Skip to content

A self-improving RLM agent for coding workflows and long-running autonomous tasks.

Rust

fix(ai): stop sending enable_thinking to Prime Inference GLM models (#2459)

rasdani63d88319bf58
diff --git a/packages/ai/.changes/prime-inference-glm-thinking-compat.md b/packages/ai/.changes/prime-inference-glm-thinking-compat.md
new file mode 100644
index 0000000000..30f48e1e97
--- /dev/null
+++ b/packages/ai/.changes/prime-inference-glm-thinking-compat.md
@@ -0,0 +1 @@
+- Fixed Prime Inference GLM models failing every request with an unsupported `enable_thinking` parameter.
diff --git a/packages/ai/scripts/generate-models.ts b/packages/ai/scripts/generate-models.ts
index 5a942ab3a7..c1e170d3cd 100644
--- a/packages/ai/scripts/generate-models.ts
+++ b/packages/ai/scripts/generate-models.ts
@@ -113,6 +113,13 @@ const ZAI_THINKING_COMPAT: OpenAICompletionsCompat = {
thinkingFormat: "zai",
};
+// Prime Inference rejects `enable_thinking` on GLM routes with a 400 on every
+// request. The routes think by default, so send no thinking parameter (and no
+// reasoning_effort, whose support is unverified).
+const PRIME_INFERENCE_ZAI_COMPAT: OpenAICompletionsCompat = {
+ supportsReasoningEffort: false,
+};
+
const PRIME_INFERENCE_BASE_URL = "https://api.pinference.ai/api/v1";
const PRIME_INFERENCE_COMPAT: OpenAICompletionsCompat = {
supportsStore: false,
@@ -448,6 +455,10 @@ function isPrimeInferenceReasoningModel(modelId: string, catalogReasoning?: bool
);
}
+function isPrimeInferenceZaiModelId(modelId: string): boolean {
+ return modelId.toLowerCase().startsWith("z-ai/glm-");
+}
+
function getPrimeInferenceCompat(modelId: string): OpenAICompletionsCompat {
const id = modelId.toLowerCase();
if (id.includes("deepseek-v4")) {
@@ -456,10 +467,10 @@ function getPrimeInferenceCompat(modelId: string): OpenAICompletionsCompat {
...DEEPSEEK_V4_COMPAT,
};
}
- if (id.startsWith("z-ai/glm-")) {
+ if (isPrimeInferenceZaiModelId(id)) {
return {
...PRIME_INFERENCE_COMPAT,
- ...ZAI_THINKING_COMPAT,
+ ...PRIME_INFERENCE_ZAI_COMPAT,
};
}
@@ -606,7 +617,11 @@ function createPrimeInferenceModel(
...(openRouter?.supportsReasoningEffort === false
? {
supportsReasoningEffort: false,
- ...(!compat.thinkingFormat ? { thinkingFormat: "openrouter" as const } : {}),
+ // GLM routes must not receive any thinking parameter; the
+ // `reasoning` object is not verified against the gateway.
+ ...(!compat.thinkingFormat && !isPrimeInferenceZaiModelId(entry.id)
+ ? { thinkingFormat: "openrouter" as const }
+ : {}),
}
: {}),
},
diff --git a/packages/ai/src/models.generated.ts b/packages/ai/src/models.generated.ts
index ff0f332401..b5ff751ed7 100644
--- a/packages/ai/src/models.generated.ts
+++ b/packages/ai/src/models.generated.ts
@@ -17924,7 +17924,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -17943,7 +17943,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -17962,7 +17962,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -17981,7 +17981,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -18000,7 +18000,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -18019,7 +18019,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -18039,7 +18039,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null},
input: ["text"],
@@ -18059,7 +18059,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"minimal":null,"low":null,"medium":null,"high":"high","xhigh":"xhigh","max":null},
input: ["text"],
@@ -18079,7 +18079,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"off":null,"minimal":null,"low":"low","medium":null,"high":"high","xhigh":null,"max":"max"},
input: ["text"],
@@ -18098,7 +18098,7 @@ export const MODELS = {
api: "openai-completions",
provider: "prime-inference",
baseUrl: "https://api.pinference.ai/api/v1",
- compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false,"thinkingFormat":"zai"},
+ compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false,"maxTokensField":"max_tokens","supportsStrictMode":false},
reasoning: true,
thinkingLevelMap: {"off":null,"minimal":null,"low":"low","medium":null,"high":"high","xhigh":null,"max":"max"},
input: ["text", "image"],
diff --git a/packages/ai/test/openai-completions-tool-choice.test.ts b/packages/ai/test/openai-completions-tool-choice.test.ts
index 04d3e0b771..17b0508e83 100644
--- a/packages/ai/test/openai-completions-tool-choice.test.ts
+++ b/packages/ai/test/openai-completions-tool-choice.test.ts
@@ -1145,6 +1145,25 @@ describe("openai-completions tool_choice", () => {
expect((payload as { reasoning_effort?: unknown }).reasoning_effort).toBe("none");
});
