@oh-my-pi/pi-catalog 18.14.0 → 18.14.2

package/CHANGELOG.md

@@ -2,6 +2,28 @@
 
 ## [Unreleased]
 
+## [18.14.2] - 2026-10-31
+
+### Breaking Changes
+
+- Removed `thinkingLoopGuard` from compat records; `thinking-loop-guard` is now a catalog axis that applies on every host, with `"semantic"` for Gemini, DeepSeek, and Grok and `"off"` for Claude
+
+### Fixed
+
+- Fixed GitHub Copilot Claude 5.x base models (`claude-opus-5`, `claude-opus-5.5`, `claude-sonnet-5`, `claude-sonnet-5.5`) reporting the 1M window of their `-1m` variants once the model list was read from cache; discovered limits now survive cached reads instead of reverting to bundled values ([#17481](https://github.com/can1357/oh-my-pi/issues/17481))
+
+## [18.14.1] - 2026-10-31
+
+### Changed
+
+- Clamping a requested thinking effort now resolves through the model's `effortMap` alias first, so the request lands on the tier the model serves instead of the nearest lower one ([#17395](https://github.com/can1357/oh-my-pi/pull/17395) by [@Maziar123](https://github.com/Maziar123))
+- Long image-heavy conversations on the official Anthropic API now keep up to 120 MB of inline images before evicting older ones, following the Messages API request limit increase from 32 MB to 128 MB
+- The official Anthropic API now records its 128 MB request limit (`request-body-byte-limit`), so the newest message's images can use the 8 MB reserve and long text cannot push images past the limit ([#17492](https://github.com/can1357/oh-my-pi/pull/17492) by [@will-bogusz](https://github.com/will-bogusz))
+
+### Fixed
+
+- Fixed Cursor's DeepSeek V4.1 Flash 1017 reporting zero cost; every effort level now bills at Cursor's API rate of $0.30/$1.20 per million tokens, with cache reads at $0.006 ([#17390](https://github.com/can1357/oh-my-pi/pull/17390) by [@eggpeat](https://github.com/eggpeat))
+
 ## [18.14.0] - 2026-10-29
 
 ### Breaking Changes

package/dist/types/discovery/cursor-proto.d.ts

@@ -2514,6 +2514,9 @@ export interface InteractionUpdate extends ProtoMessage {
     } | {
         case: "stepCompleted";
         value: StepCompletedUpdate;
+    } | {
+        case: "toolRequestsListed";
+        value: ToolRequestsListedUpdate;
     };
 }
 export declare const InteractionUpdateSchema: MessageCodec<InteractionUpdate>;
@@ -5318,6 +5321,11 @@ export interface ToolCallStartedUpdate extends ProtoMessage {
     modelCallId: string;
 }
 export declare const ToolCallStartedUpdateSchema: MessageCodec<ToolCallStartedUpdate>;
+/** Cursor agent message agent.v1.ToolRequestsListedUpdate. */
+export interface ToolRequestsListedUpdate extends ProtoMessage {
+    callCount: number;
+}
+export declare const ToolRequestsListedUpdateSchema: MessageCodec<ToolRequestsListedUpdate>;
 /** Cursor agent message agent.v1.TruncatedToolCall. */
 export interface TruncatedToolCall extends ProtoMessage {
     originalStepBlobId: Uint8Array;

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

@@ -129,9 +129,19 @@ export declare function createModelManager<TApi extends Api = Api, TModelsDevPay
  * when at least one configured remote source did not refresh successfully.
  */
 export declare function resolveProviderModels<TApi extends Api = Api, TModelsDevPayload = unknown>(options: ModelManagerOptions<TApi, TModelsDevPayload>, strategy?: ModelRefreshStrategy): Promise<ModelResolutionResult<TApi>>;
+/**
+ * Merge policy that is part of a cache's identity in
+ * {@link fingerprintStaticModels}: `authoritative` when dynamic discovery owns
+ * the catalog, `additive` when the shared catalog may only add ids. The flags
+ * compose; snapshots written under different policies never match.
+ */
+export interface StaticFingerprintPolicy {
+    authoritative?: boolean;
+    additive?: boolean;
+}
 /**
  * Return the versioned content identity for a static provider slice. Callers
  * may mutate or freeze their arrays, so every call observes current content
  * and never attaches a stale sidecar fingerprint to caller-owned data.
  */
-export declare function fingerprintStaticModels<TApi extends Api>(models: readonly (ModelSpec<TApi> | Model<TApi>)[], dynamicModelsAuthoritative?: boolean): string;
+export declare function fingerprintStaticModels<TApi extends Api>(models: readonly (ModelSpec<TApi> | Model<TApi>)[], policy?: StaticFingerprintPolicy): string;

