@oh-my-pi/pi-catalog 18.2.0 → 18.2.1

package/CHANGELOG.md

@@ -2,6 +2,33 @@
 
 ## [Unreleased]
 
+## [18.2.1] - 2026-09-15
+
+### Added
+
+- DeepSeek V4.1 Flash (`deepseek-flash`) now accepts image inputs and resolves its reasoning metadata, name, and effort ladder from upstream instead of shipping as a text-only row without them.
+
+### Fixed
+
+- Fixed Cerebras Qwen 3.8 reasoning control by enabling correct effort tiers and disable behavior
+- Fixed Cerebras Qwen models using incorrect reasoning format, enabling proper support for thinking modes
+- Fixed Command Code's `deepseek/deepseek-v4.1-flash` row exposing its documented low/high/max thinking levels and image input, and made live discovery resolve the reasoning contract its rules declare ([#1666](https://github.com/can1357/oh-my-pi/issues/1666), [#11703](https://github.com/can1357/oh-my-pi/pull/11703) by [@aliefe04](https://github.com/aliefe04)).
+- Fixed Meta Model API and Muse Code requests failing with 400 whenever omp forced a tool choice: `api.meta.ai/v1` accepts only `tool_choice: "auto"`, so `none`, `required`, and named choices (subagent final-retry `yield`, forced tools, structured output, compaction handoff) are now omitted instead of sent ([#11635](https://github.com/can1357/oh-my-pi/pull/11635) by [@quantmind-br](https://github.com/quantmind-br)).
+- Bedrock's Qwen rows no longer ask for more output tokens than the model accepts, which Bedrock rejected with a 400. ([#12117](https://github.com/can1357/oh-my-pi/pull/12117) by [@Huang-404-Q](https://github.com/Huang-404-Q))
+- Gemini 2.5 Flash Lite on Vertex AI no longer requests `maxOutputTokens=65536`, which the endpoint rejects with a 400; the output cap is clamped to 65535 ([#10595](https://github.com/can1357/oh-my-pi/pull/10595) by [@WeMingT](https://github.com/WeMingT)).
+- Fixed Meta muse-spark models on OpenRouter wedging every turn with `400 Referenced reasoning item ... was not found or has expired`: replayed Responses reasoning history is now filtered for the muse-spark family, matching the existing Anthropic-on-OpenRouter treatment ([#10675](https://github.com/can1357/oh-my-pi/pull/10675) by [@Giardi77](https://github.com/Giardi77)).
+- Fixed Ollama Cloud model discovery synthesizing a generic `minimal`/`low`/`medium`/`high` effort ladder for every thinking-capable model, which shadowed the per-model compat rules and made `max` unreachable on the DeepSeek V4 line (including the served `deepseek-v4.1-flash`, `deepseek-v4-flash:0731`, and `deepseek-v4-pro:0813` ids): discovery now leaves the ladder to the rule tree, so those models advertise the wire-exact `low`/`high`/`max` and GLM-5.3 exposes `low`/`high`/`max` ([#8334](https://github.com/can1357/oh-my-pi/issues/8334)).
+- Fixed Meta Model API and Muse Code requests failing with 400 whenever omp forced a tool choice: `api.meta.ai/v1` accepts only `tool_choice: "auto"`, so `none`, `required`, and named choices (subagent final-retry `yield`, forced tools, structured output, compaction handoff) are now omitted instead of sent.
+- Fixed Azure GPT-6 Astra Chat Completions policy to disable reasoning with function tools and encode thinking-off as `reasoning_effort: "none"` ([#11052](https://github.com/can1357/oh-my-pi/issues/11052)).
+- Fixed Z.AI and Zhipu timezone-naive quota reset timestamps resolving eight hours late by declaring their UTC+8 reset timezone ([#11014](https://github.com/can1357/oh-my-pi/issues/11014)).
+- Fixed Amazon Bedrock Claude Fable 5.1 effort metadata to expose `xhigh` and `max` instead of the unsupported `minimal` level. ([#10788](https://github.com/can1357/oh-my-pi/pull/10788) by [@voonfoo](https://github.com/voonfoo))
+- LiteLLM discovery no longer exposes known task-specific models, including embedding, media, moderation, reranking, and search models, as coding models.
+- Fixed custom `anthropic-messages` providers whose adaptive effort ladder exposes `minimal` sending `output_config.effort: "minimal"`, which the Anthropic Messages API rejects with `400 level "minimal" not supported`; the adaptive effort mapper now clamps `minimal` to `low` ([#10994](https://github.com/can1357/oh-my-pi/issues/10994)).
+- Qwen 3.8 models on Alibaba Token Plan now send the selected reasoning effort and replay reasoning history across turns.
+- Fixed DeepInfra model cost reporting so promotional pricing is reflected: the `metadata.discount` fraction is now applied to input, output, and cache-read rates ([#10935](https://github.com/can1357/oh-my-pi/issues/10935)).
+- Fixed LiteLLM model discovery leaking ClinePass display names and pricing into models with colliding ids ([#10932](https://github.com/can1357/oh-my-pi/issues/10932)).
+- Raised Cursor context windows to Cursor's documented sizes (Grok 4.5/4.6 256k, default/Auto 256k, Kimi K2.7 Code 262k, GPT-5.6 272k, Claude Opus 5 and Fable 300k by default) so compaction no longer fires too early, without lowering existing 1M Max-mode windows.
+
 ## [18.2.0] - 2026-09-15
 
