@sayknow-cli/ai 0.5.6 → 0.5.8

package/CHANGELOG.md

@@ -2,7 +2,17 @@
 
 ## [Unreleased]
 
-## [0.5.6] - 2026-08-28
+## [0.5.8] - 2026-09-10
+
+## [0.5.7] - 2026-09-10
+
+### Added
+
+- Bundled `openai-codex/gpt-6-astra` (ported from upstream #5294) so eligible ChatGPT accounts can select GPT-6 Astra without waiting for authenticated Codex discovery. The reviewed snapshot records the 272K prompt budget, 128K output cap, text-and-image input, websocket preference, low-through-max reasoning efforts, freeform `apply_patch`, and standard published pricing ($10/$50, cache read $1, cache write $12.50). Upstream's above-272K pricing tier is intentionally omitted: this fork's cost engine has no long-context tier support, so the field would be dead data. The GPT model-id parser and the generated `apply_patch` policy now recognize the `astra` variant (GPT-6), and the catalog was regenerated, which also picks up the current models.dev snapshot across gateway providers.
+
+### Fixed
+
+- Bumped the advertised Claude Code client version from 2.1.219 to 2.1.267. Anthropic gates newer models (e.g. the 5.1 line) on the advertised client version and rejected requests with HTTP 400 `claude_code_version_too_old` ("version 2.1.251 or newer is required"), so those models were unusable through the Claude Code OAuth path.
 
 ## [0.5.3] - 2026-08-28
 

package/dist/types/model-thinking.d.ts

@@ -27,6 +27,13 @@ export declare function enrichModelThinking<TApi extends Api>(model: ApiModel<TA
  * canonical rules, replacing any existing `thinking`.
  */
 export declare function refreshModelThinking<TApi extends Api>(model: ApiModel<TApi>): ApiModel<TApi>;
+/**
+ * Extract the GLM generation from a Zhipu/ZAI model id: `glm-5.3` and
+ * `glm-5.3-flash` -> 5.3, `glm-5` and `glm-5-turbo` -> 5, `glm-4.7` -> 4.7.
+ * Returns undefined for ids that are not a plain GLM generation, including the
+ * vision line (`glm-5v-turbo`), so those never inherit text-model corrections.
+ */
+export declare function glmGeneration(modelId: string): number | undefined;
 /**
  * Apply upstream metadata corrections to a mutable array of models.
  *

package/dist/types/providers/anthropic.d.ts

@@ -50,7 +50,7 @@ export declare function isAnthropicThinkingBlockMutationError(error: unknown): b
  * than only the latest one.
  */
 export declare function isAnthropicThinkingSignatureInvalidError(error: unknown): boolean;
-export declare const claudeCodeVersion = "2.1.219";
+export declare const claudeCodeVersion = "2.1.267";
 export declare const claudeCodeEntrypoint = "sdk-cli";
 export declare const claudeToolPrefix: string;
 export declare const claudeCodeSystemInstruction = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/ai",
-	"version": "0.5.6",
+	"version": "0.5.8",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://sayknow-cli.com",
 	"author": "jaybeyond",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@sayknow-cli/utils": "0.5.6",
+		"@sayknow-cli/utils": "0.5.8",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/model-thinking.ts

@@ -64,6 +64,7 @@ type SemVer = {
 type GeminiKind = "pro" | "flash";
 type AnthropicKind = "opus" | "sonnet" | "fable";
 type OpenAIVariant =
+	| "astra"
 	| "base"
 	| "codex"
 	| "codex-max"
@@ -190,6 +191,24 @@ export function refreshModelThinking<TApi extends Api>(model: ApiModel<TApi>): A
 	return { ...model, thinking: inferModelThinking(model) };
 }
 
