@oh-my-pi/pi-catalog 18.12.5 → 18.13.0

package/CHANGELOG.md

@@ -2,6 +2,61 @@
 
 ## [Unreleased]
 
+## [18.13.0] - 2026-10-26
+
+### Breaking Changes
+
+- Renamed the `task-sibling-prompt-cache` axis to `shared-prompt-cache` (catalog field `sharedPromptCache`); it now also covers forks and explicit prompt-cache keys ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+
+### Added
+
+- Added the `requiresThinkingPresence` model compat (KDL `requires-thinking-presence`) for Anthropic-compatible endpoints that reject a thinking-mode request whose current turn has an assistant message without a `thinking` block ([#16739](https://github.com/can1357/oh-my-pi/pull/16739) by [@lz37](https://github.com/lz37))
+- Added `resolveAnthropicApiVersion()`: Anthropic model discovery sends API version 2026-10-21 to the official API when `PI_ANTHROPIC_API_VERSION=2026-10-21` is set ([#16706](https://github.com/can1357/oh-my-pi/pull/16706) by [@H4vC](https://github.com/H4vC))
+- OpenCode Go `muse-spark-1.3-contributor`, `muse-spark-1.2-contributor`, `gpt-5.6-luna`, `gpt-6-luna`, `grok-4.6` (Responses) and `deepseek-v4-pro`, `omen-alpha` (Completions) opt into shared prompt caching; lanes that warmed under any key stay off ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+- Added `CatalogPolicyInputs` so catalog-policy lookups accept a reduced model view ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+- Added a catalog conformance test that fails the build when a `bundle="always"` seed row is missing from the bundled `models.json`, so a stale-snapshot regeneration can no longer silently drop an engine the way `web/deepseek` was lost in 18.12.2 ([#16610](https://github.com/can1357/oh-my-pi/issues/16610), [#17053](https://github.com/can1357/oh-my-pi/pull/17053) by [@ouyangjiahong26](https://github.com/ouyangjiahong26))
+
+### Changed
+
+- llama.cpp login instructions now list both default local base URLs (`http://127.0.0.1:8080`, then `http://127.0.0.1:9931`) ([#16675](https://github.com/can1357/oh-my-pi/pull/16675) by [@bartbroere](https://github.com/bartbroere))
+- Devin requests now identify as Devin CLI 3000.11.205, the current stable release, instead of 3000.11.3 ([#16906](https://github.com/can1357/oh-my-pi/pull/16906) by [@will-bogusz](https://github.com/will-bogusz))
+- Google Antigravity login now uses PKCE and requests the `aicode` scope like the official client; existing logins keep working, and the new scope applies at the next `/login` ([#17012](https://github.com/can1357/oh-my-pi/pull/17012) by [@apoc](https://github.com/apoc))
+- Antigravity requests identify as client 2.26.0 when the release manifest is unreachable (previously 2.19.1), and model discovery sends the account's project like the official client ([#17012](https://github.com/can1357/oh-my-pi/pull/17012) by [@apoc](https://github.com/apoc))
+- Antigravity's `User-Agent` now reports the native Antigravity build for the host OS (Linux `linux`/`amd64`, Windows `windows`/`amd64`) instead of always claiming macOS `darwin`/`arm64` ([#17012](https://github.com/can1357/oh-my-pi/pull/17012) by [@apoc](https://github.com/apoc))
+- Antigravity model discovery now keeps each account's per-tier `model_enum`, output cap and thinking level or budget, and the bundled Antigravity models carry them for offline use; cached Antigravity model lists are refreshed once so they pick them up ([#17012](https://github.com/can1357/oh-my-pi/pull/17012) by [@apoc](https://github.com/apoc))
+
+### Fixed
+
+- Fixed OpenCode Go `gemini-4-argon` requests failing with an unsupported protocol; discovered models now use native Gemini routing and low, medium, or high thinking ([#17031](https://github.com/can1357/oh-my-pi/issues/17031))
+- Fixed DeepSeek V4 Pro cost estimates using Flash rates after September 14 despite Pro billing remaining unchanged ([#16875](https://github.com/can1357/oh-my-pi/issues/16875)).
+- Anthropic refusal fallbacks from Claude Opus 5.5, Opus 5 and Sonnet 5.5 to their documented target models (Opus 4.8 and Opus 5, Opus 4.8, and Sonnet 5) now redeem fallback credit: the retry is no longer billed as a full cache write, and it is no longer skipped when the conversation holds signed thinking ([#17034](https://github.com/can1357/oh-my-pi/pull/17034) by [@sjawhar](https://github.com/sjawhar))
+- Fixed Step 5 Preview reasoning choices on OpenCode Go and Zen to offer only low, medium, and high ([#14987](https://github.com/can1357/oh-my-pi/issues/14987), [#16669](https://github.com/can1357/oh-my-pi/pull/16669) by [@SheetHappiness](https://github.com/SheetHappiness))
