@oh-my-pi/pi-catalog 18.12.3 → 18.12.4

package/CHANGELOG.md

@@ -2,6 +2,23 @@
 
 ## [Unreleased]
 
+## [18.12.4] - 2026-10-20
+
+### Added
+
+- Added native Inception support with Mercury 3, Mercury 2.5, and Mercury 2, live chat-model discovery, API-key login, reasoning controls, and Mercury 3 launch pricing while its catalog reports none; Mercury 3 waits up to 10 minutes for its first output, since high-effort reasoning streams nothing ([#16439](https://github.com/can1357/oh-my-pi/pull/16439) by [@eggpeat](https://github.com/eggpeat))
+
+### Changed
+
+- Changed Devin model discovery to send legacy Windsurf Enterprise keys unprefixed in one request, dropping the Windsurf-editor retry that returned a smaller model list ([#13677](https://github.com/can1357/oh-my-pi/pull/13677) by [@will-bogusz](https://github.com/will-bogusz))
+
+### Fixed
+
+- Factory Droid now carries Droid's 3.5 MB inline-image budget (`inline-image-byte-budget`) for its hosted routes, under their 4.5 MB request limit, and the new `request-body-byte-limit` axis (`resolveRequestBodyByteLimit`) records that limit; it applies on Factory's own hosts through pi-ai's `reachesBudgetedHost` ([#16288](https://github.com/can1357/oh-my-pi/pull/16288) by [@will-bogusz](https://github.com/will-bogusz))
+- DeepSeek V4.1 Flash images are now priced by DeepSeek's own billing (994 tokens for a snapcompact frame) instead of a GPT or Claude rate (2,882 to 3,136), so a 25-frame archive no longer counts about 50k tokens the provider never bills and triggers compaction early ([#16418](https://github.com/can1357/oh-my-pi/pull/16418) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed Amazon Bedrock Kimi K3 (`global.`/`us.moonshotai.kimi-k3`) offering Anthropic-style thinking budgets it ignores; it now offers `low`, `high` and `max` effort, defaulting to `max` ([#13697](https://github.com/can1357/oh-my-pi/pull/13697) by [@alblez](https://github.com/alblez))
+- Marked supported Bedrock Claude models as eligible for native compaction ([#16343](https://github.com/can1357/oh-my-pi/pull/16343) by [@anatoli-tsinovoy](https://github.com/anatoli-tsinovoy))
+
 ## [18.12.3] - 2026-10-19
 
 ### Changed

package/dist/types/compat/auth-ids.d.ts

@@ -1,4 +1,4 @@
 /** Every provider with an `auth/<id>.kdl` policy. */
-export type AuthProviderId = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stencil" | "stepfun" | "synthetic" | "tavily" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
+export type AuthProviderId = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "inception" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stencil" | "stepfun" | "synthetic" | "tavily" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
 /** Providers whose policy declares a `login` flow (the `/login` roster). */
-export type LoginProviderId = "abliteration" | "aiand" | "alibaba-coding-plan" | "alibaba-token-plan" | "anthropic" | "baseten" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google-antigravity" | "google-gemini-cli" | "helmcode" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax-code" | "minimax-code-cn" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stencil" | "stepfun" | "synthetic" | "tavily" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
+export type LoginProviderId = "abliteration" | "aiand" | "alibaba-coding-plan" | "alibaba-token-plan" | "anthropic" | "baseten" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google-antigravity" | "google-gemini-cli" | "helmcode" | "huggingface" | "inception" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax-code" | "minimax-code-cn" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stencil" | "stepfun" | "synthetic" | "tavily" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";

