feat(ai): add Z.AI GLM-5.3

packages/ai/CHANGELOG.md

@@ -10,9 +10,11 @@
 
 ### Fixed
 
+- Preserve GLM-5.3 Flash and Highspeed reasoning effort mappings and Z.AI thinking serialization.
 - Coalesced adjacent Anthropic user and tool-result turns without changing standalone string user-message content.
 - Anthropic prompt caching now retains the previous checkpoint while tool loops append a new result, avoiding repeated prefix reprocessing for API-key and OAuth requests.
 
+
 ### Removed
 
 ## [2026.8.29] - 2026-08-29

packages/ai/changes.md

@@ -1,5 +1,65 @@
 # changes.md — ai
 
+## 2026-08-29 - Narrow GLM-5.3 serializer matching
+
+### What changed
+
+- `src/api/openai-completions.ts`
+
+### Why
+
+- Unsupported GLM-5.3 suffixes must not inherit validated model reasoning controls.
+
+### Why an extension could not handle it
+
+- Z.AI request serialization is implemented inside the AI package provider adapter.
+
+### Expected merge conflict zones
+
+- `src/api/openai-completions.ts`
+
+## 2026-08-29 - Preserve GLM-5.3 variant reasoning controls
+
+### What changed
+
+- `packages/ai/scripts/generate-models.ts`: recognize `glm-5.3` Flash and Highspeed variants as part of the Z.AI GLM-5.3 family so generated metadata retains low/high/max reasoning effort mappings and wire support.
+- `packages/ai/src/api/openai-completions.ts`; `src/api/openai-completions.ts`: Z.AI GLM-5.3 thinking serialization.
+- src/api/openai-completions.ts
+- `src/api/openai-completions.ts`: Z.AI GLM-5.3 thinking serialization.
+
+### Why
+
+- Without family matching, explicit reasoning effort selections were dropped for the Flash and Highspeed models and their model controls exposed incorrect levels.
+
+### Why an extension could not handle it
+
+- Model-family metadata and request serialization are built into the AI package generator and provider adapter before extension code can intervene.
+
+### Expected merge conflict zones
+
+- `packages/ai/scripts/generate-models.ts`: Z.AI model-family detection and generated capability metadata.
+- `packages/ai/src/api/openai-completions.ts`; `src/api/openai-completions.ts`: Z.AI GLM-5.3 thinking serialization.
+- src/api/openai-completions.ts
+
+## 2026-08-29 - Z.AI GLM-5.3 request and catalog fixes
+
+### What changed
+
+- `src/api/openai-completions.ts`: narrow GLM-5.3 thinking serialization to validated variants.
+- GLM-5.3 reasoning-off requests now use the provider's lowest enabled effort instead of sending the rejected disabled-thinking payload; regenerated Z.AI catalogs retain the separate global and China sources and published model metadata.
+
+### Why
+
+- Unsupported GLM-5.3 suffixes must not inherit validated model reasoning controls.
+
+### Why an extension could not handle it
+
+- Z.AI request serialization is implemented inside the AI package provider adapter.
+
+### Expected merge conflict zones
+
+- `src/api/openai-completions.ts`.
+
 ## Anthropic cache checkpoints across tool loops (2026-08-29)
 
 ### What changed
@@ -17,7 +77,6 @@
 ### Expected merge conflict zones
 
 - MEDIUM: `src/api/anthropic-messages.ts` message coalescing and final cache-marker pass.
-
 ## Credential pool export wildcard (2026-08-27)
 
 ### What changed

packages/ai/scripts/generate-models.ts

@@ -1930,7 +1930,7 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 				const supportsImage = m.modalities?.input?.includes("image");
 
 				const isGlm52 = modelId === "glm-5.2" || modelId === "glm-5.2-highspeed";
-				const isGlm5x = isGlm52 || modelId === "glm-5.3";
+				const isGlm5x = isGlm52 || /^(?:glm-5\.3)(?:-(?:flash|highspeed))?$/.test(modelId);
 				const referenceCost = modelId === "glm-5.2-highspeed" ? undefined : data.zai?.models[modelId]?.cost ?? m.cost;
 
 				models.push({

packages/ai/src/api/openai-completions.ts

@@ -1094,7 +1094,7 @@ function buildParams(
 	const thinkingBudget = resolveClampedThinkingBudget(model, options, params);
 
 	if (compat.thinkingFormat === "zai" && model.reasoning) {
-		const isGlm53 = /(?:^|[/:-])glm-5\.3(?:$|[/.:_-])/.test(model.id.toLowerCase());
+		const isGlm53 = /(?:^|[/:-])glm-5\.3(?:-(?:flash|highspeed))?(?:$|[/.:_])/.test(model.id.toLowerCase());
 		const zaiParams = params as Omit<typeof params, "reasoning_effort"> & {
 			thinking?: { type: "enabled" | "disabled"; clear_thinking?: boolean };
 			reasoning_effort?: string;

packages/ai/src/changes.md

@@ -208,6 +208,26 @@ The divergence lives in core wiring, package identity, or build plumbing that ex
 