+- Fixed GPT-6 Sol/Luna API and MiniMax-M3 long-context usage estimates above their published input thresholds ([#16711](https://github.com/can1357/oh-my-pi/issues/16711))
+- Fixed Xiaomi Token Plan login and default model selection after the V2.5 retirement by using MiMo V2.6 Flash ([#16727](https://github.com/can1357/oh-my-pi/pull/16727) by [@LiPu-jpg](https://github.com/LiPu-jpg))
+- Fixed Xiaomi model discovery sending `ttp-` Token Plan keys to the pay-as-you-go host, which rejects them ([#17045](https://github.com/can1357/oh-my-pi/issues/17045))
+- Fixed GPT-5.6 Sol requests over 272K input tokens (`openai` and `openai-codex`) being priced at the retired $10/$45 long-context card instead of $8/$30 ([#16742](https://github.com/can1357/oh-my-pi/issues/16742))
+- Fixed DeepSeek's official vision requests being held to the 5-image unknown-host floor: the catalog endpoint now carries DeepSeek's documented 600-image and 40 MiB inline-image budgets, while a custom `baseUrl` keeps the conservative floor ([vision limits](https://api-docs.deepseek.com/guides/vision#limits); [#16737](https://github.com/can1357/oh-my-pi/pull/16737) by [@Fun10165](https://github.com/Fun10165))
+- Fixed Devin model discovery falling back to the two SWE-1.6 seed models when the catalog takes longer than 5s to respond; it now waits up to 15s ([#16668](https://github.com/can1357/oh-my-pi/pull/16668) by [@DusKing1](https://github.com/DusKing1))
+- Fixed Go Task keeping stale prompt-cache affinity after a model's wire model changed ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+- Fixed Devin models all showing a 200K context window: each model now reports its own (1M for Claude Opus 5.5 and the `-1m` lanes, 262K for SWE-2), and GPT and Grok models with a long-context price tier start at that tier's threshold, with `/extended-context on` opening their full window ([#16905](https://github.com/can1357/oh-my-pi/pull/16905) by [@will-bogusz](https://github.com/will-bogusz))
+- Added the `extended-cache-ttl-beta` compat axis so a route can opt out of the `extended-cache-ttl-2025-04-11` beta (Factory Droid does), and prompt-cache lifetimes and long-retention support for Factory Droid's Claude models ([#16296](https://github.com/can1357/oh-my-pi/pull/16296) by [@will-bogusz](https://github.com/will-bogusz))
+- Added Factory Droid's image models (GPT Image 2.5, GPT Image 2.5 Flare, Gemini 3 Pro Image) to the `image` role's choices, listed only while your Factory organization allows image generation and that model ([#16316](https://github.com/can1357/oh-my-pi/pull/16316) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed LiteLLM model groups with `mode: decisions` being listed as chat models; they now route to `openai-decisions` as judge models ([#16939](https://github.com/can1357/oh-my-pi/issues/16939))
+- Removed retired Anthropic models (Claude 3.x, Opus 4/4.1, Sonnet 4) from the bundled catalog and discovery ([#16755](https://github.com/can1357/oh-my-pi/issues/16755))
+- Fixed Kimi Code K3 and K3 256K (`k3`, `k3-256k`) pricing to include Kimi's $3/MTok cache-write rate, matching the Moonshot API `kimi-k3` card ([#16508](https://github.com/can1357/oh-my-pi/issues/16508), [#16903](https://github.com/can1357/oh-my-pi/pull/16903) by [@eggpeat](https://github.com/eggpeat))
+- Fixed Devin models all showing a 200K context window: each model now reports its own (1M for Claude Opus 5.5 and the `-1m` lanes, 262K for SWE-2), and GPT and Grok models with a long-context price tier start at that tier's threshold, with `/extended-context on` opening their full window ([#16905](https://github.com/can1357/oh-my-pi/pull/16905) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed Cursor Auto (`cursor/default`) being metered against the third-party ("Other Models") pool instead of the Cursor Models pool it bills to ([#17048](https://github.com/can1357/oh-my-pi/pull/17048) by [@trukhinyuri](https://github.com/trukhinyuri))
+
+## [18.12.6] - 2026-10-21
+
+### Added
+
+- Added the `prompt-cache-persistent` catalog axis (`isPromptCachePersistent`) for providers whose own endpoint never expires prompt caches; logged-in OpenAI Codex declares it
+
+### Changed
+
+- Updated the bundled Yolo-Auto catalog to its current models, `qwen3.8-27b` (now the default), `yolo`, `yolo-small`, and `yolo-128k`, and removed the retired `deepseek-flash-v4` and `qwen3.8-flash`
+
 ## [18.12.5] - 2026-10-21
 