package/dist/types/compat/image-tokenization.d.ts

@@ -22,6 +22,18 @@ export interface OpenAiPatchTokenization {
     /** The level `auto` (and an omitted `detail`) sizes like. */
     auto: "high" | "original";
 }
+/** DeepSeek ViT patches merged `downsample`² into one token, plus a newline per row and two delimiters. */
+export interface DeepSeekPatchTokenization {
+    regime: "deepseek-patch";
+    /** ViT patch edge in pixels. */
+    patchSize: number;
+    /** Patches merged per token along each axis. */
+    downsample: number;
+    /** Token budget per image, layout tokens included. */
+    maxTokens: number;
+    /** Smaller images are upscaled to this pixel area. */
+    minPixels: number;
+}
 /** One model line's image billing rule: the vendor formula and its per-model numbers. */
 export type ImageTokenization = (OpenAiPatchTokenization
 /** Anthropic 28px patches, resized to fit a padded-edge limit and a visual-token budget. */
@@ -29,7 +41,7 @@ export type ImageTokenization = (OpenAiPatchTokenization
     regime: "anthropic-patch";
     maxEdge: number;
     maxTokens: number;
-}
+} | DeepSeekPatchTokenization
 /** A fixed per-image budget regardless of pixels (Gemini 3 `media_resolution`). */
  | {
     regime: "fixed";

package/dist/types/compat/provider-ids.d.ts

@@ -1,2 +1,2 @@
 /** Every chat-model provider with a catalog entry (`default-model` in its provider rules). */
-export type KnownProvider = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "kilo" | "kimi-code" | "litellm" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "opencode-go" | "opencode-zen" | "openrouter" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stepfun" | "synthetic" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zenmux" | "zhipu-coding-plan";
+export type KnownProvider = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "inception" | "kilo" | "kimi-code" | "litellm" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "opencode-go" | "opencode-zen" | "openrouter" | "qianfan" | "qwen-portal" | "sakana" | "sambanova" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stepfun" | "synthetic" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zenmux" | "zhipu-coding-plan";

package/dist/types/compat/request-size.d.ts

@@ -2,15 +2,20 @@ import type { Model } from "../types.js";
 /**
  * Maximum encoded inline-image bytes retained per request for a deployment with a documented body-size limit.
  * The axis is a host contract: a rerouted endpoint (proxy, gateway) imposes its own limit, so the budget applies
- * only when `officialEndpoint` says the request reaches the model's official endpoint. `compat.officialEndpoint`
- * classifies only the configured `baseUrl`; pi-ai's `reachesOfficialEndpoint` also follows request-time reroutes.
+ * only when `budgetedHost` says the request reaches the host the rules describe: pi-ai's `reachesBudgetedHost`
+ * (the model's official endpoint, following request-time reroutes, or Factory's hosted API for a Factory Droid model).
  */
-export declare function resolveInlineImageByteBudget(model: Model, officialEndpoint: boolean): number | undefined;
+export declare function resolveInlineImageByteBudget(model: Model, budgetedHost: boolean): number | undefined;
+/**
+ * Largest request body the deployment accepts (`request-body-byte-limit`), or `undefined` when no rule names it.
+ * Same gate as {@link resolveInlineImageByteBudget}.
+ */
+export declare function resolveRequestBodyByteLimit(model: Model, budgetedHost: boolean): number | undefined;
 /**
  * Most images one request to this deployment may carry, or `undefined` when no rule names the host. A configured
  * `model.imageCountBudget` wins: the user knows the host behind a custom gateway. Otherwise the
- * `official-image-count-budget` axis replaces `image-count-budget` only when `officialEndpoint` says the request
- * reaches the model's official endpoint (see {@link resolveInlineImageByteBudget}); a rerouted endpoint keeps the
+ * `official-image-count-budget` axis replaces `image-count-budget` only when `budgetedHost` says the request
+ * reaches the host the rules describe (see {@link resolveInlineImageByteBudget}); a rerouted endpoint keeps the
  * provider's generic budget.
  */
-export declare function resolveImageCountBudget(model: Model, officialEndpoint: boolean): number | undefined;
+export declare function resolveImageCountBudget(model: Model, budgetedHost: boolean): number | undefined;

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

@@ -3,7 +3,7 @@ import type { FetchImpl, ModelSpec } from "../types.js";
  * Options for fetching dynamic Devin (Codeium Cascade) models from `GetCliModelConfigs`.
  */
 export interface DevinModelDiscoveryOptions {
-    /** Codeium session token carried inside protobuf `Metadata.apiKey`. */
+    /** Stored Devin credential; encoded into protobuf `Metadata.apiKey` by shape. */
     apiKey?: string;
     /** Optional Codeium API base URL override. */
     baseUrl?: string;