 ### Breaking Changes

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

@@ -29,6 +29,8 @@ export declare function cursorModelParameters(model: string): readonly {
  * preserve quota semantics for newly discovered ids.
  */
 export declare function quotaTierFor(provider: string, model: string): string | undefined;
+/** UTC offset for a provider's timezone-naive absolute retry-reset timestamps. */
+export declare function retryResetTimezoneOffsetFor(provider: string): string | undefined;
 /** Whether a provider has catalog-authored model quota scopes. */
 export declare function hasQuotaTierPolicy(provider: string): boolean;
 /** The provider-default wire model for a model-less hosted operation. */
@@ -55,11 +57,20 @@ export declare function modelLimitsFor(provider: string, model: string): {
 } | undefined;
 /** Whether a provider roster entry is a declared non-chat/unsupported SKU. */
 export declare function isExcludedModel(provider: string, model: string): boolean;
+/** Whether an exact upstream discovery mode is excluded from a provider's coding-model roster. */
+export declare function isExcludedDiscoveryMode(provider: string, mode: string): boolean;
 /**
  * Whether a provider is retired: its entire roster is excluded from the
  * generated bundle and must never be resurrected from the previous snapshot.
  */
 export declare function isRetiredProvider(provider: string): boolean;
+/**
+ * Whether a provider's bundled rows may seed cross-provider bare-id enrichment
+ * references. Gateway-namespaced providers (ClinePass) are isolated so their
+ * limits, pricing, and reasoning controls never contaminate an unrelated proxy
+ * model that happens to advertise the same bare id.
+ */
+export declare function isBareIdReferenceProvider(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/types.d.ts

