Merge pull request #144 from Codonyat/gate-deferred-tools-to-moonshot

Gate the Kimi deferred-tools handshake to Moonshot providers

commit: 27cf8b3edd2c9bc8b8506e778c0a1ea61a89c079
author: Patric Eckhart · 2mo ago · 2026-08-16 11:40 UTC
committer: GitHub · 2mo ago · 2026-08-16 11:40 UTC
merge of: e900970 + 8c2eb47 · diff against the first parent
signature: GPG UnverifiedB5690EEEBB952194 The key is unknown
refs: tag: v0.3.44

2 files · +55 -1

M packages/provider/deferred_tools_test.go +39 -0


@@ -37,6 +37,45 @@ func TestKimiK3DeferredToolsLoadAtToolResult(t *testing.T) {
37	37		}
38	38		}
39	39
	40	+	// The deferred-tools handshake is a Moonshot API extension. Another provider
	41	+	// serving the same model id must get the standard wire instead: no tools-bearing
	42	+	// system message (it carries no content, which strict schemas reject), and the
	43	+	// activated tool in the top-level definitions where every provider reads it.
	44	+	func TestKimiK3DeferredToolsSkippedForNonMoonshotProvider(t *testing.T) {
	45	+	client := NewGondola("token", "").(*openaiClient)
	46	+	tools := []Tool{
	47	+	{Name: "search_tools", Schema: json.RawMessage(`{"type":"object"}`)},
	48	+	{Name: "lookup_weather", Schema: json.RawMessage(`{"type":"object"}`), Deferred: true},
	49	+	}
	50	+	wire, err := client.buildRequest(Request{
	51	+	Model: "kimi-k3",
	52	+	Tools: tools,
	53	+	Messages: []Message{
	54	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "weather"}}},
	55	+	{Role: RoleAssistant, Content: []Content{ToolCallBlock{ID: "call-1", Name: "search_tools", Arguments: json.RawMessage(`{}`)}}},
	56	+	{Role: RoleTool, Content: []Content{ToolResultBlock{CallID: "call-1", Content: []Content{TextBlock{Text: "found"}}}}, AddedToolNames: []string{"lookup_weather"}},
	57	+	},
	58	+	})
	59	+	if err != nil {
	60	+	t.Fatal(err)
	61	+	}
	62	+	for _, message := range wire.Messages {
	63	+	if len(message.Tools) > 0 {
	64	+	t.Fatalf("message %q carries tools; deferred injection must not reach a non-Moonshot provider", message.Role)
	65	+	}
	66	+	if message.Content == nil && len(message.ToolCalls) == 0 {
	67	+	t.Fatalf("message %q has neither content nor tool calls", message.Role)
	68	+	}
	69	+	}
	70	+	var names []string
	71	+	for _, tool := range wire.Tools {
	72	+	names = append(names, tool.Function.Name)
	73	+	}
	74	+	if len(names) != 2 \|\| names[0] != "search_tools" \|\| names[1] != "lookup_weather" {
	75	+	t.Fatalf("top-level tools = %v, want [search_tools lookup_weather]", names)
	76	+	}
	77	+	}
	78	+
40	79		func TestDeferredToolsFallbackToActiveTopLevelDefinitions(t *testing.T) {
41	80		tools := []Tool{
42	81		{Name: "base", Schema: json.RawMessage(`{"type":"object"}`)},


@@ -37,6 +37,45 @@ func TestKimiK3DeferredToolsLoadAtToolResult(t *testing.T) {
37	}	37		}
38	}	38		}
39		39
		40	+	// The deferred-tools handshake is a Moonshot API extension. Another provider
		41	+	// serving the same model id must get the standard wire instead: no tools-bearing
		42	+	// system message (it carries no content, which strict schemas reject), and the
		43	+	// activated tool in the top-level definitions where every provider reads it.
		44	+	func TestKimiK3DeferredToolsSkippedForNonMoonshotProvider(t *testing.T) {
		45	+	client := NewGondola("token", "").(*openaiClient)
		46	+	tools := []Tool{
		47	+	{Name: "search_tools", Schema: json.RawMessage(`{"type":"object"}`)},
		48	+	{Name: "lookup_weather", Schema: json.RawMessage(`{"type":"object"}`), Deferred: true},
		49	+	}
		50	+	wire, err := client.buildRequest(Request{
		51	+	Model: "kimi-k3",
		52	+	Tools: tools,
		53	+	Messages: []Message{
		54	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "weather"}}},
		55	+	{Role: RoleAssistant, Content: []Content{ToolCallBlock{ID: "call-1", Name: "search_tools", Arguments: json.RawMessage(`{}`)}}},
		56	+	{Role: RoleTool, Content: []Content{ToolResultBlock{CallID: "call-1", Content: []Content{TextBlock{Text: "found"}}}}, AddedToolNames: []string{"lookup_weather"}},
		57	+	},
		58	+	})
		59	+	if err != nil {
		60	+	t.Fatal(err)
		61	+	}
		62	+	for _, message := range wire.Messages {
		63	+	if len(message.Tools) > 0 {
		64	+	t.Fatalf("message %q carries tools; deferred injection must not reach a non-Moonshot provider", message.Role)
		65	+	}
		66	+	if message.Content == nil && len(message.ToolCalls) == 0 {
		67	+	t.Fatalf("message %q has neither content nor tool calls", message.Role)
		68	+	}
		69	+	}
		70	+	var names []string
		71	+	for _, tool := range wire.Tools {
		72	+	names = append(names, tool.Function.Name)
		73	+	}
		74	+	if len(names) != 2 \|\| names[0] != "search_tools" \|\| names[1] != "lookup_weather" {
		75	+	t.Fatalf("top-level tools = %v, want [search_tools lookup_weather]", names)
		76	+	}
		77	+	}
		78	+
40	func TestDeferredToolsFallbackToActiveTopLevelDefinitions(t *testing.T) {	79		func TestDeferredToolsFallbackToActiveTopLevelDefinitions(t *testing.T) {
41	tools := []Tool{	80		tools := []Tool{
42	{Name: "base", Schema: json.RawMessage(`{"type":"object"}`)},	81		{Name: "base", Schema: json.RawMessage(`{"type":"object"}`)},