 ### Fixed

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

@@ -6,6 +6,26 @@
  * prefix check would accept lookalikes like `https://api.anthropic.com.evil.com`.
  */
 export declare function isOfficialAnthropicApiUrl(baseUrl?: string): boolean;
+/** `anthropic-version` values this codebase knows how to speak. */
+export declare const ANTHROPIC_API_VERSIONS: readonly ["2023-06-01", "2026-10-21"];
+export type AnthropicApiVersion = (typeof ANTHROPIC_API_VERSIONS)[number];
+/** Version every non-official Anthropic-compatible endpoint receives. */
+export declare const LEGACY_ANTHROPIC_API_VERSION: AnthropicApiVersion;
+/**
+ * `anthropic-version` for a request to `baseUrl`. `PI_ANTHROPIC_API_VERSION`
+ * opts the official first-party API into a newer version; every other host
+ * (Copilot, Bedrock, Foundry, Cloudflare, third-party compatibles) stays on
+ * {@link LEGACY_ANTHROPIC_API_VERSION} because those deployments gate newer
+ * versions on their own schedule (Bedrock accepts 2026-10-21 only for 4.7+).
+ * Read per call so a flipped environment takes effect without a restart.
+ */
+export declare function resolveAnthropicApiVersion(baseUrl?: string): AnthropicApiVersion;
+/**
+ * Whether Messages requests on `version` may carry `temperature`, `top_p`,
+ * or `top_k`. API version 2026-10-21 rejects all three with a 400, even at
+ * their default values.
+ */
+export declare function anthropicApiVersionAcceptsSamplingParams(version: AnthropicApiVersion): boolean;
 /**
  * Amazon Bedrock's Anthropic Messages API: the `/anthropic` path on the
  * bedrock-runtime or bedrock-mantle endpoint. It rejects the tool `strict`

package/dist/types/compat/behavior.d.ts

@@ -71,6 +71,8 @@ export declare function isRetiredProvider(provider: string): boolean;
  * model that happens to advertise the same bare id.
  */
 export declare function isBareIdReferenceProvider(provider: string): boolean;
+/** Whether a host's prefix cache should use append-only conversation context by default. */
+export declare function prefersAppendOnlyContext(provider: string): boolean;
 /** The declared subscription tier required to use a provider model id, if any. */
 export declare function planRequirementFor(provider: string, model: string): string | undefined;
 /** Cross-provider pricing-peer resolution for one provider model id. */

package/dist/types/compat/catalog-policy.d.ts

@@ -1,7 +1,7 @@
-import type { Model } from "../types.js";
+import type { CatalogPolicyInputs } from "../types.js";
 /**
  * The `catalog` axis bag authored in the KDL cascade rules for a model,
  * memoized per distinct model. Frozen: the bag is shared across callers, so it
  * must stay read-only — extract the value you need rather than handing it on.
  */
-export declare function resolveCatalogPolicy(model: Model): Readonly<Record<string, unknown>>;
+export declare function resolveCatalogPolicy(model: CatalogPolicyInputs): Readonly<Record<string, unknown>>;

package/dist/types/compat/collapse.d.ts

@@ -139,6 +139,16 @@ export declare function isCollapsedVariantSpec(spec: VariantSpecLike): boolean;
  * and for routes live discovery never advertised.
  */
 export declare function isCursorMaxModeWireId(wireModelId: string): boolean;
+/**
+ * Whether `wireModelId` is one tier of a reviewed `provider` family that serves
+ * different efforts from distinct sibling SKUs (`-low`/`-medium`/`-high`), so
+ * the SKU, not the effort, fixes the thinking strength. Decided from the
+ * declared family routes, not a collapsed row's routing: an account roster
+ * listing one tier collapses to a row routing every effort to it, yet that SKU
+ * still advertises its own fixed thinking. Families routing every effort to
+ * one `-thinking` SKU keep the effort's own budget or level.
+ */
+export declare function isReviewedTierVariantWireId(provider: string, wireModelId: string): boolean;
 /**
  * Collapse a full mixed-provider list: per provider, the compiled reviewed table, Cursor's
  * conservative live effort-sibling rule, and the automatic `X`/`X-thinking`

package/dist/types/compat/factory-droid.d.ts

@@ -45,6 +45,8 @@ export interface FactoryDroidModelPolicy {
     entitlement: FactoryDroidEntitlement;
     /** Native default with no caller effort is thinking off. */
     defaultReasoningOff: boolean;
+    /** A `generate_image` model: gated by the org's image-generation settings, served through Factory's image routes. */
+    image: boolean;
 }
 /** One registry model: its roster row and resolved policy. */
 export interface FactoryDroidRegistryModel {

package/dist/types/compat/prompt-cache-persistent.d.ts

@@ -0,0 +1,7 @@
+import type { Model } from "../types.js";
+/**
+ * Whether {@link model}'s prompt-cache entries never expire
+ * (`prompt-cache-persistent` axis). Only the provider's own endpoint
+ * qualifies: a gateway or proxy reusing the provider id makes no such promise.
+ */
+export declare function isPromptCachePersistent(model: Model): boolean;

package/dist/types/compat/resolve.d.ts

@@ -1,4 +1,4 @@
-import type { Api, CompatOf, ModelSpec, ThinkingConfig } from "../types.js";
+import type { Api, CatalogPolicyInputs, CompatOf, ModelSpec, ThinkingConfig } from "../types.js";
 import type { ModelIdentity, RequestPolicy } from "./types.js";
 /** Result of resolving one model spec through the compat engine. */
 export interface ResolvedModelPolicy<TApi extends Api = Api> {
@@ -21,7 +21,7 @@ export declare function resolveDiscoveryApi(spec: ModelSpec<Api>, providerType:
  * {@link resolveModelPolicy} without resolving the compat and thinking
  * policies, which dominate its cost on catalog-wide scans.
  */
-export declare function resolveCatalogAxes(spec: ModelSpec<Api>): Record<string, unknown>;
+export declare function resolveCatalogAxes(spec: CatalogPolicyInputs): Record<string, unknown>;
 /**
  * Resolves the full policy surface for one model spec: structured identity,
  * complete compat record, thinking metadata, and catalog-data corrections.