package/dist/types/hosts.d.ts

@@ -175,11 +175,13 @@ export declare function isAzureDeploymentsUrl(baseUrl: string): boolean;
  * - `anthropic`: the Anthropic Messages API under `/anthropic`.
  * - `openai`: the OpenAI-compatible APIs under `/openai`, plus Mantle's
  *   documented `/v1` base (`https://bedrock-mantle.<region>.api.aws/v1`).
+ * - `runtime`: the bedrock-runtime origin itself, where Converse and
+ *   InvokeModel live under `/model/<id>/…`.
  *
  * Hostnames are parsed strictly so proxies that embed these hosts in a path
  * do not match.
  */
-export declare function isBedrockRouteUrl(baseUrl: string | undefined, route: "openai" | "anthropic"): boolean;
+export declare function isBedrockRouteUrl(baseUrl: string | undefined, route: "openai" | "anthropic" | "runtime"): boolean;
 /** Amazon Bedrock's OpenAI-compatible routes; see {@link isBedrockRouteUrl}. */
 export declare function isBedrockOpenAIUrl(baseUrl: string | undefined): boolean;
 /** Alibaba DashScope consumer `compatible-mode` endpoint (rejects multimodal arrays for some text-only SKUs). */

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

@@ -419,6 +419,13 @@ export interface TogetherModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function togetherModelManagerOptions(config?: TogetherModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Builds Inception's authenticated model-discovery manager. A successful
+ * `/v1/models` fetch is the complete roster, keeping only rows that advertise
+ * `chat.completions` (the shared roster also lists FIM/edit and Decisions-only
+ * models); without an API key the bundled Mercury seeds are served.
+ */
+export declare function inceptionModelManagerOptions(config?: ModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface CoreWeaveModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -688,8 +688,8 @@ export interface AnthropicCompat {
     fastMode?: boolean;
 }
 /**
- * Compatibility settings for Bedrock Converse prompt caching. Cache pricing is
- * deliberately not used to infer these request-shape capabilities.
+ * Compatibility settings for Bedrock Converse requests and native streaming
+ * side-compaction. Cache pricing does not imply these capabilities.
  */
 export interface BedrockCompat {
     /** Explicit disabled-thinking wire form; unset preserves the provider's existing behavior. */
@@ -737,8 +737,13 @@ export interface BedrockCompat {
      * local prefix-binding safeguards. Default: true.
      */
     supportsThinkingBindingControls?: boolean;
+    /**
+     * Whether the model supports a compaction-only Bedrock Runtime
+     * InvokeModelWithResponseStream request. Ordinary turns still use Converse.
+     */
+    supportsServerCompaction?: boolean;
 }
-/** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
+/** Fully-resolved Bedrock Converse and side-compaction capabilities, materialized once by `buildModel`. */
 export interface ResolvedBedrockCompat {
     /** See {@link BedrockCompat.disabledThinking}. */
     disabledThinking?: BedrockCompat["disabledThinking"];
@@ -753,6 +758,8 @@ export interface ResolvedBedrockCompat {
     supportsForcedToolChoice: boolean;
     /** See {@link BedrockCompat.supportsThinkingBindingControls}. */
     supportsThinkingBindingControls: boolean;
+    /** See {@link BedrockCompat.supportsServerCompaction}; false unless the Bedrock model-card policy opts in. */
+    supportsServerCompaction: boolean;
     /**
      * Stream-watchdog idle-timeout fallback in ms for hosts with no keepalive
      * events; 0 disables the idle watchdog. Undefined defers to

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

@@ -6,23 +6,9 @@
  */
 /** Base host for Codeium/Windsurf's Cascade API (Connect protocol over HTTP/1.1). */
 export declare const DEVIN_DEFAULT_BASE_URL = "https://server.codeium.com";
-/** Session token as the wire format carries it: the scheme prefix is required. */
-export declare function normalizeDevinSessionToken(apiKey: string | undefined): string;
-/** Released-CLI metadata with credential bytes already encoded for the wire. */
-export declare function devinWireMetadata(apiKey: string | undefined, userJwt?: string): {
-    ideName: "devin-cli";
-    ideType: "chisel";
-    ideVersion: "3000.11.3";
-    extensionName: "chisel";
-    extensionVersion: "3000.11.3";
-    locale: "en";
-    os: "darwin" | "linux" | "windows";
-    apiKey: string;
-    userJwt: string;
-};
 /**
  * Fields for `Metadata` on released-CLI calls (`GetUserJwt`, `AssignModel`,
- * `GetChatMessage`, `GetUserStatus`) authenticated by a Devin session token.
+ * `GetChatMessage`, `GetUserStatus`).
  */
 export declare function devinCliMetadata(apiKey: string | undefined, userJwt?: string): {
     ideName: "devin-cli";

