@sayknow-cli/ai 0.5.25 → 0.6.0

package/CHANGELOG.md

@@ -2,7 +2,16 @@
 
 ## [Unreleased]
 
-## [0.5.25] - 2026-09-21
+## [0.6.0] - 2026-09-24
+### Fixed
+- Retire `anthropic/claude-3-haiku-20240307`, `claude-3-5-sonnet-20240620` and `claude-3-5-sonnet-20241022`: the Anthropic API answers 404 `not_found_error` for all three (measured 2026-09-23 with a forced tool call). Left in the catalog, the cheapest of them was the first pick for typed decisions on every registry with an Anthropic key.
+- The on-disk model cache now filters retired keys on read. A row written before a retirement carried the dead entry for up to the TTL and the catalog merge brought it straight back.
+
+## [0.5.26] - 2026-09-23
+### Fixed
+- API-key `/login` now upserts a second key instead of replacing the whole provider pool. Same-key re-login still reuses the existing row.
+- Bundle `xai/grok-4.7` and use it as the xAI default without changing explicitly saved model choices. Preserve its `low`–`xhigh` reasoning range during catalog generation. Cost metadata records the standard (<200k prompt tokens) tariff; the existing cost engine does not represent long-context price tiers.
+- Bundle `anthropic/claude-opus-5-5` (1M context, 128K output, vision, adaptive thinking). The provider default stays `claude-sonnet-4-6`: promoting a Sonnet default to an Opus would raise per-token cost for every user, so selecting Opus 5.5 stays explicit. Cost metadata records the standard tariff ($4/$20, cache read $0.20 at the Opus-5.5-specific 0.05x multiplier, 5-minute cache write $5); the cost engine represents neither the fast-mode tier ($8/$40) nor long-context tiers. Anthropic's server-side default effort (`medium`) is deliberately not encoded as `thinking.defaultLevel`, because `inferDefaultEffort` has no Anthropic rule and catalog regeneration would strip it — every sibling Opus entry omits it for the same reason.
 
 ## [0.5.12] - 2026-09-15
 ### Added

package/dist/types/auth-storage.d.ts

@@ -226,7 +226,7 @@ export interface AuthCredentialStore {
     }): Promise<void>;
     /**
      * Optional async write hook for upserting a single credential. When present,
-     * `AuthStorage.#upsertOAuthCredential` routes through this instead of the
+     * `AuthStorage.#upsertStoredCredential` routes through this instead of the
      * sync `upsertAuthCredentialForProvider`. `RemoteAuthCredentialStore` uses
      * it to send the upsert to the broker via `POST /v1/credential`.
      *
@@ -236,10 +236,11 @@ export interface AuthCredentialStore {
     upsertAuthCredentialRemote?(provider: string, credential: AuthCredential): Promise<StoredAuthCredential[]>;
     upsertAuthCredentialRemoteIfAbsent?(provider: string, credential: AuthCredential): Promise<AuthCredentialIfAbsentResult>;
     /**
-     * Optional async write hook for replace-all semantics (e.g. API-key login
-     * overwriting any previous keys for the same provider). When present,
+     * Optional async write hook for replace-all semantics (logout-style
+     * overwrite of every stored credential for a provider). When present,
      * `AuthStorage.set` routes through this instead of the sync
-     * `replaceAuthCredentialsForProvider`.
+     * `replaceAuthCredentialsForProvider`. API-key login no longer uses this
+     * path: it upserts so a second account is added instead of wiping the pool.
      */
     replaceAuthCredentialsRemote?(provider: string, credentials: AuthCredential[]): Promise<StoredAuthCredential[]>;
     /**

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

@@ -1,4 +1,4 @@
-export declare const RETIRED_MODEL_KEYS: readonly ["google-antigravity/gemini-3.1-pro-high"];
+export declare const RETIRED_MODEL_KEYS: readonly ["google-antigravity/gemini-3.1-pro-high", "anthropic/claude-3-haiku-20240307", "anthropic/claude-3-5-sonnet-20240620", "anthropic/claude-3-5-sonnet-20241022"];
 export declare function isRetiredModelKey(provider: string, modelId: string): boolean;
 export declare function isRetiredModel(model: {
     provider: string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/ai",
-	"version": "0.5.25",
+	"version": "0.6.0",
 	"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": "0.5.25",
+		"@sayknow-cli/utils": "0.6.0",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/auth-storage.ts