 ## Unreleased
 
+## 2026-08-29 - Cover GLM-5.3 generator negative variants
+
+### What changed
+
+- `packages/ai/src/api/openai-completions.ts`: narrowed the Z.AI always-enabled matcher to the exact `glm-5.3`, `glm-5.3-flash`, and `glm-5.3-highspeed` variants so unsupported variants are not forced into thinking.
+- `scripts/generate-models.ts`: generated Z.AI records for unsupported GLM-5.3 variants omit `thinkingLevelMap` and `compat.supportsReasoningEffort`.
+- `test/generate-models-strict.test.ts`: added an offline generator fixture covering `glm-5.3-turbo`, `glm-5.3-xl`, and `glm-5.3-anything-else`.
+
+### Why
+
+- Unsupported GLM-5.3 variants must not receive reasoning metadata or be forced into enabled thinking; only the validated base, Flash, and Highspeed variants should use the always-enabled Z.AI thinking path.
+
+### Why an extension could not handle it
+
+- Generated model capability metadata and OpenAI Completions request serialization are implemented inside the AI package.
+
+### Expected merge conflict zones
+
+- LOW: `api/openai-completions.ts` and the GLM-5.3 generator regression coverage.
+
 ## 2026-08-26 - Coalesce adjacent Anthropic user turns
 