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.12.3",
+	"version": "18.12.4",
 	"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": "18.12.3",
-		"@oh-my-pi/pi-utils": "18.12.3"
+		"@oh-my-pi/omptype": "18.12.4",
+		"@oh-my-pi/pi-utils": "18.12.4"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.12.3",
+		"@oh-my-pi/pi-ai": "18.12.4",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/auth-ids.ts

@@ -40,6 +40,7 @@ export type AuthProviderId =
 	| "groq"
 	| "helmcode"
 	| "huggingface"
+	| "inception"
 	| "kagi"
 	| "kilo"
 	| "kimi-code"
@@ -132,6 +133,7 @@ export type LoginProviderId =
 	| "google-gemini-cli"
 	| "helmcode"
 	| "huggingface"
+	| "inception"
 	| "kagi"
 	| "kilo"
 	| "kimi-code"

package/src/compat/axes.ts

@@ -267,7 +267,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"supports-mid-conversation-system": wire("supportsMidConversationSystem", ["anthropic"]),
 	"supports-mid-conversation-tool-changes": wire("supportsMidConversationToolChanges", ["anthropic"]),
 	"supports-per-message-effort": wire("supportsPerMessageEffort", ["anthropic"]),
-	"supports-server-compaction": wire("supportsServerCompaction", ["anthropic"]),
+	"supports-server-compaction": wire("supportsServerCompaction", ["anthropic", "bedrock"]),
 	"supports-between-tools-thinking": wire("supportsBetweenToolsThinking", ["anthropic"]),
 	"supports-thinking-binding-controls": wire("supportsThinkingBindingControls", ["anthropic", "bedrock"]),
 	"supports-turn-scoped-system": wire("supportsTurnScopedSystem", ["anthropic"]),
@@ -477,7 +477,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"image-count-budget": { key: "imageCountBudget", set: "catalog", shape: "scalar" },
 	/**
 	 * Replaces `image-count-budget` while the request reaches the official
-	 * endpoint (pi-ai `reachesOfficialEndpoint`): a limit measured or documented
+	 * endpoint (pi-ai `reachesBudgetedHost`): a limit measured or documented
 	 * for the host's own endpoint, which a custom `baseUrl` or an
 	 * `ANTHROPIC_BASE_URL`/`FOUNDRY_BASE_URL` reroute (proxy, gateway) need not share.
 	 */
@@ -493,6 +493,22 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	/** Longest image edge the host passes to the model: it downscales larger images to fit first. */
 	"image-max-edge": { key: "imageMaxEdge", set: "catalog", shape: "scalar" },
 	"inline-image-byte-budget": { key: "inlineImageByteBudget", set: "catalog", shape: "scalar" },
+	/**
+	 * Largest request body the host accepts. Where set, the inline-image byte clamp keeps the newest user message's
+	 * and newest tool result's images past `inline-image-byte-budget` while they alone fit this limit.
+	 */
+	"request-body-byte-limit": { key: "requestBodyByteLimit", set: "catalog", shape: "scalar" },
+	/**
+	 * Opt-in: the host budgets above apply whenever the request reaches the provider's own catalog endpoint
+	 * (`endpoint-hosts`, or an unset `baseUrl`), for a hosted gateway whose compat names no official endpoint.
+	 * Unassigned: they apply only on the official endpoint (pi-ai `reachesOfficialEndpoint`).
+	 */
+	"catalog-endpoint-budgets": {
+		key: "catalogEndpointBudgets",
+		set: "catalog",
+		shape: "scalar",
+		values: [true, false],
+	},
 	/**
 	 * Supported formats and size bounds for judgment-request images; `mime-types`
 	 * is comma-separated because object payloads have no array shape.

package/src/compat/image-tokenization.ts

@@ -36,11 +36,25 @@ export interface OpenAiPatchTokenization {
 	auto: "high" | "original";
 }
 