package/dist/types/compat/types.d.ts

@@ -327,6 +327,8 @@ export interface CompiledBehavior {
     pricingPeers: CompiledPricingPeer[];
     retryResetTimezones: CompiledRetryResetTimezone[];
     retiredProviders: string[];
+    /** Hosts whose prefix caches benefit from a stable append-only conversation. */
+    appendOnlyContextProviders: string[];
     referenceIsolatedProviders: string[];
     /** Built-in model selector chains per agent role, keyed by role in declaration order. */
     roleDefaults: Record<string, string[]>;

package/dist/types/discovery/antigravity.d.ts

@@ -11,6 +11,8 @@ export interface AntigravityDiscoveryApiModel {
     supportsImages?: boolean;
     supportsThinking?: boolean;
     thinkingBudget?: number;
+    /** SKU thinking level: 1 low, 2 medium, 3 high. */
+    thinkingLevel?: number;
     recommended?: boolean;
     maxTokens?: number;
     maxOutputTokens?: number;
@@ -55,8 +57,11 @@ export interface FetchAntigravityDiscoveryModelsOptions {
     token: string;
     /** Optional endpoint override. Defaults to Antigravity fallback endpoints. */
     endpoint?: string;
-    /** Deprecated and ignored for antigravity discovery parity. */
-    project?: string;
+    /**
+     * Cloud Code Assist project the account is onboarded to. Sent as the body's
+     * `project` field like the native client; the body is `{}` when absent.
+     */
+    projectId?: string;
     /** Optional user agent override. */
     userAgent?: string;
     /** Optional abort signal for request cancellation. */

package/dist/types/discovery/devin.d.ts

@@ -7,7 +7,7 @@ export interface DevinModelDiscoveryOptions {
     apiKey?: string;
     /** Optional Codeium API base URL override. */
     baseUrl?: string;
-    /** Optional request timeout in milliseconds (default 5000). */
+    /** Optional request timeout in milliseconds (default 15000). */
     timeoutMs?: number;
     /** Optional caller abort signal, combined with the internal timeout. */
     signal?: AbortSignal;

package/dist/types/provider-models/google.d.ts

@@ -21,6 +21,8 @@ export interface GoogleAntigravityAccount {
      * When any account lacks one, no model records per-account access.
      */
     accountKey?: string;
+    /** Cloud Code Assist project of the credential, sent with roster discovery like the native client. */
+    projectId?: string;
 }
 export interface GoogleAntigravityModelManagerConfig {
     /**

package/dist/types/provider-models/openai-compat.d.ts

@@ -475,8 +475,8 @@ export interface YoloAutoModelManagerConfig {
     fetch?: FetchImpl;
 }
 /**
- * Yolo-Auto model manager: OpenAI-compatible chat completions at the
- * flat-rate `deepseek-flash-v4` endpoint. Live `/v1/models` discovery replaces the
+ * Yolo-Auto model manager: flat-rate OpenAI-compatible chat completions
+ * (`qwen3.8-27b`, `yolo`, ...). Live `/v1/models` discovery replaces the
  * bundled seed once a key is stored or present in `YOLO_AUTO_API_KEY`. Models
  * the provider adds later inherit metadata (reasoning, thinking, context)
  * from the global bundled reference index, so new ids work without a per-id
@@ -560,6 +560,11 @@ export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_MAX_TOKENS = 32768;
 /** Exclude only known non-conversational modes; unknown and non-string modes remain selectable for aliases. */
 export declare function isSelectableLiteLLMModelMode(mode: unknown): boolean;
 export declare function normalizeLiteLLMManagementBaseUrl(baseUrl: string): string;
+/**
+ * Transport for a LiteLLM model group: `mode: "decisions"` answers only on the
+ * proxy's `/v1/decisions` route, OpenAI-backed groups use Responses, and every
+ * other group uses `fallbackApi`.
+ */
 export declare function resolveLiteLLMApi(entry: Record<string, unknown> | undefined, id: string, fallbackApi?: Api): Api;
 export declare function fetchLiteLLMRichModels<TApi extends Api>(options: FetchLiteLLMRichModelsOptions<TApi>): Promise<ModelSpec<TApi>[] | null>;
 export declare function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig): ModelManagerOptions<Api>;