@@ -387,7 +387,7 @@ export interface AuthCredentialStore {
 	markCredentialSuspect?(credentialId: number, opts?: { signal?: AbortSignal }): Promise<void>;
 	/**
 	 * Optional async write hook for upserting a single credential. When present,
-	 * `AuthStorage.#upsertOAuthCredential` routes through this instead of the
+	 * `AuthStorage.#upsertStoredCredential` routes through this instead of the
 	 * sync `upsertAuthCredentialForProvider`. `RemoteAuthCredentialStore` uses
 	 * it to send the upsert to the broker via `POST /v1/credential`.
 	 *
@@ -400,10 +400,11 @@ export interface AuthCredentialStore {
 		credential: AuthCredential,
 	): Promise<AuthCredentialIfAbsentResult>;
 	/**
-	 * Optional async write hook for replace-all semantics (e.g. API-key login
-	 * overwriting any previous keys for the same provider). When present,
+	 * Optional async write hook for replace-all semantics (logout-style
+	 * overwrite of every stored credential for a provider). When present,
 	 * `AuthStorage.set` routes through this instead of the sync
-	 * `replaceAuthCredentialsForProvider`.
+	 * `replaceAuthCredentialsForProvider`. API-key login no longer uses this
+	 * path: it upserts so a second account is added instead of wiping the pool.
 	 */
 	replaceAuthCredentialsRemote?(provider: string, credentials: AuthCredential[]): Promise<StoredAuthCredential[]>;
 	/**
@@ -1550,7 +1551,7 @@ export class AuthStorage {
 		};
 	}
 
-	async #upsertOAuthCredential(provider: string, credential: OAuthCredential): Promise<void> {
+	async #upsertStoredCredential(provider: string, credential: AuthCredential): Promise<void> {
 		const stored = this.#store.upsertAuthCredentialRemote
 			? await this.#store.upsertAuthCredentialRemote(provider, credential)
 			: this.#store.upsertAuthCredentialForProvider(provider, credential);
@@ -1562,6 +1563,17 @@ export class AuthStorage {
 		this.#invalidateUsageCacheForProvider(provider);
 	}
 
+	async #dropStoredApiKeysForProvider(provider: string): Promise<void> {
+		const remaining = this.#getCredentialsForProvider(provider).filter(credential => credential.type !== "api_key");
+		if (remaining.length === this.#getCredentialsForProvider(provider).length) return;
+		await this.set(provider, remaining);
+	}
+
+	async #upsertOAuthCredential(provider: string, credential: OAuthCredential): Promise<void> {
+		await this.#dropStoredApiKeysForProvider(provider);
+		await this.#upsertStoredCredential(provider, credential);
+	}
+
 	#invalidateUsageCacheForProvider(provider: string): void {
 		this.#usageRequestInFlight.clear();
 		this.#usageReportsInFlight.clear();
@@ -1696,7 +1708,7 @@ export class AuthStorage {
 		let credentials: OAuthCredentials;
 		const saveApiKeyCredential = async (apiKey: string): Promise<void> => {
 			const newCredential: ApiKeyCredential = { type: "api_key", key: apiKey };
-			await this.set(provider, newCredential);
+			await this.#upsertStoredCredential(provider, newCredential);
 		};
 		const manualCodeInput = () => ctrl.onPrompt({ message: "Paste the authorization code (or full redirect URL):" });
 		switch (provider) {
@@ -2048,13 +2060,6 @@ export class AuthStorage {
 			}
 		}
 		const newCredential: OAuthCredential = { type: "oauth", ...credentials };
-		if (provider === "xai") {
-			const existingOAuthCredentials = this.#getCredentialsForProvider(provider).filter(
-				(credential): credential is OAuthCredential => credential.type === "oauth",
-			);
-			await this.set(provider, [...existingOAuthCredentials, newCredential]);
-			return;
-		}
 		await this.#upsertOAuthCredential(provider, newCredential);
 	}
 

package/src/model-cache.ts

@@ -4,6 +4,7 @@
  */
 import { Database } from "bun:sqlite";
 import { getModelDbPath } from "@sayknow-cli/utils/dirs";
+import { isRetiredModelKey } from "./model-retirements";
 import type { Api, Model } from "./types";
 
 const CACHE_SCHEMA_VERSION = 3;
@@ -96,7 +97,12 @@ export function readModelCache<TApi extends Api>(
 		if (!row || row.version !== CACHE_SCHEMA_VERSION) {
 			return null;
 		}
-		const models = JSON.parse(row.models) as Model<TApi>[];
+		// A row written before a model was retired still carries it for up to the TTL;
+		// the bundled catalog filters retirements, so the cache must too or the merge
+		// brings the dead entry straight back.
+		const models = (JSON.parse(row.models) as Model<TApi>[]).filter(
+			model => !isRetiredModelKey(model.provider ?? providerId, model.id),
+		);
 		const ageMs = now() - row.updated_at;
 		const fresh = Number.isFinite(ageMs) && ageMs >= 0 && ageMs <= ttlMs;
 		return {

