@oh-my-pi/pi-catalog 18.2.11 → 18.3.1

package/src/provider-models/cache-provider-id.ts

@@ -175,6 +175,11 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 			const baseUrl = options.baseUrl ?? getDefaultModelDiscoveryBaseUrl(providerId)!;
 			return `vllm:models-v2:${Bun.hash(baseUrl).toString(36)}`;
 		}
+		case "devin":
+			// v2: rows cached before Fusion pairings carried their lead uid as
+			// `requestModelId` send the composite uid and fail with
+			// `permission_denied: no API providers are available`.
+			return "devin:models-v2";
 		default:
 			return providerId;
 	}

package/src/provider-models/openai-compat.ts

@@ -4710,8 +4710,8 @@ export function sakanaModelManagerOptions(config?: SakanaModelManagerConfig): Mo
 
 const AIAND_DEFAULT_BASE_URL = "https://api.aiand.com/v1";
 
-/** `reasoning_efforts` wire values ai& reports, mapped onto pi effort levels. */
-const AIAND_EFFORT_BY_WIRE_VALUE: Record<string, Effort> = {
+/** Effort wire values discovery endpoints report (`reasoning_efforts`, `thinking`), mapped onto pi effort levels. */
+const EFFORT_BY_WIRE_VALUE: Record<string, Effort> = {
 	minimal: Effort.Minimal,
 	low: Effort.Low,
 	medium: Effort.Medium,
@@ -4728,18 +4728,23 @@ function normalizeAiandBaseUrl(baseUrl: string | undefined): string {
 
 const AIAND_STATIC_MODEL_IDS = seedModels("aiand").map(model => model.id);
 
-function mapAiandThinking(entry: OpenAICompatibleModelRecord): ThinkingConfig | undefined {
-	const efforts = Array.isArray(entry.reasoning_efforts)
-		? entry.reasoning_efforts.flatMap(value =>
-				typeof value === "string" && AIAND_EFFORT_BY_WIRE_VALUE[value] ? [AIAND_EFFORT_BY_WIRE_VALUE[value]] : [],
+/** Parse a discovered effort list, dropping unknown wire values; empty when absent or unrecognized. */
+function parseWireEfforts(value: unknown): Effort[] {
+	return Array.isArray(value)
+		? value.flatMap(item =>
+				typeof item === "string" && EFFORT_BY_WIRE_VALUE[item] ? [EFFORT_BY_WIRE_VALUE[item]] : [],
 			)
 		: [];
+}
+
+function mapAiandThinking(entry: OpenAICompatibleModelRecord): ThinkingConfig | undefined {
+	const efforts = parseWireEfforts(entry.reasoning_efforts);
 	if (efforts.length === 0) {
 		return undefined;
 	}
 	const defaultLevel =
 		typeof entry.reasoning_effort_default === "string"
-			? AIAND_EFFORT_BY_WIRE_VALUE[entry.reasoning_effort_default]
+			? EFFORT_BY_WIRE_VALUE[entry.reasoning_effort_default]
 			: undefined;
 	return {
 		mode: "effort",
@@ -4896,8 +4901,26 @@ function mapYoloAutoModel(
 	reference: ModelSpec<"openai-completions"> | undefined,
 ): ModelSpec<"openai-completions"> {
 	const model = mapWithBundledReference(entry, defaults, reference);
+	// `/v1/models` advertises each reasoning model's accepted effort ladder in
+	// `thinking`; it is authoritative over the seed/reference ladder. Rows
+	// without the field (e.g. `yolo-small`) keep the reference surface.
+	const efforts = parseWireEfforts(entry.thinking);
+	const thinking: ThinkingConfig | undefined =
+		efforts.length > 0
+			? {
+					...model.thinking,
+					mode: model.thinking?.mode ?? "effort",
+					efforts,
+					defaultLevel:
+						model.thinking?.defaultLevel && efforts.includes(model.thinking.defaultLevel)
+							? model.thinking.defaultLevel
+							: undefined,
+				}
+			: model.thinking;
 	return {
 		...model,
+		...(efforts.length > 0 && { reasoning: true }),
+		thinking,
 		// Flat-rate and no-store are provider-wide: they must win whether the
 		// reference came from the yolo bundle, the global index, or nowhere.
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },

package/src/provider-models/special.ts

@@ -77,7 +77,8 @@ export function openaiCodexModelManagerOptions(
 
 /**
  * Merge complete per-account Codex catalogs into one authoritative list,
- * deduped by model id (first account to expose an id wins).
+ * deduped by model id. The first account to expose an id supplies its spec;
+ * account access is merged from every account whose catalog lists that id.
  *
  * Returns `null` when any account's fetch failed transiently, so a partial list
  * cannot replace the previous or bundled authoritative catalog. An account
@@ -102,7 +103,15 @@ function unionCodexModels(
 		}
 		catalogs++;
 		for (const model of result.models) {
-			if (!byId.has(model.id)) byId.set(model.id, model);
+			const existing = byId.get(model.id);
+			if (!existing) {
+				byId.set(model.id, model);
+			} else if (model.accountAccess) {
+				byId.set(model.id, {
+					...existing,
+					accountAccess: { ...existing.accountAccess, ...model.accountAccess },
+				});
+			}
 		}
 	}
 	return catalogs > 0 ? [...byId.values()] : null;
@@ -336,6 +345,7 @@ export function devinModelManagerOptions(config: DevinModelManagerConfig = {}):
 	const staticModels = seedModels<"devin-agent">("devin");
 	return {
 		providerId: "devin",
+		cacheProviderId: resolveModelCacheProviderId("devin"),
 		// A configured host serves its own Cascade deployment; keep the seed on it.
 		staticModels:
 			baseUrl === undefined || baseUrl === DEVIN_DEFAULT_BASE_URL