+/**
+ * Extract the GLM generation from a Zhipu/ZAI model id: `glm-5.3` and
+ * `glm-5.3-flash` -> 5.3, `glm-5` and `glm-5-turbo` -> 5, `glm-4.7` -> 4.7.
+ * Returns undefined for ids that are not a plain GLM generation, including the
+ * vision line (`glm-5v-turbo`), so those never inherit text-model corrections.
+ */
+export function glmGeneration(modelId: string): number | undefined {
+	const match = modelId.toLowerCase().match(/^glm-(\d+)(?:\.(\d+))?(?![\d.v])/);
+	if (!match) return undefined;
+	const [, major, minor] = match;
+	return minor === undefined ? Number(major) : Number(`${major}.${minor}`);
+}
+
+function glmGenerationAtLeast(modelId: string, minimum: number): boolean {
+	const generation = glmGeneration(modelId);
+	return generation !== undefined && generation >= minimum;
+}
+
 /**
  * Apply upstream metadata corrections to a mutable array of models.
  *
@@ -425,13 +444,31 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	if (parsedModel.family === "openai") {
 		applyOpenAICatalogPolicy(model, parsedModel);
 	}
-	// GLM-5.2 (Zhipu/ZAI): ships a 1M lossless context window, but the bundled
-	// catalog copied GLM-5.1's 200K and that stale value survives generate-models
-	// (provider-scoped models bypass the models.dev refresh in applyGlobalModelsDevFallback).
-	// Pin to the true 1M so context-cap / auto-compaction thresholds aren't tripped ~5x early.
-	if (model.provider === "zai" && model.id === "glm-5.2") {
+	// GLM-5.2 and newer (Zhipu/ZAI): ship a 1M lossless context window, but the
+	// bundled catalog copied GLM-5.1's 200K and that stale value survives
+	// generate-models (provider-scoped models bypass the models.dev refresh in
+	// applyGlobalModelsDevFallback). Pin the true 1M so context-cap /
+	// auto-compaction thresholds aren't tripped ~5x early.
+	//
+	// Matched by generation rather than by exact id so the next GLM point
+	// release inherits the correct window instead of silently regressing to
+	// whatever the catalog happens to carry. The correction only raises a
+	// stale-low window: a future generation that genuinely ships more than 1M
+	// keeps its larger catalog value instead of being clamped down to 1M.
+	if (model.provider === "zai" && glmGenerationAtLeast(model.id, 5.2) && model.contextWindow < 1_000_000) {
 		model.contextWindow = 1_000_000;
 	}
+	// Alibaba Token Plan's qwen3.8-max-preview is routed through the OpenAI
+	// Responses API. models.dev recently started publishing the whole
+	// alibaba-token-plan provider, and its blanket chat-completions mapping wins
+	// the generate-models merge, so a regeneration silently reroutes this one
+	// paid model. Pin the routing that ships today and is covered by
+	// register-builtins' Responses lazy-path watchdog until the endpoint is
+	// re-verified; the catalog's other metadata (vision input, 131K output cap)
+	// is still inherited.
+	if (model.provider === "alibaba-token-plan" && model.id === "qwen3.8-max-preview") {
+		model.api = "openai-responses" as typeof model.api;
+	}
 	// MiniMax-M3: MiniMax exposes a 1M context tier, but usage beyond 512K is
 	// billed separately. Keep bundled/default metadata at the billing-safe 512K
 	// unless an explicit paid-tier contract is added.
@@ -468,7 +505,10 @@ function inferGeneratedApplyPatchToolType(
 	model: ApiModel<Api>,
 	parsedModel: ParsedModel,
 ): ApiModel<Api>["applyPatchToolType"] {
-	if (parsedModel.family !== "openai" || parsedModel.version.major !== 5) {
+	if (
+		parsedModel.family !== "openai" ||
+		(parsedModel.version.major !== 5 && !(parsedModel.version.major === 6 && parsedModel.variant === "astra"))
+	) {
 		return undefined;
 	}
 	if (model.provider === "openai" && model.api === "openai-responses") {
@@ -778,7 +818,7 @@ function parseAnthropicModel(modelId: string): AnthropicModel | null {
 
 function parseOpenAIModel(modelId: string): OpenAIModel | null {
 	const match =
-		/gpt-(\d+(?:\.\d+){0,2})(?:-(codex-spark|codex-mini|codex-max|codex|luna|mini|max|nano|sol|terra))?$/.exec(
+		/gpt-(\d+(?:\.\d+){0,2})(?:-(astra|codex-spark|codex-mini|codex-max|codex|luna|mini|max|nano|sol|terra))?$/.exec(
 			modelId,
 		);
 	if (!match) {