+ it("sends no thinking parameter to Prime Inference GLM routes but keeps the z.ai toggle", async () => {
+ const context = { messages: [{ role: "user" as const, content: "Hi", timestamp: Date.now() }] };
+ const payloads = new Map<string, Record<string, unknown>>();
+ for (const model of [getModel("prime-inference", "z-ai/glm-5.3")!, getModel("zai", "glm-5.3")!]) {
+ await streamSimple(model, context, {
+ apiKey: "test",
+ reasoning: "high",
+ onPayload: (params: unknown) => {
+ payloads.set(model.id, params as Record<string, unknown>);
+ },
+ }).result();
+ }
+ const prime = payloads.get("z-ai/glm-5.3");
+ expect(prime?.enable_thinking).toBeUndefined();
+ expect(prime?.reasoning_effort).toBeUndefined();
+ expect(prime?.reasoning).toBeUndefined();
+ expect(payloads.get("glm-5.3")?.enable_thinking).toBe(true);
+ });
+
it("serializes explicit off only for models that allow disabling reasoning", async () => {
const baseModel = getModel("openrouter", "deepseek/deepseek-r1")!;
const effortCompat = { ...baseModel.compat, supportsReasoningEffort: true };
diff --git a/packages/coding-agent/src/core/model-resolver.ts b/packages/coding-agent/src/core/model-resolver.ts
index c8f57f7745..5d3bcfaf5b 100644
--- a/packages/coding-agent/src/core/model-resolver.ts
+++ b/packages/coding-agent/src/core/model-resolver.ts
@@ -174,9 +174,8 @@ function buildFallbackModel(provider: string, modelId: string, availableModels:
// Daemon-created sessions (rlm.create_session) re-resolve their model from
// provider/id strings in a registry that has not refreshed the team-authorized
// private catalog, so unknown private ids reach this fallback. They must
- // inherit a private-route template: the public provider default carries the
- // zai thinking format, whose enable_thinking parameter the private endpoint
- // rejects with a 400 on every request.
+ // inherit a private-route template: public-route limits and thinking-level
+ // maps do not describe private routes.
if (isPrivatePrimeInferenceModel({ provider, id: modelId })) {
const privateTemplate = providerModels.find((m) => isPrivatePrimeInferenceModel(m));
if (!privateTemplate) return undefined;
diff --git a/packages/coding-agent/test/create-session-private-model-params.test.ts b/packages/coding-agent/test/create-session-private-model-params.test.ts
index 784a05c1e6..05c222a1cb 100644
--- a/packages/coding-agent/test/create-session-private-model-params.test.ts
+++ b/packages/coding-agent/test/create-session-private-model-params.test.ts
@@ -78,7 +78,7 @@ describe("created-session private model request params", () => {
expect("enable_thinking" in params).toBe(false);
});
- it("keeps enable_thinking for public zai-format Prime Inference models", async () => {
+ it("omits enable_thinking for public Prime Inference z-ai models", async () => {
const model = getModels("prime-inference").find(
(candidate) => candidate.id === "z-ai/glm-5.3",
) as Model<"openai-completions">;
@@ -90,6 +90,6 @@ describe("created-session private model request params", () => {
).result();
const params = mockState.lastParams as { enable_thinking?: boolean };
- expect(params.enable_thinking).toBe(true);
+ expect("enable_thinking" in params).toBe(false);
});
});
diff --git a/packages/coding-agent/test/model-resolver.test.ts b/packages/coding-agent/test/model-resolver.test.ts
index aaec47c7b2..3647c351cc 100644
--- a/packages/coding-agent/test/model-resolver.test.ts
+++ b/packages/coding-agent/test/model-resolver.test.ts
@@ -168,16 +168,16 @@ describe("resolveCliModel", () => {
expect(priv.model?.provider).toBe("prime-inference");
expect(priv.model?.baseUrl).toBe("https://api.pinference.ai/api/v1");
const privateModel = priv.model as Model<"openai-completions">;
- // The public provider default carries the zai thinking format; a private
- // route must not inherit it (enable_thinking is a provider 400 there).
+ // Prime Inference rejects enable_thinking, so no route may carry the zai
+ // thinking format.
expect(privateModel.compat?.thinkingFormat).toBeUndefined();
- // The zai thinkingLevelMap would coerce thinking "off" to "low".
+ // The public template's thinkingLevelMap would coerce thinking "off" to "low".
expect(getSupportedThinkingLevels(privateModel).includes("off")).toBe(true);
const pub = resolveCliModel({ cliProvider: "prime-inference", cliModel: "z-ai/glm-9", modelRegistry: registry });
expect(pub.error).toBeUndefined();
expect(pub.model?.id).toBe("z-ai/glm-9");
- expect((pub.model as Model<"openai-completions">).compat?.thinkingFormat).toBe("zai");
+ expect((pub.model as Model<"openai-completions">).compat?.thinkingFormat).toBeUndefined();
});
});