test(anthropic): pin tool translation and upstream error shapes

Cover four zero-coverage functions on the /v1/messages route Claude
Code drives on nearly every call: openAIToolsFromAnthropic,
openAIToolChoiceFromAnthropic, responseTranslationError.Error, and
writeAnthropicProviderUpstreamError.

Add table-driven tests for every tool_choice form (auto, any,
tool-by-name, unrecognized, absent, non-map, nil), a multi-tool
mapping test that checks field order, description omission, and
malformed-entry skipping, a full request-to-response round trip
through a forced tool call, and a test proving no-tools/no-tool-choice
omits both keys rather than sending an empty array.

Add direct tests for the Anthropic-shaped upstream error envelope
(status passthrough, error.code, decoded error.upstream) and for the
translation-failure error prefix.

Every test was confirmed to fail against a real mutation of the
target function before being kept; each mutation and its revert are
recorded in the test's own doc comment.

routes_messages_test.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -679,3 +680,68 @@ func TestHandleMessages_AnthropicProvider_Passthrough_HeaderAllowlist(t *testing
 		t.Errorf("upstream x-api-key = %q, want the gateway's own configured key sk-ant-up, never the client's sk-alice", apiKeyValues[0])
 	}
 }
+
+// TestResponseTranslationError_ErrorPrefixesTranslateResponse proves
+// (*responseTranslationError).Error prefixes the wrapped error's own
+// message with "translate response: " — the text a log line built from
+// this error carries, distinguishing a translation failure (the upstream
+// already answered and was billed) from a connectivity failure in the
+// same log stream.
+//
+// MUTATION PROVEN: changing the method body to `return e.err.Error()`
+// (dropping the prefix) made this test fail — confirmed via `go test -run
+// TestResponseTranslationError_ErrorPrefixesTranslateResponse`, then
+// reverted with `git checkout -- routes_messages.go`.
+func TestResponseTranslationError_ErrorPrefixesTranslateResponse(t *testing.T) {
+	err := &responseTranslationError{err: errors.New("boom")}
+	if got, want := err.Error(), "translate response: boom"; got != want {
+		t.Errorf("Error() = %q, want %q", got, want)
+	}
+}
+
+// TestWriteAnthropicProviderUpstreamError_EmbedsStatusTypeAndDecodedBody
+// proves writeAnthropicProviderUpstreamError writes perr's own status
+// through as the HTTP status (not a fixed one), wraps it in the Anthropic
+// {"type":"error","error":{...}} envelope with error.type "upstream_error"
+// and error.code as perr's status stringified, and embeds perr.body —
+// decoded as JSON when it parses as one — under error.upstream
+// (decodeUpstreamErrorBody, routes_unified.go, shared with this
+// function's OpenAI-shaped counterpart writeProviderUpstreamError).
+//
+// MUTATION PROVEN: replacing `w.WriteHeader(perr.status)` with a
+// hardcoded `w.WriteHeader(http.StatusInternalServerError)` made the
+// status assertion fail — confirmed, then reverted. Separately proven:
+// removing `"upstream": decodeUpstreamErrorBody(perr)` from the envelope
+// made the upstream-body assertion fail — confirmed, then reverted.
+func TestWriteAnthropicProviderUpstreamError_EmbedsStatusTypeAndDecodedBody(t *testing.T) {
+	perr := &providerHTTPError{status: http.StatusServiceUnavailable, body: []byte(`{"error":{"message":"overloaded"}}`)}
+	rec := httptest.NewRecorder()
+
+	writeAnthropicProviderUpstreamError(rec, "anthropic", perr)
+
+	if rec.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status = %d, want 503 (perr's own status)", rec.Code)
+	}
+	var got map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &got); err != nil {
+		t.Fatalf("response is not valid JSON: %v, body=%s", err, rec.Body.String())
+	}
+	if got["type"] != "error" {
+		t.Errorf(`top-level "type" = %v, want "error"`, got["type"])
+	}
+	errObj, _ := got["error"].(map[string]any)
+	if errObj["type"] != "upstream_error" {
+		t.Errorf(`error.type = %v, want "upstream_error"`, errObj["type"])
+	}
+	if errObj["message"] != "anthropic upstream error" {
+		t.Errorf(`error.message = %v, want "anthropic upstream error"`, errObj["message"])
+	}
+	if errObj["code"] != "503" {
+		t.Errorf(`error.code = %v, want "503"`, errObj["code"])
+	}
+	upstream, _ := errObj["upstream"].(map[string]any)
+	inner, _ := upstream["error"].(map[string]any)
+	if inner["message"] != "overloaded" {
+		t.Errorf(`error.upstream decoded body = %v, want the parsed provider body {"error":{"message":"overloaded"}}`, errObj["upstream"])
+	}
+}