package/dist/types/types.d.ts

@@ -20,12 +20,12 @@ export declare const RUNNER_API_KINDS: Record<Exclude<(typeof RUNNER_APIS)[numbe
 /** Kind a runner API serves; `undefined` for chat transports and multi-kind `local-inference`. */
 export declare function runnerApiKind(api: Api): ModelKind | undefined;
 /** Catalog APIs `generate_image` runs through a pi-ai image client; the hosted Responses pair needs a carrier model. */
-export declare const IMAGE_GENERATION_APIS: readonly ["openai-images", "openrouter-images", "google-generative-ai", "google-gemini-cli", "openai-responses", "openai-codex-responses"];
+export declare const IMAGE_GENERATION_APIS: readonly ["openai-images", "openrouter-images", "google-generative-ai", "google-gemini-cli", "openai-responses", "openai-codex-responses", "factory-droid-agent"];
 export type ImageGenerationApi = (typeof IMAGE_GENERATION_APIS)[number];
 /**
  * Kinds a model on `api` may declare: a runner api serves its own kind; a chat
  * transport serves `chat` and `tiny`, plus `image` when `generate_image` runs it
- * (hosted Responses image tool, Gemini image models). `undefined` for
+ * (hosted Responses image tool, Gemini and Factory Droid image models). `undefined` for
  * `local-inference`, which hosts several kinds chosen by the model itself.
  */
 export declare function servedKinds(api: Api): readonly ModelKind[] | undefined;
@@ -550,6 +550,8 @@ export interface AnthropicCompat {
     supportsEagerToolInputStreaming?: boolean;
     /** Whether long prompt-cache retention (`ttl: "1h"`) is supported. Default: true for canonical Anthropic API. */
     supportsLongCacheRetention?: boolean;
+    /** Whether a `ttl: "1h"` breakpoint on API-key requests adds the `extended-cache-ttl-2025-04-11` beta. Unset: true (Vertex rawPredict URLs never send it). */
+    extendedCacheTtlBeta?: boolean;
     /**
      * Whether mid-conversation `role: "system"` messages are accepted in the
      * `messages` array. When unset, auto-detected from model and deployment policy.
@@ -608,6 +610,21 @@ export interface AnthropicCompat {
      * Default: auto-detected from provider/baseUrl and `model.reasoning`.
      */
     replayUnsignedThinking?: boolean;
+    /**
+     * Whether a thinking-mode history must carry a `thinking` block on EVERY
+     * assistant turn of the current turn. DeepSeek's Anthropic endpoint rejects
+     * the whole request (`The content[].thinking in the thinking mode must be
+     * passed back to the API.`) when any assistant turn after the last user text
+     * lacks one, including turns another upstream in the same gateway model group
+     * produced without reasoning. Presence is all it validates (the block text may
+     * be a placeholder, the signature may be empty), and only `thinking` counts —
+     * `redacted_thinking` does not, and that endpoint rejects the block outright.
+     * A model group that may fail over to such an endpoint must set this: the
+     * stricter member defines the replay contract for the group.
+     * Default: false (official Anthropic and lax proxies tolerate a turn without
+     * thinking and the block is not invented for them).
+     */
+    requiresThinkingPresence?: boolean;
     /**
      * Whether the endpoint requires `thinking.type: "enabled"` whenever the
      * model reasons. Use for models that reject omitted or disabled thinking.
@@ -975,9 +992,11 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
  */
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard" | "bedrockMessagesApi" | "effortBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard" | "bedrockMessagesApi" | "effortBeta" | "extendedCacheTtlBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
     /** Effort-beta override; undefined keeps the transport's legacy heuristic. */
     effortBeta?: AnthropicCompat["effortBeta"];
+    /** Extended-cache-ttl beta opt-out; undefined sends it where a 1h breakpoint needs it. */
+    extendedCacheTtlBeta?: AnthropicCompat["extendedCacheTtlBeta"];
     /** Disabled-thinking wire form; undefined keeps the direct-provider behavior. */
     disabledThinking?: AnthropicCompat["disabledThinking"];
     /** Effort pinned on `between_tools` turns; undefined leaves effort untouched. */
@@ -1204,6 +1223,13 @@ export interface ModelAccountAccess {
      * Absent when the backend reported no program metadata.
      */
     cyberPrograms?: readonly string[];
+    /**
+     * Antigravity request constants this account's `fetchAvailableModels`
+     * roster advertised, keyed by upstream SKU id. Takes precedence over
+     * {@link Model.antigravityWireProfiles}, which keeps only the first
+     * account's values when multi-account rosters are unioned.
+     */
+    antigravityWireProfiles?: Readonly<Record<string, AntigravityModelWireProfile>>;
 }
 /** One Cursor `RequestedModel.parameters` entry recovered from rich discovery. */
 export interface CursorModelParameter {
@@ -1216,6 +1242,20 @@ export interface CursorModelRoute {
     parameters: readonly CursorModelParameter[];
     maxMode?: boolean;
 }
+/**
+ * Antigravity Cloud Code Assist request constants for one upstream SKU, as the
+ * native client sends them; see {@link Model.antigravityWireProfiles}.
+ */
+export interface AntigravityModelWireProfile {
+    /** Opaque `labels.model_enum` token (`fetchAvailableModels` `model`). */
+    modelEnum?: string;
+    /** Fixed `generationConfig.maxOutputTokens` the client always sends. */
+    maxOutputTokens?: number;
+    /** SKU thinking level; the native client sends it in preference to a budget. */
+    thinkingLevel?: "LOW" | "MEDIUM" | "HIGH";
+    /** SKU thinking budget (`-1` is dynamic), sent when no level is advertised. */
+    thinkingBudget?: number;
+}
 export interface Model<TApi extends Api = Api> {
     id: string;
     /** Role-specific runner capability; omitted for ordinary chat models. */
@@ -1329,6 +1369,13 @@ export interface Model<TApi extends Api = Api> {
      * their single wire id) and on bundled snapshots that predate discovery.
      */
     cursorMaxModeRoutes?: Readonly<Record<string, boolean>>;
+    /**
+     * Per-wire-id Antigravity request constants recorded from the account's
+     * `fetchAvailableModels` roster, keyed by upstream SKU id. A collapsed row
+     * carries one entry per member it routes to; a raw row carries its own.
+     * Absent on bundled snapshots that predate discovery.
+     */
+    antigravityWireProfiles?: Readonly<Record<string, AntigravityModelWireProfile>>;
     /**
      * Per-account availability recorded by multi-account discovery: provider
      * account key (Codex: ChatGPT `chatgpt_account_id`; Antigravity: login
@@ -1525,3 +1572,11 @@ export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "co
     /** Sparse compatibility overrides; resolved into `Model.compat` by `buildModel`. */
     compat?: CompatConfigOf<TApi>;
 }
+/**
+ * The model facts the catalog policy cascade reads: the resolve-target inputs
+ * (`providerType ?? provider`, the request adapter, the wire model id, the
+ * reasoning flag) plus the unverified-thinking correction applied before the
+ * cascade. Callers holding a reduced model view — request adapters that only
+ * need policy — supply exactly these instead of asserting a full `Model`.
+ */
+export type CatalogPolicyInputs = Pick<ModelSpec<Api>, "provider" | "providerType" | "api" | "id" | "reasoning" | "requestModelId" | "thinking">;