 ### What changed

packages/ai/src/providers/data/.manifest.json

@@ -1 +1 @@
-{"schemaVersion":3,"generatedAt":"2026-08-29T04:42:44.820Z","structureHash":"3c455a74fcf9a66fed76da0baeafab7eec982373b540beefb9d588b18318c185","files":{"alibaba-token-plan.json":"9f9e8f88b1faa2e4134a0d8bcda600531d4d044d64b7ec3d2373f1db46dcf452","amazon-bedrock.json":"eb1770cb9edee5efb9646da27043a83cc80fa5e186dce60fcafa681222b794c2","ant-ling.json":"4979fe79d99ed97382d7ce40170c6132e932e77906de1eb54464c042e7d63633","anthropic.json":"8c1136a83a3b1ff0e61ab461ee221b1de7680784e774a60e1520af14a44bb0f1","azure-openai-responses.json":"37d12c5cbfc29bfd9a802a0516f5d3209320bd4aa8ef9cd7c219af208244afa3","baseten.json":"e4d1ec84e8180a53e21c1511b8d78e497784515c33ba8397382ac3450c8b126e","cerebras.json":"38272a7525d6680c71bb4bb3c4085e871768b400832e2013b0e7fd92f8fd4f05","cloudflare-ai-gateway.json":"38d37fefcc6437904f42437881e782d31a79982942c25a48cb8fe9a255afbaa2","cloudflare-workers-ai.json":"0ed67191adec10707a8ce1ed2c99a949dd06ec537de15d6b57e830575d248b86","deepseek.json":"e3106724723fd82d48a3ff685da9908a3a7b7b4fd84cc3cfe97a6a23067b6029","fireworks.json":"21dbcb60719ae099beda44a8fe03dc725b00bf3767c32bd1027d288061d30ffe","github-copilot.json":"61dec9e46df61a9b0bf6c7618651f97bfa9ae609d4c216b5725f014942ca3001","google-vertex.json":"fb7586fc955da7b7d7d11d641954f496fbbb42cf3e4cbba66d7116ea3b37d5ed","google.json":"b5572904165d37ab594eafc3ab7f29774e35747b5dd211144ec01532db9217c1","groq.json":"f1eb899453e70b5e4479fe6eb78869ee99121fa6d4a019a5a05ab42d2ada8992","huggingface.json":"7616a851e53d254e375b760abe2b6148e8961a102bd5c7af31829a0f2dc8e5c5","kimi-coding.json":"88150f8970f83e17968b02f3b33fac803b96ebf1474e76437689b7d0f1b69363","minimax-cn.json":"65d37e55a6d9585855f94b3fb7739308d107fe248ebd9e1001d1c5d4606ad2c4","minimax.json":"c7dc5765995ae77242916430929c7cc9a782589f7e483e13f7e01663190d028e","mistral.json":"e3a109306644bf2f2fa38fece50b2d3ce1dbe3b87b1f1b9e2d70cce7eb5c0ea1","moonshotai-cn.json":"dbf1b71ee36812fe64b014177ff455809a8dd852bbbd584422645beb85635127","moonshotai.json":"2d7518b4769e97ac900366175c8b60ff86a02675bb6f6cc2fffcc6761fda6faa","nvidia.json":"62fa810bece1c90454b481db5a3ff78bc66dc28dab9c65a655397eb1dfe2de73","openai-codex.json":"4e4afe2929b2b20a751f2453b219deae4e5b9fd85e147bd0403601f183cfc15f","openai.json":"b014fb0676fec97264eb13b367b0ed73b99915f45410a3dd9a7397a0360c2a4b","opencode-go.json":"1b707b5fdaf0bbe9b436a2fa662b9ade1f1c7db0ffc0631916428547483acaac","opencode.json":"ab2bf29b6c0b097c9dca13faab7515365787da523e6e942b49002a2612b99d26","opengateway.json":"1702a5df9877110c4d71a6ff3760cd10b378bfbd32d6322c05ce59dc1cc0f8e4","openrouter.json":"569ee081f6d5e995a23f5bb5b900f872ba8479b61a95faf02c38a8410b71a21b","qwen-token-plan-cn.json":"e62dde34edb603376a1dc60b2a74a9c2605b9e81d81df55cc259796c45aa4793","qwen-token-plan-individual.json":"5649fd4075ada766ec65a7a162cd65f096a7ecc29960ace04b2e6a6a4a995a41","qwen-token-plan.json":"fe914313157ff41ab2f71977f0fdcfedb3a8a205d1092e3180f6799e5d7942fd","together.json":"c2b03c476d6b4d4059ed18a80f34d274c754dc84fe2591ae79fe2986512f2a51","vercel-ai-gateway.json":"df4810da53726a9465ec6ddb6c5ec6d66955552f6ca7aa5d01f18f3ffe02cbe7","xai.json":"9973aae022723732b8a7ea12107e6fb8b3f1d0e087ec6bbc2ee18f0820595e19","xiaomi-token-plan-ams.json":"568b9992cf320536a65826e26a414afa5c0d17f083a01af4bd96658d9dd3e705","xiaomi-token-plan-cn.json":"c423857d0bbb0b126394eaf1ebb2a91bf2d1193cc5a4fe77843ff0f4536a0a80","xiaomi-token-plan-sgp.json":"008f1eb41037e93c5ae478f12815cd67bac1978f131ffefa602775bb60b63fea","xiaomi.json":"37acf451bb4364bdca2ecea6b7d962590d9974a87e80cfe34bb800be6a95c13b","zai-coding-cn.json":"b3fe0a69833cb573ee73fa4ed3f620732b3193fc32f5433d2fa9971c2f5ce299","zai.json":"6634870e983bdae76dddbf2e721e972fbc4b5974467b4b99edaab5bd9c1c1e3f"}}