+/** DeepSeek ViT patches merged `downsample`² into one token, plus a newline per row and two delimiters. */
+export interface DeepSeekPatchTokenization {
+	regime: "deepseek-patch";
+	/** ViT patch edge in pixels. */
+	patchSize: number;
+	/** Patches merged per token along each axis. */
+	downsample: number;
+	/** Token budget per image, layout tokens included. */
+	maxTokens: number;
+	/** Smaller images are upscaled to this pixel area. */
+	minPixels: number;
+}
+
 /** One model line's image billing rule: the vendor formula and its per-model numbers. */
 export type ImageTokenization = (
 	| OpenAiPatchTokenization
 	/** Anthropic 28px patches, resized to fit a padded-edge limit and a visual-token budget. */
 	| { regime: "anthropic-patch"; maxEdge: number; maxTokens: number }
+	| DeepSeekPatchTokenization
 	/** A fixed per-image budget regardless of pixels (Gemini 3 `media_resolution`). */
 	| { regime: "fixed"; tokens: number }
 ) & {
@@ -54,7 +68,10 @@ const UNCLASSIFIED: Pick<ModelIdentity, "class"> = { class: "unknown" };
 function isImageTokenization(value: unknown): value is ImageTokenization {
 	return (
 		isRecord(value) &&
-		(value.regime === "openai-patch" || value.regime === "anthropic-patch" || value.regime === "fixed")
+		(value.regime === "openai-patch" ||
+			value.regime === "anthropic-patch" ||
+			value.regime === "deepseek-patch" ||
+			value.regime === "fixed")
 	);
 }
 
@@ -169,6 +186,41 @@ function anthropicTokens(size: ImageSize, maxEdge: number, maxTokens: number): n
 	return patches(lo, short(lo));
 }
 
+/**
+ * DeepSeek visual tokens: the image is upscaled to `minPixels`, padded to
+ * whole patches, and when its token grid busts `maxTokens` shrunk to the
+ * largest aspect-preserving grid that fits. Port of `plan_image_grid` in the
+ * reference processor (`max_wh_ratio` is unset in the V4.1 config):
+ * <https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/inference/image_processor.py>
+ */
+function deepseekTokens(size: ImageSize, rule: DeepSeekPatchTokenization): number {
+	const { patchSize, downsample, maxTokens, minPixels } = rule;
+	const cell = patchSize * downsample;
+	const spanTokens = (height: number, width: number) => {
+		const rows = Math.ceil(Math.floor(height / patchSize) / downsample);
+		const cols = Math.ceil(Math.floor(width / patchSize) / downsample);
+		return rows * (cols + 1) + 2;
+	};
+	let { width, height } = size;
+	if (width * height < minPixels) {
+		const scale = Math.sqrt(minPixels / (width * height));
+		width = Math.trunc(width * scale);
+		height = Math.trunc(height * scale);
+	}
+	const tokens = spanTokens(Math.ceil(height / patchSize) * patchSize, Math.ceil(width / patchSize) * patchSize);
+	if (tokens <= maxTokens) return tokens;
+	const aspect = height / width;
+	const maxCols = Math.sqrt((maxTokens - 2) / aspect + 0.25) - 0.5;
+	const maxRows = maxCols * aspect;
+	if (maxCols < 1) return spanTokens(Math.floor((maxTokens - 2) / 2) * cell, cell);
+	if (maxRows < 1) return spanTokens(cell, (maxTokens - 3) * cell);
+	const beta = Math.min((Math.floor(maxCols) * cell) / width, (Math.floor(maxRows) * cell) / height);
+	return spanTokens(
+		Math.floor((height * beta) / patchSize) * patchSize,
+		Math.floor((width * beta) / patchSize) * patchSize,
+	);
+}
+
 /** Billed input tokens for one image of `size` under `rule`; only `openai-patch` reads `detail`. */
 export function imageTokens(rule: ImageTokenization, size: ImageSize, detail?: ImageDetail): number {
 	switch (rule.regime) {
@@ -180,5 +232,7 @@ export function imageTokens(rule: ImageTokenization, size: ImageSize, detail?: I
 		}
 		case "anthropic-patch":
 			return anthropicTokens(size, rule.maxEdge, rule.maxTokens);
+		case "deepseek-patch":
+			return deepseekTokens(size, rule);
 	}
 }