package/dist/types/wire/devin.d.ts

@@ -13,9 +13,9 @@ export declare const DEVIN_DEFAULT_BASE_URL = "https://server.codeium.com";
 export declare function devinCliMetadata(apiKey: string | undefined, userJwt?: string): {
     ideName: "devin-cli";
     ideType: "chisel";
-    ideVersion: "3000.11.3";
+    ideVersion: "3000.11.205";
     extensionName: "chisel";
-    extensionVersion: "3000.11.3";
+    extensionVersion: "3000.11.205";
     locale: "en";
     os: "darwin" | "linux" | "windows";
     apiKey: string;

package/dist/types/wire/gemini-headers.d.ts

@@ -1,4 +1,5 @@
 import type { FetchImpl } from "@oh-my-pi/pi-utils";
+import type { AntigravityModelWireProfile } from "../types.js";
 /**
  * Build a User-Agent string that identifies as Gemini CLI to unlock higher rate limits.
  * Uses the same format as the official Gemini CLI (v0.35+):
@@ -15,16 +16,19 @@ export declare const getGeminiCliHeaders: (modelId?: string) => {
  * (and its @google/genai → google-auth-library dependency chain) into the startup
  * parse graph.
  *
- * Format captured from the real 2.8.0 `antigravity/hub` client:
- * `antigravity/hub/2.8.0 (aidev_client; os_type=darwin; arch=arm64; cl=963137146)`.
+ * Format captured from the real 2.26.0 `antigravity/hub` client:
+ * `antigravity/hub/2.26.0 (aidev_client; os_type=darwin; arch=arm64; cl=1005016104)`.
  * The backend gates newer models (e.g. gemini-3.7-flash) on the client version,
  * so the version tracks the latest Antigravity release via the update manifest
  * (see {@link ensureAntigravityVersion}) with `DEFAULT_ANTIGRAVITY_VERSION` as
- * the offline fallback. os_type/arch are pinned to the darwin/arm64 reference
- * client the version and manifest are captured from, independent of the host
- * platform. Overrides: PI_AI_ANTIGRAVITY_VERSION / _CL / _OS / _ARCH.
+ * the offline fallback. os_type/arch name the native build for the host OS:
+ * each 2.26.0 installer ships one language-server build whose Go target is
+ * stamped into the header (macOS darwin/arm64, Linux linux/amd64, Windows
+ * windows/amd64; no other OS/arch builds exist); see
+ * {@link getAntigravityPlatform}. The changelist is identical across those
+ * builds. Overrides: PI_AI_ANTIGRAVITY_VERSION / _CL / _OS / _ARCH.
  */
-export declare const DEFAULT_ANTIGRAVITY_VERSION = "2.19.1";
+export declare const DEFAULT_ANTIGRAVITY_VERSION = "2.26.0";
 /** Current Antigravity client version: env override → manifest-discovered → pinned fallback. */
 export declare function getAntigravityVersion(): string;
 /**
@@ -42,24 +46,39 @@ export declare function parseAntigravityManifestVersion(yamlText: string): strin
  * timeout; `signal` ends this caller's wait without cancelling the lookup.
  */
 export declare function ensureAntigravityVersion(fetcher?: FetchImpl, signal?: AbortSignal): Promise<void>;
+/**
+ * Native build the host would run. OSes without a native build report the
+ * Linux build, the closest client that exists.
+ */
+export declare function getAntigravityPlatform(platform?: NodeJS.Platform): {
+    os: string;
+    arch: string;
+};
 /** Antigravity `User-Agent` header value; rebuilt when the discovered version changes. */
 export declare function getAntigravityUserAgent(): string;
 /**
- * Per-wire-id Antigravity Cloud Code Assist request constants, captured from the
- * real `antigravity/hub` client against `daily-cloudcode-pa`. `modelEnum` is the
- * opaque `labels.model_enum` token the client tags each request with — optional
- * because Anthropic-backed wire ids (e.g. `claude-sonnet-4-6`,
- * `claude-opus-4-6-thinking`) are accepted without one; the label is purely
- * telemetry. `maxOutputTokens` is the fixed `generationConfig.maxOutputTokens`
- * the backend enforces regardless of the thinking budget (Claude caps at
- * 64000, Gemini accepts the discovered cap). Keyed by the routed upstream wire
- * id (post effort-routing), not the collapsed logical id. Checkpoint-only ids
- * (e.g. `gemini-3.1-flash-lite`) are intentionally absent — this provider only
- * emits agent requests.
+ * Client headers for every Antigravity Cloud Code Assist request. The native
+ * client is Go `net/http`, which offers `Accept-Encoding: gzip` and no `Accept`;
+ * pinning the encoding replaces the runtime's `gzip, deflate, br, zstd` (Bun
+ * still decompresses the gzip body). Bun always adds `Accept: *\/*` itself, so
+ * callers must not set a more specific `Accept`.
+ */
+export declare function getAntigravityHeaders(userAgent?: string): {
+    "User-Agent": string;
+    "Accept-Encoding": string;
+};
+/**
+ * Offline fallback for {@link Model.antigravityWireProfiles}: per-SKU request
+ * constants for rows that carry no discovered profile. Keyed by the routed
+ * upstream wire id (post effort-routing), not the collapsed logical id.
+ * `modelEnum` is the opaque `labels.model_enum` token; the backend accepts
+ * requests without it. Discovery (and the bundled snapshot baked from it)
+ * owns current SKUs; do not extend this table.
  */
-export interface AntigravityModelWireProfile {
-    modelEnum?: string;
-    maxOutputTokens: number;
-}
 export declare const ANTIGRAVITY_MODEL_WIRE_PROFILES: Readonly<Record<string, AntigravityModelWireProfile>>;
-export declare function getAntigravityModelWireProfile(wireModelId: string): AntigravityModelWireProfile | undefined;
+/**
+ * Request constants for `wireModelId`: the first discovered profile map that
+ * lists it (callers pass the serving account's before the row's), else the
+ * offline fallback table.
+ */
+export declare function getAntigravityModelWireProfile(wireModelId: string, ...discovered: (Readonly<Record<string, AntigravityModelWireProfile>> | undefined)[]): AntigravityModelWireProfile | undefined;