package/src/types.ts

@@ -20,7 +20,8 @@ export type KnownApi =
 	| "ollama-chat"
 	| "cursor-agent"
 	| "gitlab-duo-agent"
-	| "devin-agent";
+	| "devin-agent"
+	| "apple-foundation-models";
 export type Api = KnownApi | (string & {});
 
 /** Catalog kinds used to isolate role-specific runners from session chat models. */
@@ -478,6 +479,13 @@ export interface OpenAICompat {
 	 * sent as the top-level `reasoning.effort`.
 	 */
 	supportsConfigurationUpdate?: boolean;
+	/**
+	 * Whether the Responses WebSocket accepts `response.steer`, which queues user
+	 * input into the in-flight response (GPT-6 family). Default: rule-detected.
+	 * Set `false` for proxies that reject the event; steering then waits for the
+	 * next request boundary.
+	 */
+	supportsSteering?: boolean;
 	/** Whether streamed reasoning deltas for the same field may repeat the full cumulative text snapshot. Default: false. */
 	reasoningDeltasMayBeCumulative?: boolean;
 	/** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
@@ -519,17 +527,17 @@ export interface AnthropicCompat {
 	/** Whether thinking requests may include `context_management` and its beta header. Default: true. */
 	supportsContextManagement?: boolean;
 	/**
-	 * Whether the model lineage supports Anthropic server-side compaction
-	 * (`compact-2026-01-12`: the `compact_20260112` edit and replayed
-	 * `compaction` blocks). Rule-owned per model line; the beta covers the
-	 * adaptive-thinking generation onward and rejects older lines. Default: false.
+	 * Whether the model and host support Anthropic on-demand compaction
+	 * (`compact-2026-09-04` requests and signed replay). Enabled on Opus 4.6+,
+	 * Sonnet 4.6+, Fable/Mythos 5+ on supported hosts. Default: false.
 	 */
 	supportsServerCompaction?: boolean;
 	/**
 	 * Whether the model is served by the first-party Anthropic provider (its
 	 * default route is the official API). Rule-owned on the provider; the
 	 * compaction transport pairs it with a per-request effective-URL check
-	 * because reroutes leave it stale-true. Default: false.
+	 * because reroutes leave it stale-true. Vertex is selected by its provider
+	 * contract instead. Default: false.
 	 */
 	firstPartyProvider?: boolean;
 	/**
@@ -874,6 +882,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "strictResponsesPairing"
 			| "supportsImageDetailOriginal"
 			| "supportsConfigurationUpdate"
+			| "supportsSteering"
 			| "stripImageInput"
 			| "thinkingLoopGuard"
 			| "whenThinking"
@@ -913,6 +922,11 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
 	 * the item type with 400.
 	 */
 	supportsConfigurationUpdate: boolean;
+	/**
+	 * Whether the WebSocket transport may send `response.steer` to deliver user
+	 * input into the in-flight response. Rule-owned: GPT-6 family.
+	 */
+	supportsSteering: boolean;
 	/** Inject the `# Juice: 0 !important` developer item when reasoning is forced off (gpt-5.6+). */
 	requiresReasoningOffJuiceInstruction: boolean;
 	/**
@@ -1163,6 +1177,16 @@ export type ModelTokenizer =
 	| "kimi-k2"
 	| "glm5";
 
+/** One account's discovered entitlements on a model; see {@link Model.accountAccess}. */
+export interface ModelAccountAccess {
+	/**
+	 * Codex `available_access_programs.cyber`: cyber access programs this account
+	 * may request on the model (`standard`, `daybreak_blue`, `daybreak_red`).
+	 * Absent when the backend reported no program metadata.
+	 */
+	cyberPrograms?: readonly string[];
+}
+
 // Model interface for the unified model system
 export interface Model<TApi extends Api = Api> {
 	id: string;
@@ -1257,6 +1281,15 @@ export interface Model<TApi extends Api = Api> {
 	 * their single wire id) and on bundled snapshots that predate discovery.
 	 */
 	cursorMaxModeRoutes?: Readonly<Record<string, boolean>>;
+	/**
+	 * Per-account availability recorded by multi-account discovery: provider
+	 * account id (Codex: ChatGPT `chatgpt_account_id`) → that account's
+	 * entitlements on this model. An account appears only when its own catalog
+	 * lists the model, so credential selection can route account-gated models
+	 * (e.g. `gpt-daybreak-blue-latest`) straight to eligible accounts. Absent on
+	 * bundled/config rows and on single-account discovery.
+	 */
+	accountAccess?: Readonly<Record<string, ModelAccountAccess>>;
 	cost: ModelCost;
 	/** Premium Copilot requests charged per user-initiated request (defaults to 1). */
 	premiumMultiplier?: number;