannot establish that they preserve every field and type. Existing global spill retention limits and persistence-failure behavior remain outside this change.

Validation

• CI follow-up: pnpm --filter @moonshot-ai/agent-core-v2 test src/human/test/llm/thinking.test.ts — 15 tests passed after correcting an existing test listener from llm.delta to llm.streaming.part. The requester already emits the latter; production behavior is unchanged.
• pnpm --filter @moonshot-ai/agent-core-v2 test test/agent/mcp test/mcpCore test/agent/toolExecutor test/agent/toolResultTruncation test/os/backends/node-local/tools/read.test.ts --disableConsoleIntercept --maxWorkers=4 — 419 tests passed across 15 files.
• After replacing the touched fixture's project identifiers with neutral examples, pnpm --filter @moonshot-ai/agent-core-v2 test test/agent/mcp/output.test.ts — 68 tests passed.
• pnpm --filter @moonshot-ai/agent-core-v2 typecheck — passed.
• pnpm --filter @moonshot-ai/agent-core-v2 lint:imports — passed.
• pnpm lint — passed with warnings and no errors.
• pnpm --dir docs run build — passed.
• Read-only pre-PR audit found no new internal identifiers, credentials, unrelated files, or blocking findings.

Checklist

• I have read the CONTRIBUTING document.
• I have linked the related issue and explained the problem.
• I have added tests that prove the behavior works.
• Ran gen-changesets guidance and included a CLI patch changeset.
• Reviewed gen-docs guidance and synchronized the affected documentation with translate-docs.

[changeset-bot]

🦋 Changeset detected

Latest commit: f88f2c7

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[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 2026-09-08T13:10:58.725821Z	7bbf860	PR opened
🔒 Security Review	✅ Completed 2026-09-08T13:15:17.883568Z	7bbf860	PR opened

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.

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@f88f2c7

npx https://pkg.pr.new/@moonshot-ai/kimi-code@f88f2c7

commit: f88f2c7