+{"schemaVersion":3,"generatedAt":"2026-08-29T12:46:36.807Z","structureHash":"3c455a74fcf9a66fed76da0baeafab7eec982373b540beefb9d588b18318c185","files":{"alibaba-token-plan.json":"9f9e8f88b1faa2e4134a0d8bcda600531d4d044d64b7ec3d2373f1db46dcf452","amazon-bedrock.json":"eb1770cb9edee5efb9646da27043a83cc80fa5e186dce60fcafa681222b794c2","ant-ling.json":"4979fe79d99ed97382d7ce40170c6132e932e77906de1eb54464c042e7d63633","anthropic.json":"8c1136a83a3b1ff0e61ab461ee221b1de7680784e774a60e1520af14a44bb0f1","azure-openai-responses.json":"37d12c5cbfc29bfd9a802a0516f5d3209320bd4aa8ef9cd7c219af208244afa3","baseten.json":"e4d1ec84e8180a53e21c1511b8d78e497784515c33ba8397382ac3450c8b126e","cerebras.json":"38272a7525d6680c71bb4bb3c4085e871768b400832e2013b0e7fd92f8fd4f05","cloudflare-ai-gateway.json":"38d37fefcc6437904f42437881e782d31a79982942c25a48cb8fe9a255afbaa2","cloudflare-workers-ai.json":"0ed67191adec10707a8ce1ed2c99a949dd06ec537de15d6b57e830575d248b86","deepseek.json":"e3106724723fd82d48a3ff685da9908a3a7b7b4fd84cc3cfe97a6a23067b6029","fireworks.json":"21dbcb60719ae099beda44a8fe03dc725b00bf3767c32bd1027d288061d30ffe","github-copilot.json":"61dec9e46df61a9b0bf6c7618651f97bfa9ae609d4c216b5725f014942ca3001","google-vertex.json":"fb7586fc955da7b7d7d11d641954f496fbbb42cf3e4cbba66d7116ea3b37d5ed","google.json":"b5572904165d37ab594eafc3ab7f29774e35747b5dd211144ec01532db9217c1","groq.json":"f1eb899453e70b5e4479fe6eb78869ee99121fa6d4a019a5a05ab42d2ada8992","huggingface.json":"7616a851e53d254e375b760abe2b6148e8961a102bd5c7af31829a0f2dc8e5c5","kimi-coding.json":"88150f8970f83e17968b02f3b33fac803b96ebf1474e76437689b7d0f1b69363","minimax-cn.json":"65d37e55a6d9585855f94b3fb7739308d107fe248ebd9e1001d1c5d4606ad2c4","minimax.json":"c7dc5765995ae77242916430929c7cc9a782589f7e483e13f7e01663190d028e","mistral.json":"e3a109306644bf2f2fa38fece50b2d3ce1dbe3b87b1f1b9e2d70cce7eb5c0ea1","moonshotai-cn.json":"dbf1b71ee36812fe64b014177ff455809a8dd852bbbd584422645beb85635127","moonshotai.json":"2d7518b4769e97ac900366175c8b60ff86a02675bb6f6cc2fffcc6761fda6faa","nvidia.json":"62fa810bece1c90454b481db5a3ff78bc66dc28dab9c65a655397eb1dfe2de73","openai-codex.json":"4e4afe2929b2b20a751f2453b219deae4e5b9fd85e147bd0403601f183cfc15f","openai.json":"b014fb0676fec97264eb13b367b0ed73b99915f45410a3dd9a7397a0360c2a4b","opencode-go.json":"1b707b5fdaf0bbe9b436a2fa662b9ade1f1c7db0ffc0631916428547483acaac","opencode.json":"ab2bf29b6c0b097c9dca13faab7515365787da523e6e942b49002a2612b99d26","opengateway.json":"1702a5df9877110c4d71a6ff3760cd10b378bfbd32d6322c05ce59dc1cc0f8e4","openrouter.json":"065d4eb3817363dc217cc9bea7130a0cf30c0ad934819b63a5a4c29ef9919943","qwen-token-plan-cn.json":"e62dde34edb603376a1dc60b2a74a9c2605b9e81d81df55cc259796c45aa4793","qwen-token-plan-individual.json":"5649fd4075ada766ec65a7a162cd65f096a7ecc29960ace04b2e6a6a4a995a41","qwen-token-plan.json":"fe914313157ff41ab2f71977f0fdcfedb3a8a205d1092e3180f6799e5d7942fd","together.json":"c2b03c476d6b4d4059ed18a80f34d274c754dc84fe2591ae79fe2986512f2a51","vercel-ai-gateway.json":"df4810da53726a9465ec6ddb6c5ec6d66955552f6ca7aa5d01f18f3ffe02cbe7","xai.json":"9973aae022723732b8a7ea12107e6fb8b3f1d0e087ec6bbc2ee18f0820595e19","xiaomi-token-plan-ams.json":"568b9992cf320536a65826e26a414afa5c0d17f083a01af4bd96658d9dd3e705","xiaomi-token-plan-cn.json":"c423857d0bbb0b126394eaf1ebb2a91bf2d1193cc5a4fe77843ff0f4536a0a80","xiaomi-token-plan-sgp.json":"008f1eb41037e93c5ae478f12815cd67bac1978f131ffefa602775bb60b63fea","xiaomi.json":"37acf451bb4364bdca2ecea6b7d962590d9974a87e80cfe34bb800be6a95c13b","zai-coding-cn.json":"d3c969020a7ab978497837a3301c7bca364d704082832221cfe61e6596333e56","zai.json":"f42790c77fb4681a656897a9d860b939916c0ddf979a5f16856bce9573cdf64e"}}