[coffeerunhobby]

Description

Codex and other Responses API clients send each MCP server's tools as one namespace entry,
{"type": "namespace", "name": "mcp__docs", "tools": [...]}, and route a returned call by its
(namespace, name) pair. convertORToolsToOpenAIFormat keeps only function tools, so the model
never sees an MCP tool, and Codex reports that the tools are not available.

This PR flattens each namespace member to ns_{namespace}___{name}, the form llama.cpp's server
uses (ggml-org/llama.cpp#31568). That flat name goes to the model and the grammar. A
function_call item that carries a namespace in the input, or in a replayed stored response, is
flattened the same way. Returned function_call items are split back into namespace and name
where outputs leave: in buildORResponse and in normalizeORStreamEvent. That covers
non-streaming, streaming, background, WebSocket and stored responses.

Notes for Reviewers

• The name decodes by itself, so no per-request state is kept. The one assumption, the same as
  llama.cpp's: a namespace does not contain ___.
• Namespace members are client tools: the client that declared them runs the calls. They are never
  matched against LocalAI's server-side MCP tools (which are injected as plain function tools under
  their own names), even if the names coincide, so a client-declared tool cannot trigger a server tool.
• Out of scope, possible follow-ups:
  • client-executed tool_search and additional_tools (tools loaded during a turn);
  • tool_choice naming a namespaced function;
  • the 64-character function-name limit some templates have (Codex connector names can be long).
• Swagger: only the two new fields (tools on schema.ORFunctionTool, namespace on
  schema.ORItemParam). Running swag init on current master also picks up unrelated
  locked/lockedReason drift, which is left out.

Tests: go test ./core/http/endpoints/openresponses/ ./core/schema/ (87/87 and 83/83). The new
specs cover the tool conversion, replaying a namespaced call, and a streamed call coming back with
its namespace in output_item.added, output_item.done and response.completed. They fail with
the change reverted.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable

[coffeerunhobby]

I almost forgot: for context, all others are adding the same support for namespace tools :) We are just lucky we see around the corner:

• vLLM: [Frontend] Support OpenAI Responses API namespace tools vllm-project/vllm#47024 (merged 2026-07-06); the harmony path is in [Bugfix][Frontend] Support namespace tools for harmony/gpt-oss models in Responses API vllm-project/vllm#49581
• llama.cpp server: server : implement namespaced tools in responses API (rebase of #23235) ggml-org/llama.cpp#31568 (open); this PR uses the same ns_{namespace}___{name} naming
• Ollama: responses: preserve namespace tool identity ollama/ollama#16263 (open)
• LiteLLM: fix(responses): keep flattened namespace tool names within 64 characters BerriAI/litellm#46243 (open)