package/dist/types/wire/xiaomi-token-plan.d.ts

@@ -0,0 +1,2 @@
+/** Identifies Xiaomi Token Plan keys so login and discovery use regional hosts instead of the pay-as-you-go host. */
+export declare function isXiaomiTokenPlanKey(apiKey: string): boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.12.5",
+	"version": "18.13.0",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -93,12 +93,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.12.5",
-		"@oh-my-pi/pi-utils": "18.12.5"
+		"@oh-my-pi/omptype": "18.13.0",
+		"@oh-my-pi/pi-utils": "18.13.0"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.12.5",
+		"@oh-my-pi/pi-ai": "18.13.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/build.ts

@@ -19,7 +19,16 @@ import {
 	NO_RECURRING_TARIFF,
 	parseDatedCostCards,
 } from "./pricing";
-import { type Api, type EffectiveTokenCost, MODEL_KINDS, type Model, type ModelCost, type ModelSpec } from "./types";
+import {
+	type AntigravityModelWireProfile,
+	type Api,
+	type EffectiveTokenCost,
+	MODEL_KINDS,
+	type Model,
+	type ModelAccountAccess,
+	type ModelCost,
+	type ModelSpec,
+} from "./types";
 import { cleanModelName } from "./utils";
 
 function numberField(source: object, key: string): number | undefined {
@@ -353,6 +362,42 @@ export function applyCatalogCorrections(
 	}
 }
 