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

@@ -32,6 +32,10 @@ export declare function withoutUnverifiedThinking<T extends {
 /**
  * Clamps a requested thinking level against explicit model metadata.
  *
+ * An unserved level first follows `thinking.effortMap` when the map sends it to
+ * a served tier that reaches the wire unchanged (a gateway answering `medium`
+ * as `high`); otherwise it climbs down to the nearest lower tier.
+ *
  * Non-reasoning models always resolve to `undefined`.
  */
 export declare function clampThinkingLevelForModel<TApi extends Api>(model: ApiModel<TApi> | undefined, requested: Effort | undefined): Effort | undefined;

package/dist/types/types.d.ts

@@ -477,8 +477,6 @@ export interface OpenAICompat {
     clampOutputToModelMax?: boolean;
     /** Strip image inputs before encoding (text-only serving of a multimodal id). */
     stripImageInput?: boolean;
-    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
-    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
 }
 /**
  * Compatibility settings for anthropic-messages API.
@@ -676,8 +674,6 @@ export interface AnthropicCompat {
      * contract override it. Default: 10485760.
      */
     maxImagePayloadBytes?: number;
-    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
-    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
     /**
      * Drop the enabled `thinking` config and replayed thinking blocks when the
      * conversation is not thinking-led (a proxy contract for non-interleaved
@@ -887,8 +883,6 @@ export interface ResolvedOpenAISharedCompat {
     toolSchemaFlavor?: OpenAICompat["toolSchemaFlavor"];
     /** Strip image inputs before encoding (text-only serving of a multimodal id). */
     stripImageInput: boolean;
-    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
-    thinkingLoopGuard?: OpenAICompat["thinkingLoopGuard"];
     /** Flatten/reject leftover root `anyOf`/`oneOf` unions in strict tool schemas (xAI's function-calling validator 400s on them). */
     rejectRootObjectUnion: boolean;
     /** Retry without strict tools when the host rejects a strict grammar as too large (OpenRouter-Anthropic compiled-grammar overflow). */
@@ -900,7 +894,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "stripThinkingHistory" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "stripThinkingHistory" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -992,7 +986,7 @@ 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" | "extendedCacheTtlBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "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. */
@@ -1016,8 +1010,6 @@ export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdle
      * request builder.
      */
     maxImagePayloadBytes?: number;
-    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
-    thinkingLoopGuard?: AnthropicCompat["thinkingLoopGuard"];
     /**
      * Bedrock's Anthropic Messages API (see {@link AnthropicCompat.bedrockMessagesApi}).
      * `true` when detected or configured; otherwise unset, so rows baked before
@@ -1116,16 +1108,13 @@ export interface GoogleCompat {
      * {@link AnthropicCompat.maxImagePayloadBytes}).
      */
     maxImagePayloadBytes?: number;
-    /** Thinking-loop watchdog guard family applied to streamed reasoning. */
-    thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
     /** See {@link BedrockCompat.supportsSamplingParams}. */
     supportsSamplingParams?: boolean;
 }
 /** Fully-resolved google-API compat view, materialized once by `buildModel`. */
-export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel" | "supportsSamplingParams" | "maxImageDimension" | "maxImagePayloadBytes">> & {
+export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "antigravityUsageLabel" | "supportsSamplingParams" | "maxImageDimension" | "maxImagePayloadBytes">> & {
     streamFirstEventTimeoutMs?: number;
     streamIdleTimeoutMs?: number;
-    thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
     antigravityUsageLabel?: string;
     supportsSamplingParams?: boolean;
     maxImageDimension?: number;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.14.0",
+	"version": "18.14.2",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -88,12 +88,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.14.0",
-		"@oh-my-pi/pi-utils": "18.14.0"
+		"@oh-my-pi/omptype": "18.14.2",
+		"@oh-my-pi/pi-utils": "18.14.2"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.14.0",
+		"@oh-my-pi/pi-ai": "18.14.2",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -317,11 +317,6 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"strip-image-input": wire("stripImageInput", [...OAI, "anthropic", "google"]),
 	"supports-forced-tool-choice": wire("supportsForcedToolChoice", [...OAI, "anthropic", "bedrock"]),
 	"supports-sampling-params": wire("supportsSamplingParams", [...OAI, "anthropic", "bedrock", "devin", "google"]),
-	"thinking-loop-guard": wire("thinkingLoopGuard", [...OAI, "anthropic", "google"], "scalar", [
-		"gemini",
-		"deepseek",
-		"xai",
-	]),
 