package/src/model-retirements.ts

@@ -1,6 +1,18 @@
-// Retired from advertised catalogs because Cloud Code Assist rejects live calls
-// with HTTP 400. The callable high-thinking path is gemini-3.1-pro-low:high.
-export const RETIRED_MODEL_KEYS = ["google-antigravity/gemini-3.1-pro-high"] as const;
+// Retired from advertised catalogs because the provider rejects live calls.
+//
+// google-antigravity/gemini-3.1-pro-high: Cloud Code Assist answers HTTP 400. The
+// callable high-thinking path is gemini-3.1-pro-low:high.
+//
+// anthropic/claude-3-*: the Anthropic API answers 404 `not_found_error` for all three
+// (measured 2026-09-23 with a forced tool call). Left in the catalog, the cheapest of
+// them ($0.25/MTok) was the first pick for typed decisions, so the "small hosted model"
+// fallback silently answered nothing on every registry that had an Anthropic key.
+export const RETIRED_MODEL_KEYS = [
+	"google-antigravity/gemini-3.1-pro-high",
+	"anthropic/claude-3-haiku-20240307",
+	"anthropic/claude-3-5-sonnet-20240620",
+	"anthropic/claude-3-5-sonnet-20241022",
+] as const;
 
 const RETIRED_MODEL_KEY_SET = new Set<string>(RETIRED_MODEL_KEYS);
 

package/src/model-thinking.ts

@@ -475,12 +475,12 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	if (model.provider !== "opencode-go" && model.id === "minimax-m3") {
 		model.contextWindow = 512_000;
 	}
-	// xAI Grok 4.5/4.6: official reasoning docs (docs.x.ai) expose
-	// low/medium/high, with xhigh only on grok-4.6+. Catalog snapshots and
+	// xAI Grok 4.5/4.6/4.7: official reasoning docs (docs.x.ai) expose
+	// low/medium/high, with xhigh on the reviewed 4.6 and 4.7 models. Catalog snapshots and
 	// openai-compat historically clamped everything to high because
 	// supportsReasoningEffort was false for Grok; pin the documented ranges so
 	// profile suffixes and requireSupportedEffort stay honest across regen.
-	if (model.provider === "xai" && /^grok-4\.6(?:$|[-.])/.test(model.id)) {
+	if (model.provider === "xai" && /^grok-4\.(?:6|7)(?:$|[-.])/.test(model.id)) {
 		if (model.thinking) {
 			model.thinking = {
 				...model.thinking,

package/src/models.json

@@ -5005,6 +5005,31 @@
 				"maxLevel": "max"
 			}
 		},
+		"claude-opus-5-5": {
+			"id": "claude-opus-5-5",
+			"name": "Anthropic Opus 5.5",
+			"api": "anthropic-messages",
+			"provider": "anthropic",
+			"baseUrl": "https://api.anthropic.com",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 4,
+				"output": 20,
+				"cacheRead": 0.2,
+				"cacheWrite": 5
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 128000,
+			"thinking": {
+				"mode": "anthropic-adaptive",
+				"minLevel": "minimal",
+				"maxLevel": "max"
+			}
+		},
 		"claude-sonnet-4-0": {
 			"id": "claude-sonnet-4-0",
 			"name": "Anthropic Sonnet 4 (latest)",
@@ -90861,6 +90886,32 @@
 				"defaultLevel": "high"
 			}
 		},
+		"grok-4.7": {
+			"id": "grok-4.7",
+			"name": "Grok 4.7",
+			"api": "openai-completions",
+			"provider": "xai",
+			"baseUrl": "https://api.x.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 2,
+				"output": 6,
+				"cacheRead": 0.5,
+				"cacheWrite": 0
+			},
+			"contextWindow": 500000,
+			"maxTokens": 500000,
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "low",
+				"maxLevel": "xhigh",
+				"defaultLevel": "high"
+			}
+		},
 		"grok-beta": {
 			"id": "grok-beta",
 			"name": "Grok Beta",

package/src/provider-models/descriptors.ts

@@ -166,7 +166,7 @@ export const PROVIDER_DESCRIPTORS: readonly ProviderDescriptor[] = [
 		config => fuguModelManagerOptions(config),
 		catalog("Sakana Fugu", ["FUGU_API_KEY"]),
 	),
-	descriptor("xai", "grok-4.6", config => xaiModelManagerOptions(config)),
+	descriptor("xai", "grok-4.7", config => xaiModelManagerOptions(config)),
 	catalogDescriptor(
 		"deepseek",
 		"deepseek-v4-pro",