translate_anthropic_test.go

@@ -856,3 +856,175 @@ func TestAnthropicUsagePayload_TotalInputTokens_FoldsCacheCounters(t *testing.T)
 		t.Errorf("totalInputTokens() = %d, want %d", got, want)
 	}
 }
+
+// --- openAIToolsFromAnthropic / openAIToolChoiceFromAnthropic: the
+// Anthropic tools[]/tool_choice request translators, on the newest,
+// heaviest-traffic path (/v1/messages -> a non-anthropic provider). Claude
+// Code sends a tools[] array on nearly every call. Both functions were
+// previously ZERO-COVERED: every existing openAIRequestFromAnthropic test
+// above never set "tools" or "tool_choice" on its request map. ---
+
+// TestOpenAIToolsFromAnthropic_MapsFieldsAndSkipsMalformedEntries proves
+// each Anthropic tool{name, description, input_schema} maps IN ORDER to
+// an OpenAI {"type":"function","function":{"name","description",
+// "parameters"}} entry, that "description" is OMITTED (not set to a zero
+// value) when the source tool carries none, and that a non-map entry in
+// toolsRaw is skipped rather than panicking or producing a garbage entry.
+//
+// MUTATION PROVEN: changing `"parameters": tm["input_schema"]` to
+// `"parameters": tm["name"]` made this test fail (parameters carried the
+// tool's name string instead of its JSON schema) — confirmed via `go test
+// -run TestOpenAIToolsFromAnthropic`, then reverted with `git checkout --
+// translate_anthropic.go`.
+func TestOpenAIToolsFromAnthropic_MapsFieldsAndSkipsMalformedEntries(t *testing.T) {
+	schema := map[string]any{"type": "object", "properties": map[string]any{"city": map[string]any{"type": "string"}}}
+	toolsRaw := []any{
+		map[string]any{"name": "get_weather", "description": "Get the weather", "input_schema": schema},
+		map[string]any{"name": "no_desc_tool", "input_schema": map[string]any{"type": "object"}},
+		"not a map", // malformed entry: must be skipped, not panic
+		map[string]any{"name": "empty_schema_tool", "input_schema": map[string]any{}},
+	}
+
+	out := openAIToolsFromAnthropic(toolsRaw)
+	if len(out) != 3 {
+		t.Fatalf("len(out) = %d, want 3 (the malformed entry must be skipped)", len(out))
+	}
+
+	first, _ := out[0].(map[string]any)
+	if first["type"] != "function" {
+		t.Errorf(`out[0]["type"] = %v, want "function"`, first["type"])
+	}
+	fn, _ := first["function"].(map[string]any)
+	assert.Equal(t, "get_weather", fn["name"])
+	assert.Equal(t, "Get the weather", fn["description"])
+	assert.Equal(t, schema, fn["parameters"])
+
+	second, _ := out[1].(map[string]any)
+	fn2, _ := second["function"].(map[string]any)
+	if _, hasDesc := fn2["description"]; hasDesc {
+		t.Errorf(`function["description"] present (%v) for a tool with no description, want the key omitted`, fn2["description"])
+	}
+
+	third, _ := out[2].(map[string]any)
+	fn3, _ := third["function"].(map[string]any)
+	assert.Equal(t, "empty_schema_tool", fn3["name"])
+	assert.Equal(t, map[string]any{}, fn3["parameters"])
+}
+
+// TestOpenAIToolChoiceFromAnthropic_AllForms proves each of Anthropic's
+// tool_choice shapes maps to OpenAI's, per the function's own doc
+// comment: "auto"->"auto", "any"->"required" (OpenAI's force-some-tool
+// value), {"type":"tool","name":n}->{"type":"function","function":
+// {"name":n}}, and an unrecognized type, an absent type, a non-map value,
+// or nil all map to a Go nil — the signal openAIRequestFromAnthropic's own
+// caller-side `if choice := ...; choice != nil` uses to decide whether to
+// set "tool_choice" on the OpenAI request at all.
+//
+// MUTATION PROVEN: changing the "any" case to `return "auto"` (collapsing
+// it into the "auto" case, so a forced tool call silently became merely
+// advisory) made this test fail — confirmed, then reverted.
+func TestOpenAIToolChoiceFromAnthropic_AllForms(t *testing.T) {
+	cases := []struct {
+		tc   any
+		want any
+		name string
+	}{
+		{name: "auto", tc: map[string]any{"type": "auto"}, want: "auto"},
+		{name: "any becomes OpenAI's required", tc: map[string]any{"type": "any"}, want: "required"},
+		{name: "tool-by-name", tc: map[string]any{"type": "tool", "name": "get_weather"}, want: map[string]any{"type": "function", "function": map[string]any{"name": "get_weather"}}},
+		{name: "unrecognized type is nil", tc: map[string]any{"type": "bogus"}, want: nil},
+		{name: "absent type is nil", tc: map[string]any{}, want: nil},
