@oh-my-pi/pi-catalog 18.2.10 → 18.3.0

package/src/provider-models/descriptors.ts

@@ -61,6 +61,7 @@ import {
 	siliconflowModelManagerOptions,
 	singularityApiDevModelManagerOptions,
 	singularityApiTechModelManagerOptions,
+	stepfunModelManagerOptions,
 	syntheticModelManagerOptions,
 	togetherModelManagerOptions,
 	umansModelManagerOptions,
@@ -141,6 +142,7 @@ const MODEL_MANAGER_FACTORIES: Readonly<Partial<Record<KnownProvider, ModelManag
 	"siliconflow-cn": config => siliconflowCnModelManagerOptions(config),
 	"singularityapi-dev": config => singularityApiDevModelManagerOptions(config),
 	"singularityapi-tech": config => singularityApiTechModelManagerOptions(config),
+	stepfun: config => stepfunModelManagerOptions(config),
 	synthetic: config => syntheticModelManagerOptions(config),
 	together: config => togetherModelManagerOptions(config),
 	typesafe: config => typesafeModelManagerOptions(config),

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

@@ -4956,6 +4956,86 @@ export function yoloAutoModelManagerOptions(
 }
 
 // ---------------------------------------------------------------------------
+// 16.9 StepFun
+// ---------------------------------------------------------------------------
+
+/**
+ * StepFun discovery configuration: the API key plus optional base-URL and
+ * fetch overrides. Consumed by {@link stepfunModelManagerOptions}, and exported
+ * for extensions and tests that construct the manager directly.
+ */
+export interface StepfunModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+	fetch?: FetchImpl;
+}
+
+/** StepFun's `/v1/models` row shape beyond the generic OpenAI-compatible fields. */
+interface StepfunModelRecord extends OpenAICompatibleModelRecord {
+	/** Reasoning tiers the endpoint advertises for the model, e.g. `["low","medium","high"]`. */
+	reasoning_effort_support_list?: unknown;
+}
+
+/**
+ * Translate StepFun's per-model `reasoning_effort_support_list` into a ladder.
+ * Every advertised value that names an OMP tier maps verbatim, in OMP's tier
+ * order; a row advertising nothing (or only tiers this client does not know)
+ * resolves to no thinking, so the wire path never sends a `reasoning_effort`
+ * the endpoint rejects. Same shape as `mapOpenRouterThinking` for OpenRouter's
+ * `reasoning.supported_efforts`.
+ */
+function mapStepfunThinking(entry: StepfunModelRecord): ThinkingConfig | undefined {
+	const advertised = Array.isArray(entry.reasoning_effort_support_list)
+		? entry.reasoning_effort_support_list.filter((value): value is string => typeof value === "string")
+		: [];
+	const efforts = THINKING_EFFORTS.filter(effort => advertised.includes(effort));
+	return efforts.length === 0 ? undefined : { mode: "effort", efforts };
+}
+
+/**
+ * Whether a StepFun `/v1/models` id is a chat model omp can route. StepFun's
+ * roster interleaves its audio and image SKUs with the chat models; the
+ * exclusion policy itself lives in `runtime/behavior.kdl` (`exclude-models
+ * provider="stepfun"`), not here.
+ */
+export function isStepfunChatModelId(id: string): boolean {
+	const normalized = id.trim().toLowerCase();
+	if (!normalized) return false;
+	return !isExcludedModel("stepfun", normalized);
+}
+
+/**
+ * StepFun model manager: plain OpenAI-compatible chat completions at
+ * `api.stepfun.ai/v1`. A successful `/v1/models` snapshot is authoritative over
+ * the bundled seed rows (`providers/stepfun.kdl`), so a model StepFun retires
+ * leaves the picker instead of lingering as a dead seed row, while models added
+ * later become selectable without an omp release.
+ */
+export function stepfunModelManagerOptions(
+	config?: StepfunModelManagerConfig,
+): ModelManagerOptions<"openai-completions"> {
+	return createOpenAICompatibleModelManagerOptions({
+		api: "openai-completions",
+		providerId: "stepfun",
+		defaultBaseUrl: "https://api.stepfun.ai/v1",
+		config,
+		requireApiKey: true,
+		filterModel: (_entry, model) => isStepfunChatModelId(model.id),
+		mapModel: (entry, model, reference) => {
+			const mapped = mapWithBundledReference(entry, model, reference);
+			// A model StepFun ships later has no bundled reference, so it starts
+			// from the generic defaults (`reasoning: false`, no thinking) and
+			// `mergeDynamicModels` adds it verbatim — the endpoint's own tiers are
+			// then the only thing that can give it a reasoning dial.
+			if (reference) return mapped;
+			const thinking = mapStepfunThinking(entry);
+			return thinking === undefined ? mapped : { ...mapped, reasoning: true, thinking };
+		},
+		// Must live on the manager options, not only the KDL descriptor:
+		// `createModelManager()` prunes the bundled slice from this flag.
+		dynamicModelsAuthoritative: true,
+	});
+}
 
 // ---------------------------------------------------------------------------
 // 17. Qwen Portal
@@ -5936,6 +6016,9 @@ export interface GithubCopilotModelManagerConfig {
 	fetch?: FetchImpl;
 }
 
+// Copilot ids whose cached route predates the Responses pin: a cache written by
+// an older build still says openai-completions, which Copilot answers with 400
+// unsupported_api_for_model (#7096, #8807, #12901).
 const COPILOT_CACHE_INVALIDATED_MODEL_IDS = [
 	"gpt-6-astra",
 	"gpt-6-astra-1m",
@@ -5943,6 +6026,8 @@ const COPILOT_CACHE_INVALIDATED_MODEL_IDS = [
 	"grok-4.5-1m",
 	"grok-4.6",
 	"grok-4.6-1m",
+	"grok-4.7",
+	"grok-4.7-1m",
 	"mai-code-1-flash-picker",
 ];
 
@@ -6211,7 +6296,13 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
 													supportsReasoningEffort: false,
 												},
 											}
-										: {}),
+										: // The bundled row for an id whose route later moved to
+											// Responses/Messages still carries this chat-completions
+											// block, and `supportsReasoningEffort: false` suppresses
+											// the effort dial on a transport that supports it
+											// (grok-4.7, #12901). Compat is transport-scoped: let the
+											// rules resolve it for the route actually in use.
+											{ compat: undefined }),
 								}
 							: {
 									...defaults,

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;

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. */
@@ -519,17 +520,17 @@ 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;
 	/**
@@ -1163,6 +1164,16 @@ 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 +1268,15 @@ 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;