@@ -272,6 +272,11 @@ export interface CompiledExcludeModels {
     provider: string;
     match: CompiledMatchList;
 }
+/** Exact upstream discovery modes excluded from one provider's coding-model roster. */
+export interface CompiledExcludeDiscoveryModes {
+    provider: string;
+    modes: string[];
+}
 /** Provider plan-requirement tiers keyed by matcher token lists. */
 export interface CompiledPlanRequirement {
     provider: string;
@@ -289,6 +294,11 @@ export interface CompiledPricingPeer {
         peerId: string;
     }[];
 }
+/** Provider timezone assumption for offset-less absolute retry-reset timestamps. */
+export interface CompiledRetryResetTimezone {
+    provider: string;
+    offset: string;
+}
 /** Compiled runtime behavior vocabulary (`runtime/behavior.kdl`). */
 export interface CompiledBehavior {
     openaiResponsesHeuristic?: CompiledResponsesHeuristic;
@@ -299,10 +309,13 @@ export interface CompiledBehavior {
     hostedDefaults: CompiledHostedDefault[];
     apiRoutes: CompiledApiRoutes[];
     modelLimits: CompiledModelLimits[];
+    excludeDiscoveryModes: CompiledExcludeDiscoveryModes[];
     excludeModels: CompiledExcludeModels[];
     planRequirements: CompiledPlanRequirement[];
     pricingPeers: CompiledPricingPeer[];
+    retryResetTimezones: CompiledRetryResetTimezone[];
     retiredProviders: string[];
+    referenceIsolatedProviders: string[];
 }
 /**
  * A string setting from `auth/*.kdl` that may be overridden by environment
@@ -647,4 +660,12 @@ export interface ResolvedAxes {
     wire: Record<string, unknown>;
     thinking: Record<string, unknown>;
     catalog: Record<string, unknown>;
+    /**
+     * Reasoning capability after the exact-model effort upgrade: `true` when the
+     * target reported reasoning or an exact rule declares a ladder for it (the
+     * reviewed correction to metadata-less discovery rows). Compat resolvers
+     * read this instead of the raw spec flag, or one id resolves two different
+     * wire contracts depending on whether it came from discovery or the bake.
+     */
+    reasoning: boolean;
 }

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

@@ -1,3 +1,4 @@
+import type { FetchImpl } from "@oh-my-pi/pi-utils";
 import { type VariantCollapseTable } from "../compat/collapse.js";
 import type { ModelSpec } from "../types.js";
 export declare const ANTIGRAVITY_PRIMARY_ENDPOINT = "https://daily-cloudcode-pa.googleapis.com";
@@ -37,6 +38,7 @@ export interface AntigravityDiscoveryAgentModelSort {
 export interface AntigravityDiscoveryApiResponse {
     models?: Record<string, AntigravityDiscoveryApiModel>;
     agentModelSorts?: AntigravityDiscoveryAgentModelSort[];
+    imageGenerationModelIds?: string[];
 }
 /**
  * Options for fetching Antigravity discovery models.
@@ -53,7 +55,7 @@ export interface FetchAntigravityDiscoveryModelsOptions {
     /** Optional abort signal for request cancellation. */
     signal?: AbortSignal;
     /** Optional fetch implementation override for tests. */
-    fetcher?: typeof fetch;
+    fetcher?: FetchImpl;
     /**
      * Hand collapse table to apply to the discovered list. Defaults to the
      * Antigravity (budget-transport) table; `googleGeminiCli` passes the
@@ -68,3 +70,10 @@ export interface FetchAntigravityDiscoveryModelsOptions {
  * Returns `[]` only when the endpoint responds successfully with no usable models.
  */
 export declare function fetchAntigravityDiscoveryModels(options: FetchAntigravityDiscoveryModelsOptions): Promise<ModelSpec<"google-gemini-cli">[] | null>;
+/** Advertised image model and serving endpoint for one Antigravity account. */
+export interface AntigravityImageModel {
+    id: string;
+    endpoint: string;
+}
+/** Resolves the first image-generation model advertised by an Antigravity account. */
+export declare function fetchAntigravityImageModel(options: FetchAntigravityDiscoveryModelsOptions): Promise<AntigravityImageModel | null>;

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

@@ -35,6 +35,13 @@ export declare function mapEffortToGoogleThinkingLevel<TApi extends Api>(effort:
 /**
  * Maps a normalized thinking effort to Anthropic adaptive effort values via
  * the model's baked `thinking.effortMap` (identity for unmapped efforts).
+ *
+ * The Anthropic adaptive wire vocabulary has no `minimal` tier (valid values:
+ * `low`, `medium`, `high`, `xhigh`, `max`); a model ladder that exposes
+ * `minimal` — custom `anthropic-messages` providers do, unlike the built-in
+ * Claude ladders — would otherwise serialize `output_config.effort: "minimal"`
+ * and 400 (`level "minimal" not supported`). Clamp it to the lowest real tier,
+ * mirroring {@link mapEffortToGoogleThinkingLevel}'s `minimal` handling.
  */
 export declare function mapEffortToAnthropicAdaptiveEffort<TApi extends Api>(model: ApiModel<TApi>, effort: Effort): "low" | "medium" | "high" | "xhigh" | "max" | "adaptive";
 /**

