feat(agent-core-v2): support per-server deferred disclosure of MCP tools

[sailist]

Related Issue

No linked issue — internal feature request (per-server control over deferred MCP tool disclosure).

Problem

The experimental progressive tool disclosure (tool-select flag + a model that declares dynamically_loaded_tools) defers every MCP server's tools: none of them appear in the top-level tools[], and the model must load each one on demand via select_tools. There is no way to pin a specific server's tools inline — for example a core server whose tools the model should see without a select_tools round-trip, while the rest stay deferred to preserve the prompt cache.

What changed

• MCP server entries (all three transports, plus session-ephemeral and plugin-declared servers) accept a new optional deferred boolean. With the tool-select gate active, servers default to deferred as before; deferred: false keeps that server's tools inline in the top-level tool list. The field is ignored while the gate is off, and v1 config parsing ignores the unknown key as well.
• The resolved server view now carries the flag to the agent MCP adapter, which registers every MCP tool with an explicit disclosure — deferred by default, inline for opted-out servers. The synthetic OAuth authenticate tool stays deferred, matching current behavior.
• Tool selection now keys purely on the registered disclosure instead of hard-coding source === 'mcp' as dynamically loadable. Default behavior is unchanged; the deferral decision is now an explicit registration-time declaration.
• Tests: resolved-view passthrough, inline registration for deferred: false servers, and shaping/manifest behavior for inline tools; existing fixtures updated to the explicit-disclosure registration contract. Docs (en/zh) document the field; changeset included.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue (no issue — internal request, see Problem).
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.

[changeset-bot]

🦋 Changeset detected

Latest commit: fc3b658

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

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@fc3b658

npx https://pkg.pr.new/@moonshot-ai/kimi-code@fc3b658

commit: fc3b658

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b18d344bdf

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Mirror deferred in the public client schemas

When callers configure an MCP server through @moonshot-ai/klient or the v2 node SDK, this option cannot reach the new engine field: packages/klient/src/contract/mcp.ts omits deferred, and Klient's Zod input/output parsing strips unknown properties before sending or returning configs; the node SDK also routes these operations through that contract. Consequently session-ephemeral, add/update, and test calls silently revert to the default deferred behavior, while reading and subsequently updating an existing deferred: false config can erase the setting. Add the field to the mirrored client schema/types and preserve it through node SDK validation.

AGENTS.md reference: AGENTS.md:L30-L30

Useful? React with 👍 / 👎.