+/**
+ * Lower discovered Antigravity per-SKU output caps to the host's
+ * `max-tokens-ceiling`. The transport sends a profile's `maxOutputTokens`
+ * verbatim, so a roster value above the ceiling (Claude 4.x advertises 65,536;
+ * Cloud Code Assist 400s past 64,000, #3067) must be corrected here as the
+ * row's `maxTokens` is.
+ */
+function capAntigravityWireProfiles(
+	model: Pick<Model<Api>, "antigravityWireProfiles" | "accountAccess">,
+	ceiling: unknown,
+): void {
+	if (typeof ceiling !== "number") return;
+	const cap = (
+		profiles: Readonly<Record<string, AntigravityModelWireProfile>> | undefined,
+	): Readonly<Record<string, AntigravityModelWireProfile>> | undefined => {
+		if (profiles === undefined) return undefined;
+		let capped: Record<string, AntigravityModelWireProfile> | undefined;
+		for (const [wireId, profile] of Object.entries(profiles)) {
+			if (profile.maxOutputTokens === undefined || profile.maxOutputTokens <= ceiling) continue;
+			capped ??= { ...profiles };
+			capped[wireId] = { ...profile, maxOutputTokens: ceiling };
+		}
+		return capped ?? profiles;
+	};
+	model.antigravityWireProfiles = cap(model.antigravityWireProfiles);
+	if (model.accountAccess === undefined) return;
+	let accountAccess: Record<string, ModelAccountAccess> | undefined;
+	for (const [accountKey, access] of Object.entries(model.accountAccess)) {
+		const profiles = cap(access.antigravityWireProfiles);
+		if (profiles === access.antigravityWireProfiles) continue;
+		accountAccess ??= { ...model.accountAccess };
+		accountAccess[accountKey] = { ...access, antigravityWireProfiles: profiles };
+	}
+	if (accountAccess !== undefined) model.accountAccess = accountAccess;
+}
+
 /**
  * Direct first-party OpenAI Responses endpoints (api.openai.com, Azure OpenAI
  * deployments). Gates GA computer-use detection; identity supplies the model
@@ -445,6 +490,7 @@ export function buildModel<TApi extends Api>(spec: ModelSpec<TApi>): Model<TApi>
 	};
 	applyCatalogAssignments(model, policy.catalog);
 	applyCatalogCorrections(model, policy.catalog);
+	capAntigravityWireProfiles(model, policy.catalog.maxTokensCeiling);
 	// Configured kinds and lifetimes replace catalog values rather than merging with them.
 	if (spec.kindConfig !== undefined) model.kind = spec.kindConfig;
 	if (spec.promptCacheConfig !== undefined) model.promptCache = { ...spec.promptCacheConfig };

package/src/compat/anthropic.ts

@@ -5,6 +5,7 @@
  * pi-ai both consume these; URL matching is the one detection surface that
  * stays in code.
  */
+import { $env, logger } from "@oh-my-pi/pi-utils";
 import { hostMatchesUrl, isBedrockRouteUrl } from "../hosts";
 
 const OFFICIAL_ANTHROPIC_URL = "https://api.anthropic.com";
@@ -22,6 +23,52 @@ export function isOfficialAnthropicApiUrl(baseUrl?: string): boolean {
 	return lower === OFFICIAL_ANTHROPIC_URL || lower.startsWith(`${OFFICIAL_ANTHROPIC_URL}/`);
 }
 
+/** `anthropic-version` values this codebase knows how to speak. */
+export const ANTHROPIC_API_VERSIONS = ["2023-06-01", "2026-10-21"] as const;
+export type AnthropicApiVersion = (typeof ANTHROPIC_API_VERSIONS)[number];
+
+/** Version every non-official Anthropic-compatible endpoint receives. */
+export const LEGACY_ANTHROPIC_API_VERSION: AnthropicApiVersion = "2023-06-01";
+
+let warnedInvalidApiVersion: string | undefined;
+
+function isAnthropicApiVersion(value: string): value is AnthropicApiVersion {
+	return (ANTHROPIC_API_VERSIONS as readonly string[]).includes(value);
+}
+
+/**
+ * `anthropic-version` for a request to `baseUrl`. `PI_ANTHROPIC_API_VERSION`
+ * opts the official first-party API into a newer version; every other host
+ * (Copilot, Bedrock, Foundry, Cloudflare, third-party compatibles) stays on
+ * {@link LEGACY_ANTHROPIC_API_VERSION} because those deployments gate newer
+ * versions on their own schedule (Bedrock accepts 2026-10-21 only for 4.7+).
+ * Read per call so a flipped environment takes effect without a restart.
+ */
+export function resolveAnthropicApiVersion(baseUrl?: string): AnthropicApiVersion {
+	if (!isOfficialAnthropicApiUrl(baseUrl)) return LEGACY_ANTHROPIC_API_VERSION;
+	const configured = $env.PI_ANTHROPIC_API_VERSION?.trim();
+	if (!configured) return LEGACY_ANTHROPIC_API_VERSION;
+	if (isAnthropicApiVersion(configured)) return configured;
+	if (warnedInvalidApiVersion !== configured) {
+		warnedInvalidApiVersion = configured;
+		logger.warn("anthropic: unsupported PI_ANTHROPIC_API_VERSION; using legacy version", {
+			configured,
+			supported: ANTHROPIC_API_VERSIONS,
+			using: LEGACY_ANTHROPIC_API_VERSION,
+		});
+	}
+	return LEGACY_ANTHROPIC_API_VERSION;
+}
+
+/**
+ * Whether Messages requests on `version` may carry `temperature`, `top_p`,
+ * or `top_k`. API version 2026-10-21 rejects all three with a 400, even at
+ * their default values.
+ */
+export function anthropicApiVersionAcceptsSamplingParams(version: AnthropicApiVersion): boolean {
+	return version === "2023-06-01";
+}
+
 const CLOUDFLARE_ANTHROPIC_GATEWAY_URL_MARKER = /gateway\.ai\.cloudflare\.com\/.+\/anthropic(?:\/|$)/i;
 const VERTEX_ANTHROPIC_URL_MARKER = /aiplatform\.googleapis\.com\/.+\/publishers\/anthropic\//i;
 const BEDROCK_ANTHROPIC_URL_MARKER = /(?:^|\/\/|\.)bedrock-runtime\.[a-z0-9-]+\.amazonaws\.com/i;