[SaladDay]

Claude Code now admits the structured output and tool search combinations its native SDK supports.

• Structured output (text.format json_schema) works with multi_agent, environment Skills, Plugins and capability directories, MCP and tool_search. Natively the StructuredOutput tool is always added; the blocker was the bridge's exact tool-inventory check, which now counts the native StructuredOutput and ToolSearch tools beside the declared MCP inventory.
• tool_search works with MCP, Skills, Plugins and structured output. The MCP profile's allowed tools now include ToolSearch. tool_search with multi_agent stays rejected while Subagents cannot use function tools, which tool search needs.
• The bridge no longer rejects tool search for some providers, beta settings or claude-3(-5)-haiku model names; the native CLI decides that itself. ENABLE_TOOL_SEARCH=true stays as the mapping of the official tool.
• MCP allowed_tools names are no longer limited by a regex; native normalizes any name, and the adapter still rejects collisions.
• The deprecated, natively ignored Agent mode input is no longer denied.

Docs: harness-capabilities.md, execution-tools.md and the bridge README, with zh mirrors.

Tests: declaration tests for each admitted combination, bridge tests (181), and a gated live test against MiniMax-M3 (Anthropic-compatible): structured output with multi_agent, with a function tool, with MCP, and with tool_search and a deferred function each returned schema-valid JSON.

---

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