 	// ── thinking control surface ──
 	"thinking-default-level": { key: "defaultLevel", set: "thinking", shape: "scalar", values: EFFORTS },
@@ -456,6 +451,19 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 		shape: "scalar",
 		values: ["output", "signature"],
 	},
+	/**
+	 * Stream loop guard for the model line, on every host and API (read
+	 * through `loopGuardLevel` in pi-ai `utils/thinking-loop`). `semantic`
+	 * adds near-duplicate and stalled-wording paragraph detection to the
+	 * exact repeated-cycle check; `off` disables the guard. Unassigned: exact
+	 * repeated cycles only.
+	 */
+	"thinking-loop-guard": {
+		key: "thinkingLoopGuard",
+		set: "catalog",
+		shape: "scalar",
+		values: ["off", "semantic"],
+	},
 	"delegation-bias": { key: "delegationBias", set: "catalog", shape: "scalar", values: DELEGATION_BIASES },
 	"discovery-api": { key: "discoveryApi", set: "catalog", shape: "scalar" },
 	"edit-prompt-variant": { key: "editPromptVariant", set: "catalog", shape: "scalar", values: ["full", "compact"] },
@@ -502,8 +510,9 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"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.
+	 * Largest request body the host accepts. Where set, the inline-image byte clamp keeps images within what the
+	 * request's other bytes leave of it, and keeps the newest user message's and newest tool result's images past
+	 * `inline-image-byte-budget` while they fit that room.
 	 */
 	"request-body-byte-limit": { key: "requestBodyByteLimit", set: "catalog", shape: "scalar" },
 	/**

package/src/compat/resolve.ts

@@ -604,7 +604,6 @@ function detectOpenAICompat(
 		zaiReasoningEffortDialect: false,
 		clampOutputToModelMax: d.isLocalOpenAICompatBackend,
 		stripImageInput: false,
-		thinkingLoopGuard: undefined,
 		rejectRootObjectUnion: false,
 		retryWithoutStrictOnGrammarError: false,
 		supportsPromptCacheKey: false,
@@ -762,7 +761,6 @@ function resolveOpenAIResponsesPolicy(
 		supportsSteering: false,
 		requiresReasoningOffJuiceInstruction: false,
 		stripImageInput: false,
-		thinkingLoopGuard: undefined,
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
 		supportsSamplingParams: true,
@@ -936,7 +934,6 @@ function resolveAnthropicPolicy(
 		// user-authored `compat` block to override them.
 		maxImageDimension: undefined,
 		maxImagePayloadBytes: undefined,
-		thinkingLoopGuard: undefined,
 		// Present as keys so models.yml `compat` can set them; unset unless a rule assigns them.
 		stripThinkingHistory: undefined,
 		fastMode: undefined,

package/src/compat/rules/README.md

@@ -227,7 +227,7 @@ One object axis carries a computed form: `long-context-cost` accepts either the
 
 `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, DeepSeek's API, 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.
 
-`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`.
+`request-body-byte-limit` is the largest request body the host accepts (Factory Droid's hosted routes: 4,500,000 bytes; Anthropic's Messages API: 128,000,000 bytes), under the same gate. Where it is set, the byte clamp also keeps inline images within what the request's other serialized bytes leave of the limit, and 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, or a first-party provider such as DeepSeek whose `openai-completions` compat resolves no `officialEndpoint`. Unassigned, they apply only on the official endpoint.
 

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

@@ -29,6 +29,11 @@ class "anthropic" {
 	// back as text (`reasoning_extraction`, issue #6093).
 	classifier-cot-extraction #true
 	thinking-replay-tokens "output"
+	// Summarized thinking repeats one progress line verbatim while long hidden
+	// reasoning runs ("Still writing the parser..." six-plus times back to
+	// back), which reads as an exact cycle. Claude has no runaway-loop failure
+	// that warrants the guard, so it never discards a turn here.
+	thinking-loop-guard "off"
 	// verified: billing matched this formula on Opus 4.7+ and Fable 5.1.
 	image-tokenization {
 		regime "anthropic-patch"

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

@@ -102,7 +102,7 @@ class "deepseek" {
 	// leaks it into content — hosted gateways, local backends, and custom
 	// providers alike. On hosts that parse it the healer sees no markup.
 	stream-markup-healing-pattern "dsml"
-	thinking-loop-guard "deepseek"
+	thinking-loop-guard "semantic"
 	// V4 reasoning reaches omp as its plaintext trace on every host, and the
 	// revision pins the weights (flash 4.0 / 4.1, pro 4.0). Earlier lineages
 	// classify too coarsely to carry: `r1` also matches the R1 distills.