package/src/compat/provider-ids.ts

@@ -39,6 +39,7 @@ export type KnownProvider =
 	| "groq"
 	| "helmcode"
 	| "huggingface"
+	| "inception"
 	| "kilo"
 	| "kimi-code"
 	| "litellm"

package/src/compat/request-size.ts

@@ -9,11 +9,11 @@ const IMAGE_BUDGET_CACHE_MAX = 8192;
 /**
  * Maximum encoded inline-image bytes retained per request for a deployment with a documented body-size limit.
  * The axis is a host contract: a rerouted endpoint (proxy, gateway) imposes its own limit, so the budget applies
- * only when `officialEndpoint` says the request reaches the model's official endpoint. `compat.officialEndpoint`
- * classifies only the configured `baseUrl`; pi-ai's `reachesOfficialEndpoint` also follows request-time reroutes.
+ * only when `budgetedHost` says the request reaches the host the rules describe: pi-ai's `reachesBudgetedHost`
+ * (the model's official endpoint, following request-time reroutes, or Factory's hosted API for a Factory Droid model).
  */
-export function resolveInlineImageByteBudget(model: Model, officialEndpoint: boolean): number | undefined {
-	if (!officialEndpoint) return undefined;
+export function resolveInlineImageByteBudget(model: Model, budgetedHost: boolean): number | undefined {
+	if (!budgetedHost) return undefined;
 	const key = `${model.provider} ${model.id} ${model.api}`;
 	let budget = imageBudgetCache.get(key);
 	if (budget === undefined) {
@@ -25,16 +25,26 @@ export function resolveInlineImageByteBudget(model: Model, officialEndpoint: boo
 	return budget ?? undefined;
 }
 
+/**
+ * Largest request body the deployment accepts (`request-body-byte-limit`), or `undefined` when no rule names it.
+ * Same gate as {@link resolveInlineImageByteBudget}.
+ */
+export function resolveRequestBodyByteLimit(model: Model, budgetedHost: boolean): number | undefined {
+	if (!budgetedHost) return undefined;
+	const value = resolveCatalogPolicy(model).requestBodyByteLimit;
+	return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : undefined;
+}
+
 /**
  * Most images one request to this deployment may carry, or `undefined` when no rule names the host. A configured
  * `model.imageCountBudget` wins: the user knows the host behind a custom gateway. Otherwise the
- * `official-image-count-budget` axis replaces `image-count-budget` only when `officialEndpoint` says the request
- * reaches the model's official endpoint (see {@link resolveInlineImageByteBudget}); a rerouted endpoint keeps the
+ * `official-image-count-budget` axis replaces `image-count-budget` only when `budgetedHost` says the request
+ * reaches the host the rules describe (see {@link resolveInlineImageByteBudget}); a rerouted endpoint keeps the
  * provider's generic budget.
  */
-export function resolveImageCountBudget(model: Model, officialEndpoint: boolean): number | undefined {
+export function resolveImageCountBudget(model: Model, budgetedHost: boolean): number | undefined {
 	if (model.imageCountBudget !== undefined) return model.imageCountBudget;
 	const policy = resolveCatalogPolicy(model);
-	const value = (officialEndpoint && policy.officialImageCountBudget) || policy.imageCountBudget;
+	const value = (budgetedHost && policy.officialImageCountBudget) || policy.imageCountBudget;
 	return typeof value === "number" && Number.isInteger(value) && value > 0 ? value : undefined;
 }

package/src/compat/resolve.ts

@@ -952,6 +952,7 @@ function resolveBedrockPolicy(spec: ModelSpec<"bedrock-converse-stream">, axes:
 	// no-checkpoint shape.
 	const compat: ResolvedBedrockCompat = {
 		promptCacheMode: "none",
+		supportsServerCompaction: false,
 		supportsLongPromptCacheRetention: false,
 		promptCacheMinimumTokens: 0,
 		promptCacheMaximumCheckpoints: 0,

package/src/compat/rules/README.md

@@ -225,9 +225,13 @@ One object axis carries a computed form: `long-context-cost` accepts either the
 
 `max-tokens-fallback` seeds `maxTokens` only when the row reports none, and `max-tokens-ceiling` lowers a larger reported `maxTokens` to the host's output cap (an unknown budget stays unknown unless a fallback seeded it). Both apply in `buildModel` after `limits-patch`, so discovery mappers pass upstream output budgets through verbatim. See the Kimi rules in [`providers/fireworks.kdl`](providers/fireworks.kdl).
 