+		{name: "non-map value is nil", tc: "auto", want: nil},
+		{name: "nil value is nil", tc: nil, want: nil},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := openAIToolChoiceFromAnthropic(c.tc)
+			assert.Equal(t, c.want, got)
+		})
+	}
+}
+
+// TestOpenAIRequestFromAnthropic_ToolsAndToolChoiceRoundTrip is the
+// end-to-end regression for openAIToolsFromAnthropic and
+// openAIToolChoiceFromAnthropic wired through their only real caller,
+// openAIRequestFromAnthropic: an Anthropic /v1/messages request carrying a
+// multi-tool tools[] array and a tool_choice forcing one specific tool
+// translates to a request shape a real OpenAI chat-completions endpoint
+// accepts, and the tool_calls[] such an endpoint would answer with
+// translates back to Anthropic's own tool_use content-block shape — the
+// full round trip Claude Code actually drives on this route.
+func TestOpenAIRequestFromAnthropic_ToolsAndToolChoiceRoundTrip(t *testing.T) {
+	req := map[string]any{
+		"model": "claude-x",
+		"tools": []any{
+			map[string]any{"name": "get_weather", "description": "Get the weather for a city", "input_schema": map[string]any{
+				"type": "object", "properties": map[string]any{"city": map[string]any{"type": "string"}}, "required": []any{"city"},
+			}},
+			map[string]any{"name": "get_time", "input_schema": map[string]any{"type": "object"}},
+		},
+		"tool_choice": map[string]any{"type": "tool", "name": "get_weather"},
+		"messages":    []any{map[string]any{"role": "user", "content": "What's the weather in London?"}},
+	}
+
+	out, _, err := openAIRequestFromAnthropic(req)
+	if err != nil {
+		t.Fatalf("openAIRequestFromAnthropic: %v", err)
+	}
+
+	tools, _ := out["tools"].([]any)
+	if len(tools) != 2 {
+		t.Fatalf(`out["tools"] = %v, want 2 entries`, out["tools"])
+	}
+	firstFn, _ := tools[0].(map[string]any)["function"].(map[string]any)
+	assert.Equal(t, "get_weather", firstFn["name"])
+	secondFn, _ := tools[1].(map[string]any)["function"].(map[string]any)
+	assert.Equal(t, "get_time", secondFn["name"])
+
+	wantChoice := map[string]any{"type": "function", "function": map[string]any{"name": "get_weather"}}
+	assert.Equal(t, wantChoice, out["tool_choice"])
+
+	// A real OpenAI-shaped response answering that forced tool call...
+	const body = `{"id":"chatcmpl-1","choices":[{"index":0,"message":{"content":null,"tool_calls":[{"id":"call_1","function":{"name":"get_weather","arguments":"{\"city\":\"London\"}"}}]},"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":10,"completion_tokens":5}}`
+	anthropicOut, err := anthropicResponseFromOpenAI([]byte(body), "claude-x")
+	if err != nil {
+		t.Fatalf("anthropicResponseFromOpenAI: %v", err)
+	}
+	// ...must translate back to Anthropic's tool_use shape.
+	assert.Equal(t, "tool_use", anthropicOut["stop_reason"])
+	content, _ := anthropicOut["content"].([]any)
+	if len(content) != 1 {
+		t.Fatalf("content = %v, want one tool_use block", anthropicOut["content"])
+	}
+	block, _ := content[0].(map[string]any)
+	assert.Equal(t, "get_weather", block["name"])
+	input, _ := block["input"].(map[string]any)
+	assert.Equal(t, "London", input["city"])
+}
+
+// TestOpenAIRequestFromAnthropic_NoTools_OmitsToolsAndToolChoiceFields
+// proves a request that never set Anthropic's "tools"/"tool_choice"
+// fields produces an OpenAI request with NEITHER key present at all — not
+// an empty "tools":[] array, which some OpenAI-compatible upstreams treat
+// differently from the field's outright absence.
+//
+// MUTATION PROVEN: changing `if toolsRaw, ok := req["tools"].([]any); ok`
+// to always assign `out["tools"] = openAIToolsFromAnthropic(toolsRaw)`
+// regardless of ok made this test fail (an empty "tools" key appeared) —
+// confirmed, then reverted.
+func TestOpenAIRequestFromAnthropic_NoTools_OmitsToolsAndToolChoiceFields(t *testing.T) {
+	req := map[string]any{
+		"model":    "claude-x",
+		"messages": []any{map[string]any{"role": "user", "content": "hi"}},
+	}
+	out, _, err := openAIRequestFromAnthropic(req)
+	if err != nil {
+		t.Fatalf("openAIRequestFromAnthropic: %v", err)
+	}
+	if _, ok := out["tools"]; ok {
+		t.Errorf(`out["tools"] = %v, want the key absent when no tools were requested`, out["tools"])
+	}
+	if _, ok := out["tool_choice"]; ok {
+		t.Errorf(`out["tool_choice"] = %v, want the key absent when no tool_choice was requested`, out["tool_choice"])
+	}
+}