[pd95]

Problem

Fixes #15921.

Ollama flattens namespace tools for its internal chat representation and must
recover the original (namespace, name) pair when returning a Responses tool
call. Losing that pair prevents clients such as Codex from routing the call to
the requested MCP tool. The Codex multi-agent failure reported in #18284 is a
real-world example of this round-trip symptom.

Current main expands namespace declarations and restores ordinary namespaced
calls, but its lookup is not injective. For example,
namespace a.b / member c, namespace a / member b.c, and flat function
a.b.c all produce the same internal name. Resolution then depends on
declaration order and can return the wrong tool identity to the client. Mixed
web-search output also bypasses request-scoped namespace restoration.

What this changes

• builds one request-scoped, bidirectional mapping for declared tools,
  tool-search results, and tool-call history;
• rejects duplicate or colliding identities instead of selecting one by
  declaration order;
• canonicalizes dotted, colon, and Codex-style aliases when they are
  unambiguous;
• restores the exact namespace and child name in streaming, non-streaming, and
  mixed web-search Responses output.

Scope

The change is limited to Responses request/output translation and the existing
web-search response assembly path.

Validation

go test ./openai ./middleware -count=1 passes. Focused tests cover namespace
round trips, collision rejection independent of declaration order, legacy
aliases, tool-search results, history, streaming, non-streaming, and mixed
web-search output. Repository lint and go build ./... also pass.

Runtime validation passes non-streaming, streaming, and namespaced-history
Responses requests with gpt-oss:20b.

AI assistance

Codex materially assisted with implementation review, tests, validation, and
this PR description. The contributor reviewed the final changes and test
results.

[pd95]

@drifkin would you mind taking a look at this PR (#16263) when you have a chance?

This PR adds support for namespaced tool calls in the OpenAI-compatible Responses API, addressing #15921. It translates type: "namespace" tools into Ollama’s existing flat internal tool naming, then restores
the namespace field on Responses API function call output items so clients like Codex can resolve MCP tools correctly.

If the approach looks reasonable, I’d appreciate a review or help routing it toward merge.

[pd95]

@drifkin, a quick update following the namespace support added by @dhiltgen in #17593—thanks for providing that foundation.

I have rebased and substantially narrowed this PR so it no longer duplicates the basic namespace expansion. It is now a single commit containing only the remaining correctness pieces:

• use one request-scoped, bidirectional namespace/name resolver;
• reject duplicate or ambiguous flattened identities instead of resolving them by declaration order;
• canonicalize dotted, colon, and Codex-style aliases when unambiguous;
• restore namespace identity for mixed web-search/function-call output;
• cover streaming, non-streaming, history, tool-search output, and collision cases with focused tests.

The PR is based directly on current main, and its focused tests, repository lint, build, and runtime namespace round-trip checks pass.

Would you be willing to take another look at the updated, reduced diff?