feat(ai-proxy): translate tool use for the Anthropic provider

CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **plugins/ai-proxy**: tool-use translation for the Anthropic provider on both Chat Completions and the Responses API. Previously the Anthropic translation dropped the client's `tools`/`tool_choice` entirely (the model never saw the tools, so it could never call them) and ignored `tool_use` blocks in the response (a tool-calling turn came back malformed, with `finish_reason: "tool_calls"` but no `tool_calls`). Now: `tools` and `tool_choice` are mapped to Anthropic's `tools`/`tool_choice` (`parameters` → `input_schema`, `"required"` → `any`, `parallel_tool_calls: false` → `disable_parallel_tool_use`); assistant `tool_calls` and `role:"tool"` history messages are translated to `tool_use`/`tool_result` blocks; and Anthropic `tool_use` responses are translated back to OpenAI `tool_calls` / Responses `function_call`. Codex freeform `custom` tools (e.g. `apply_patch`), `local_shell`, and hosted server tools have no Anthropic representation and are now rejected on the Responses path with `400 custom_tools_not_supported_for_provider` instead of being dropped silently. Shared mapping lives in a new `protocols::tools` module. OpenAI/Ollama remain passthrough.
 - **plugins/ai-proxy**: configurable credential attachment via a new `auth` field on targets, routes, and the flat config. `auth` is orthogonal to `provider` (which selects the wire protocol): `bearer` → `Authorization: Bearer`, `api_key` → `x-api-key`, `{ header: "Name" }` → an arbitrary credential header, `{ query: "param" }` → key in the query string. When omitted it defaults to the provider's convention (bearer for OpenAI/Ollama, `x-api-key` for Anthropic), so existing configs are unchanged. This lets OpenAI-compatible endpoints with non-standard credential headers (e.g. Brave AI Grounding's `X-Subscription-Token`, Azure OpenAI's `api-key`) be configured without a dedicated provider type. Internally, the three previously hardcoded auth call sites (OpenAI transport, Anthropic transport, `/v1/models` aggregator) now share a single `apply_auth` implementation.
 
 ## [0.8.1] - 2026-07-15

docs/guide/dispatchers.md

@@ -913,6 +913,15 @@ OpenAI-compatible providers (OpenAI, Ollama) stream natively via `host_http_stre
 
 Anthropic streaming is buffered: the dispatcher waits for the full response and returns it non-streamed (a warning is logged). True SSE translation is deferred per ADR-0024 / ADR-0030 §2.
 
+#### Tool use (Anthropic translation)
+
+For OpenAI/Ollama the tool surface is passthrough. For the **Anthropic** provider the dispatcher translates tools in both directions on Chat Completions and Responses:
+
+- **Request:** the client's `tools` are mapped to Anthropic's `tools` (`function.parameters` / flat `parameters` → `input_schema`), and `tool_choice` is mapped (`"auto"`/`"none"` → same, `"required"` → `any`, `{function:{name}}`/`{name}` → `{type:"tool", name}`); `parallel_tool_calls: false` becomes `disable_parallel_tool_use: true`.
+- **Response:** Anthropic `tool_use` blocks become OpenAI `tool_calls` (Chat Completions) / `function_call` items (Responses); the assistant `content` is `null` on a tool-only turn, and `finish_reason` is `tool_calls`.
+- **History:** an assistant `tool_calls` message → `tool_use` blocks, and `role:"tool"` messages → `tool_result` blocks (consecutive results merged into one user turn).
+- **Unsupported tool types:** Codex freeform `custom` tools (e.g. `apply_patch`), `local_shell`, and hosted server tools have no Anthropic representation. On the Responses path they are rejected with `400 custom_tools_not_supported_for_provider` rather than dropped silently. Route those to OpenAI.
+
 #### Responses API specifics
 
 `POST /v1/responses` is **stateless-only** (ADR-0030 §2):