-`inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history, and snapcompact gives its archive frames half of it (`frameDataBytesBudget`). The budget applies only while the request reaches the official endpoint (the Anthropic API, first-party OpenAI, or the Codex backend at its default URL), which callers decide with pi-ai's `reachesOfficialEndpoint`: `compat.officialEndpoint` for the configured `baseUrl`, and for first-party Anthropic models also the `ANTHROPIC_BASE_URL` or Foundry `FOUNDRY_BASE_URL` reroute applied at request time. A custom or rerouted endpoint imposes its own body limit.
+`inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history, and snapcompact gives its archive frames half of it (`frameDataBytesBudget`). The budget applies only while the request reaches the host the rules describe (the Anthropic API, first-party OpenAI, the Codex backend at its default URL, or Factory Droid's own global or EU API host), which callers decide with pi-ai's `reachesBudgetedHost`: `compat.officialEndpoint` for the configured `baseUrl`, for first-party Anthropic models also the `ANTHROPIC_BASE_URL` or Foundry `FOUNDRY_BASE_URL` reroute applied at request time, and for Factory Droid models the `baseUrl` origin. A custom or rerouted endpoint imposes its own body limit.
 
-`image-count-budget` is the most images one request may carry, archive frames, screenshots and imaged tool results combined. The coding agent reads it through `resolveImageCountBudget`, drops the oldest images past it, and caps snapcompact frames at it; unassigned hosts get a floor of 5. `official-image-count-budget` replaces it while the request reaches the official endpoint (same gate as `inline-image-byte-budget`), for limits established on the host's own endpoint (Anthropic: 600, 100 on 200k-window models; Codex: 2,000); a custom or rerouted endpoint keeps `image-count-budget` (Anthropic 90, Codex 200).
+`request-body-byte-limit` is the largest request body the host accepts (Factory Droid's hosted routes: 4,500,000 bytes), under the same gate. Where it is set, the byte clamp keeps the inline images of the newest user message and then of the newest image-bearing tool result even past `inline-image-byte-budget`, each only while the kept images fit the limit less the request's other serialized bytes, and evicts older images instead; read through `resolveRequestBodyByteLimit`.
+
+`catalog-endpoint-budgets #true` makes these host budgets apply whenever a request reaches the provider's own catalog endpoint (an unset `baseUrl`, or an `endpoint-hosts` host; see `isCatalogEndpointUrl`), for a hosted gateway such as Factory Droid whose compat names no official endpoint. Unassigned, they apply only on the official endpoint.
+
+`image-count-budget` is the most images one request may carry, archive frames, screenshots and imaged tool results combined. The coding agent reads it through `resolveImageCountBudget`, drops the oldest images past it, and caps snapcompact frames at it; unassigned hosts get a floor of 5. `official-image-count-budget` replaces it while the request reaches the host the rules describe (same gate as `inline-image-byte-budget`), for limits established on the host's own endpoint (Anthropic: 600, 100 on 200k-window models; Codex: 2,000); a custom or rerouted endpoint keeps `image-count-budget` (Anthropic 90, Codex 200).
 
 `endpoint-hosts` lists the HTTPS hostnames a provider serves besides its bundled base URL's host, such as Copilot's per-plan `api.<plan>.githubcopilot.com` hosts. That host plus these define the provider's own endpoint (`isCatalogEndpointUrl`), which separates first-party traffic from a custom model or provider override that keeps a catalog provider id but points `baseUrl` at a gateway, proxy, or local server. Connection prewarming and request compression act only on the provider's own endpoint.
 