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

@@ -149,7 +149,7 @@ class "gemini" {
 		}
 	}
 
-	thinking-loop-guard "gemini"
+	thinking-loop-guard "semantic"
 }
 
 // Google Search grounding is a Gemini 2+ lineage capability on every Gemini transport.

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

@@ -67,7 +67,7 @@ class "xai" {
 			}
 		}
 	}
-	thinking-loop-guard "xai"
+	thinking-loop-guard "semantic"
 }
 
 // Same-provider swap targets for hosted search and image generation.

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

@@ -78,12 +78,16 @@ provider "anthropic" {
 	// Code's `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS` kill switch for gateways
 	// that reject experimental beta fields.
 	honors-claude-code-beta-kill-switch #true
-	// Messages requests cap at 32 MB. Reserve 8 MB for text, tools, and JSON
-	// framing when retaining inline base64 images in a long conversation.
-	// Applied only on the official endpoint; a custom baseUrl has its own limit.
+	// Messages requests cap at 128 MB. Reserve 8 MB for text, tools, and JSON
+	// framing when retaining inline base64 images in a long conversation; the
+	// body limit keeps images within what the other bytes leave, and lets the
+	// newest user message's and tool result's images use the reserve.
+	// Applied only on the official endpoint; a custom baseUrl has its own limit
+	// (Bedrock 20 MB; Vertex, Foundry, and Claude Platform on AWS 30-32 MB).
 	// https://platform.claude.com/docs/en/api/overview#request-size-limits
 	on-api "anthropic-messages" {
-		inline-image-byte-budget 24000000
+		inline-image-byte-budget 120000000
+		request-body-byte-limit 128000000
 	}
 	// The official endpoint carries at most 600 images and documents per
 	// request, 100 on models with a 200k-token window: every Claude line before

package/src/compat/rules/providers/cursor.kdl

@@ -387,6 +387,16 @@ provider "cursor" {
 			cache-write 0
 		}
 	}
+	// DeepSeek V4.1 Flash 1017: $0.30/$1.20, cache read $0.006, no write fee.
+	// The bare id and every reasoning-effort wire id share one card.
+	models "deepseek-v41-flash-1017*" {
+		cost-patch {
+			input 0.3
+			output 1.2
+			cache-read 0.006
+			cache-write 0
+		}
+	}
 	// Gemini 3 Flash: $0.50/$3.00, cache read $0.05.
 	models "gemini-3-flash" {
 		cost-patch {

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

@@ -7,8 +7,8 @@ provider "deepseek" {
 
 	// DeepSeek's vision guide documents 600 images per request and a 48 MiB
 	// request body. OMP reserves 8 MiB for text, tools, and the JSON envelope
-	// (48 MiB - 8 MiB = 40 MiB); Anthropic's 24,000,000-byte image budget under
-	// its 32 MB body uses an 8 MB reserve. DeepSeek documents MiB, so this uses
+	// (48 MiB - 8 MiB = 40 MiB); Anthropic's 120,000,000-byte image budget under
+	// its 128 MB body uses an 8 MB reserve. DeepSeek documents MiB, so this uses
 	// an 8 MiB reserve. The split is OMP's conservative margin, not a
 	// DeepSeek-documented figure. A custom baseUrl keeps the conservative
 	// catalog floor because the gateway's own limits are unknown.

package/src/compat/rules/providers/openai-codex.kdl

@@ -16,7 +16,7 @@ provider "openai-codex" {
 
 	// Policy budgets: the Codex backend documents no request-size or image
 	// limit, and requests with several thousand images fail mid-stream at
-	// random. Inline images keep Anthropic's 24 MB; the official endpoint's
+	// random. Inline images keep a 24 MB budget; the official endpoint's
 	// image count stays well below those failures. A custom baseUrl keeps 200.
 	inline-image-byte-budget 24000000
 	image-count-budget 200