M packages/provider/openai.go +16 -1


@@ -276,7 +276,7 @@ func (c openaiClient) buildRequest(req Request) (oaiRequest, error) {
276	276		out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})
277	277		}
278	278
279		-	deferredMode := isKimiDeferredModel(req.Model)
	279	+	deferredMode := supportsDeferredTools(c.Name()) && isKimiDeferredModel(req.Model)
280	280		toolByName := make(map[string]Tool, len(req.Tools))
281	281		for _, t := range req.Tools {
282	282		toolByName[t.Name] = t

@@ -392,6 +392,21 @@ func isKimiDeferredModel(model string) bool {
392	392		return model == "kimi-k3" \|\| strings.HasSuffix(model, "/kimi-k3")
393	393		}
394	394
	395	+	// supportsDeferredTools reports whether a provider speaks Moonshot's
	396	+	// deferred-tools handshake, where a tool that becomes available mid-conversation
	397	+	// is delivered on a system message carrying a `tools` array.
	398	+	//
	399	+	// That message is a Moonshot API extension, not OpenAI wire format: it has no
	400	+	// content, which a strict chat-completions schema rejects outright. Gondola,
	401	+	// which serves kimi-k3 on top of Venice, answers "messages.N: Invalid input"
	402	+	// and the conversation cannot recover, because the injection is rebuilt from
	403	+	// message history on every later turn. Providers outside this list take the
	404	+	// standard path instead, where an activated deferred tool joins the top-level
	405	+	// tools array.
	406	+	func supportsDeferredTools(provider string) bool {
	407	+	return provider == "moonshotai" \|\| provider == "moonshotai-cn"
	408	+	}
	409	+
395	410		func makeOAITool(t Tool) oaiTool {
396	411		var tool oaiTool
397	412		tool.Type = "function"


@@ -276,7 +276,7 @@ func (c openaiClient) buildRequest(req Request) (oaiRequest, error) {
276		out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})	276		out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})
277		}	277		}
278			278
279	-	deferredMode := isKimiDeferredModel(req.Model)	279	+	deferredMode := supportsDeferredTools(c.Name()) && isKimiDeferredModel(req.Model)
280		toolByName := make(map[string]Tool, len(req.Tools))	280		toolByName := make(map[string]Tool, len(req.Tools))
281		for _, t := range req.Tools {	281		for _, t := range req.Tools {
282		toolByName[t.Name] = t	282		toolByName[t.Name] = t

@@ -392,6 +392,21 @@ func isKimiDeferredModel(model string) bool {
392		return model == "kimi-k3" \|\| strings.HasSuffix(model, "/kimi-k3")	392		return model == "kimi-k3" \|\| strings.HasSuffix(model, "/kimi-k3")
393		}	393		}
394			394
			395	+	// supportsDeferredTools reports whether a provider speaks Moonshot's
			396	+	// deferred-tools handshake, where a tool that becomes available mid-conversation
			397	+	// is delivered on a system message carrying a `tools` array.
			398	+	//
			399	+	// That message is a Moonshot API extension, not OpenAI wire format: it has no
			400	+	// content, which a strict chat-completions schema rejects outright. Gondola,
			401	+	// which serves kimi-k3 on top of Venice, answers "messages.N: Invalid input"
			402	+	// and the conversation cannot recover, because the injection is rebuilt from
			403	+	// message history on every later turn. Providers outside this list take the
			404	+	// standard path instead, where an activated deferred tool joins the top-level
			405	+	// tools array.
			406	+	func supportsDeferredTools(provider string) bool {
			407	+	return provider == "moonshotai" \|\| provider == "moonshotai-cn"
			408	+	}
			409	+
395		func makeOAITool(t Tool) oaiTool {	410		func makeOAITool(t Tool) oaiTool {
396		var tool oaiTool	411		var tool oaiTool
397		tool.Type = "function"	412		tool.Type = "function"