package/src/compat/rules/auth/_order.kdl

@@ -5,7 +5,7 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"cursor" "devin" "google-antigravity" "google-gemini-cli" "openai-codex-device" "xai" "xai-oauth" \
 	"gitlab-duo" "gitlab-duo-agent" "alibaba-coding-plan" "alibaba-token-plan" "aiand" "abliteration" "zhipu-coding-plan" \
 	"umans" "qwen-portal" "sakana" "sambanova" "minimax-code" "minimax-code-cn" "xiaomi" "xiaomi-token-plan-sgp" \
-	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "factory-droid" "commandcode" "charm-hyper" "deepseek" "stepfun" "muse-code" "meta" "moonshot" \
+	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "factory-droid" "commandcode" "charm-hyper" "deepseek" "stepfun" "muse-code" "meta" "moonshot" "inception" \
 	"cerebras" "baseten" "helmcode" "fireworks" "together" "nvidia" "novita" "deepinfra" "huggingface" "perplexity" \
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
 	"vercel-ai-gateway" "cloudflare-ai-gateway" "cloudflare-workers-ai" "snowflake" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \

package/src/compat/rules/auth/inception.kdl

@@ -0,0 +1,11 @@
+auth "inception" {
+	name "Inception"
+	login "api-key" {
+		auth-url "https://platform.inceptionlabs.ai/dashboard/api-keys"
+		instructions "Create or copy your key from the Inception dashboard"
+		prompt "Paste your Inception API key" placeholder="sk_..."
+		// Validate against inference, not /models: Inception's models endpoints
+		// are public and answer 200 for any key. Costs a few tokens.
+		validate "chat-completions" label="Inception" base-url="https://api.inceptionlabs.ai/v1" model="mercury-3" max-tokens-field="max_completion_tokens"
+	}
+}

package/src/compat/rules/classes/anthropic.kdl

@@ -163,34 +163,36 @@ class "anthropic" {
 	// Foundry / Claude Platform on AWS are first-party Anthropic routes whose
 	// effective URLs are checked at request time.
 	on "anthropic" "google-vertex" "amazon-bedrock" "bedrock-mantle" {
-		family "fable" {
-			revision ">=5" {
-				supports-server-compaction #true
+		on-api "anthropic-messages" {
+			family "fable" {
+				revision ">=5" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "mythos" {
-			revision ">=5" {
-				supports-server-compaction #true
+			family "mythos" {
+				revision ">=5" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "opus" {
-			revision ">=4.6" {
-				supports-server-compaction #true
+			family "opus" {
+				revision ">=4.6" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "sonnet" {
-			revision ">=4.6" {
-				supports-server-compaction #true
+			family "sonnet" {
+				revision ">=4.6" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "haiku" {
-			revision ">=5.5 <6" {
+			family "haiku" {
+				revision ">=5.5 <6" {
+					supports-server-compaction #true
+				}
+			}
+			models "claude-mythos-preview" {
 				supports-server-compaction #true
 			}
 		}
-		models "claude-mythos-preview" {
-			supports-server-compaction #true
-		}
 	}
 	// Sonnet 4.5 and newer end generation with `model_context_window_exceeded`
 	// when prompt + max_tokens overruns the window, instead of a 400, so the

package/src/compat/rules/classes/deepseek.kdl

@@ -108,4 +108,35 @@ class "deepseek" {
 	revision ">=4" {
 		portable-reasoning #true
 	}
+	// V4.1 Flash's vision config (patch 14, 3x3 merge, 1024-token cap, 544²
+	// floor). verified: OpenRouter billing matched it exactly on Fireworks and
+	// DeepInfra and within one token on Together, 256px to 4000px.
+	// Later Flash releases stay on their wire fallback until measured.
+	family "flash" {
+		revision ">=4.1 <4.2" {
+			image-tokenization {
+				regime "deepseek-patch"
+				verified #true
+				patch-size 14
+				downsample 3
+				max-tokens 1024
+				min-pixels 295936
+			}
+		}
+	}
+	// DeepSeek's API retired the V4.0 vision preview and serves its legacy id
+	// with the current Flash (api-docs.deepseek.com/guides/vision); hosts that
+	// still run the preview keep their fallback.
+	on "deepseek" {
+		models "deepseek-v4-flash-vision-exp" {
+			image-tokenization {
+				regime "deepseek-patch"
+				verified #true
+				patch-size 14
+				downsample 3
+				max-tokens 1024
+				min-pixels 295936
+			}
+		}
+	}
 }