[argszero]

Summary

tool_choice was carried across protocols by exactly one direction. map_tool_choice()
was called only from anthropic_to_openai_chat_req, so every other cross-protocol request
translation silently dropped the field: responses -> chat, responses -> anthropic,
chat -> anthropic, chat -> responses and anthropic -> responses all lost it. Measured
through the real translators: a request that asks for tool_choice: "required", or names a
function, arrives upstream with no tool preference at all — so a model may answer in prose and
an agent client waits forever for a tool call it never gets.

This is not a lossy mapping by intent: map_tool_choice exists for this field, which is
proof the author wants it translated; the other directions simply have no counterpart. The
omission was invisible because grep -rn tool_choice src/ returns six lines and none of them
is a test.

Reachability (real determine_forwarding_protocol, all 12 shipped [[plans]]): every plan has
an openai_chat endpoint, so a chat inbound is always a passthrough — but a responses
inbound is converted on 11 of 12 plans (7 through the responses -> chat -> anthropic
chain, 4 through responses -> chat).

Related Issue

No issue exists for this; the finding came from a routine sweep of the request-side
translation. Nothing is fabricated or linked.

Changes

• src/protocol.rs only: the shape knowledge for tool_choice now has a single home.
  • Added a protocol-agnostic intermediate representation ToolChoice
    (None / Auto / Required / Function(name)) with parse_tool_choice(from, raw) and
    render_tool_choice(to, &tc), plus copy_tool_choice(body, from, to, result) — the shared
    tail every request translator now calls.
  • Wired the four cross directions: chat -> anthropic, chat -> responses,
    responses -> chat, and anthropic -> chat (which previously used map_tool_choice;
    map_tool_choice is now deleted — it had no other caller).
  • The two chained directions (responses -> anthropic = responses -> chat ∘
    chat -> anthropic; anthropic -> responses = anthropic -> chat ∘ chat -> responses)
    are fixed by the same rings.
  • Shapes: chat "none"|"auto"|"required" / {"type":"function","function":{"name":X}};
    anthropic {"type":"none"|"auto"|"any"} / {"type":"tool","name":X};
    responses "none"|"auto"|"required" / {"type":"function","name":X}.
  • A shape the source protocol cannot mean (e.g. {"type":"allowed_tools"}, an unknown
    scalar, a named tool with no name) is left absent instead of being copied verbatim —
    the old _ => clone() arm could forward a value the target protocol does not accept, and
    omitting the key keeps "absent" distinguishable from an explicit null.
• 涉及配置/数据结构的改动已同步示例文件 — n/a (no config, no schema, no ui/, no i18n keys)

Tests

• cargo test 全部通过 — 205 passed (was 202)
• cargo fmt --check 通过
• 新增/更新了单元测试

Added three tests:

• tool_choice_survives_every_cross_protocol_direction — per direction, three scalars plus one
  named tool, each expected value written out literally (not derived from the implementation),
  so a missing ring cannot be masked by a sibling.
• tool_choice_survives_the_two_chained_directions — the two chains asserted at the composition
  entry point (transform_request), because a chain's benefit is a composition and is not
  visible from the individual functions.
• tool_choice_is_not_invented — negative control: no tool_choice in the source ⇒ no key in
  the target; unrecognised shape ⇒ absent, never fabricated.

Teeth proof (each mutation applied alone, its own derived expected red set, restore in
finally with an mtime refresh): baseline passed=205 red=[]; removing any one of the four
call sites ⇒ passed=203 with exactly the two tests above red, and the failure output is the
defect (assertion left == right failed: responses->chat 源 "auto"; left: Null, right: String("auto")). An unrecognised shape is asserted by construction, not by omission.

Checklist

• 分支命名符合约定 (fix/request-tool-choice-cross-protocol)
• Commit message 使用 Conventional Commits 格式 (fix(protocol): ...)
• 单一职责，改动最小化 (one field's mapping in one file; no unrelated refactor)