feat(0.22.1): mcpToolsForRuntimeMcp() — project MCP tools to OpenAI shape

[tangletools]

Summary

• New helper mcpToolsForRuntimeMcp() projects the 5 agent-runtime MCP delegation tools (delegate_code, delegate_research, delegate_feedback, delegation_status, delegation_history) into the OpenAI Chat Completions tools[] shape consumed by createOpenAICompatibleBackend.
• mcpToolsForRuntimeMcpSubset(names) filters that projection to a curated subset.
• Helper sources name + description + input schema from the canonical DELEGATE_*_TOOL_NAME / DELEGATE_*_DESCRIPTION / DELEGATE_*_INPUT_SCHEMA constants so the projection cannot drift from the server's own validators.
• Re-exported from the package root and from /mcp so eval / orchestrator code can pass tools: mcpToolsForRuntimeMcp() directly.
• Bump to 0.22.1; README gains a "Surfacing the tools through createOpenAICompatibleBackend" section.

Sandbox-SDK callers do not need this helper — the sandbox runtime mounts MCP servers natively and the in-sandbox harness discovers tools through the runtime, not through an OpenAI tools[] array.

Why

0.22.0 added tools + toolChoice to createOpenAICompatibleBackend, but the backend does not auto-discover tools from a parent sandbox's AgentProfile.mcpServers. Eval harnesses that wire tcloud / cli-bridge against an MCP-mounted runtime were not surfacing the delegation tools to the LLM. The gtm-agent canonical-delegation eval observed delegate_research: 0/3, delegate_code: 0/1, composite=0.07 because the model had no idea the tools existed.

This helper closes the projection gap. The companion PR in gtm-agent wires it through canonical-delegation so the eval can fire delegations end-to-end.

Test plan

• pnpm typecheck
• pnpm test — 267 tests pass, 11 new in tests/mcp/openai-tools.test.ts
• pnpm build — tsup clean (helper bundled into dist/mcp/index.js + package root)
• pnpm exec biome check src tests — clean
• Live smoke from gtm-agent eval once both PRs merge + 0.22.1 publishes

[drewstone]

Obsolete — solving wrong problem. The eval gap is that it bypasses the sandbox harness entirely; the right fix is routing eval through ensureWorkspaceSandbox + streamSandboxPrompt (or running real multi-shot with inline tool execution). Closing.