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

@@ -523,6 +523,8 @@ export interface FetchLiteLLMRichModelsOptions<TApi extends Api> {
 }
 export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_CONTEXT_WINDOW = 128000;
 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;
 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>;

package/dist/types/types.d.ts

@@ -278,6 +278,12 @@ export interface OpenAICompat {
      * Default: auto-detected (DeepSeek reasoning models).
      */
     disableReasoningOnToolChoice?: boolean;
+    /**
+     * Disable reasoning whenever the request advertises function tools.
+     * Use for model surfaces that reject every tools-plus-reasoning combination.
+     * Default: false.
+     */
+    disableReasoningWithTools?: boolean;
     /** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */
     openRouterRouting?: OpenRouterRouting;
     /** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */
@@ -372,12 +378,24 @@ export interface OpenAICompat {
     strictResponsesPairing?: boolean;
     /** Whether the Responses API accepts the `detail: "original"` image hint. Default: auto-detected (false for GitHub Copilot, which rejects it with a 400). */
     supportsImageDetailOriginal?: boolean;
+    /**
+     * Whether the Responses endpoint accepts `configuration_update` input items
+     * that change `reasoning.effort` mid-conversation while the request-level
+     * effort stays pinned for prompt caching (GPT-6 Astra). Default:
+     * rule-detected (`true` for `gpt-6-astra` on any host, `false` otherwise).
+     * Set `false` for custom `openai-responses` / `openai-codex-responses`
+     * endpoints that reject the item type with HTTP 400; effort changes are then
+     * sent as the top-level `reasoning.effort`.
+     */
+    supportsConfigurationUpdate?: boolean;
     /** Whether streamed reasoning deltas for the same field may repeat the full cumulative text snapshot. Default: false. */
     reasoningDeltasMayBeCumulative?: boolean;
     /** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
     stripDeepseekSpecialTokens?: boolean;
     /** Heal leaked chat-template/tool-call/thinking markup from visible content deltas. Default: auto-detected. */
     streamMarkupHealingPattern?: OpenAIStreamMarkupHealingPattern;
+    /** Whether this wire may revise already-streamed text (`stream-revision` axis). Unassigned: append-only. */
+    streamRevision?: "none" | "possible";
     /** Treat an empty length-finished stream as a context-window error. Default: auto-detected. */
     emptyLengthFinishIsContextError?: boolean;
     /** Normalize tool call ids to OpenAI's 40-character limit. Default: auto-detected. */
@@ -548,6 +566,8 @@ export interface AnthropicCompat {
 export interface BedrockCompat {
     /** Whether this endpoint accepts no checkpoints, automatic caching, or explicit cachePoint blocks. */
     promptCacheMode?: "none" | "automatic" | "explicit";
+    /** Whether this wire may revise already-streamed text (`stream-revision` axis). Unassigned: append-only. */
+    streamRevision?: "none" | "possible";
     /** Whether explicit cachePoint blocks accept `ttl: "1h"`; omitted TTL means Bedrock's 5-minute default. */
     supportsLongPromptCacheRetention?: boolean;
     /**
@@ -571,6 +591,8 @@ export interface BedrockCompat {
 /** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
 export interface ResolvedBedrockCompat {
     promptCacheMode: NonNullable<BedrockCompat["promptCacheMode"]>;
+    /** See {@link BedrockCompat.streamRevision}. */
+    streamRevision?: BedrockCompat["streamRevision"];
     supportsLongPromptCacheRetention: boolean;
     promptCacheMinimumTokens: number;
     promptCacheMaximumCheckpoints: number;
@@ -632,6 +654,7 @@ export interface ResolvedOpenAISharedCompat {
     filterReasoningHistory: boolean;
     disableReasoningOnForcedToolChoice: boolean;
     disableReasoningOnToolChoice: boolean;
+    disableReasoningWithTools?: boolean;
     supportsToolChoice: boolean;
     supportsForcedToolChoice: boolean;
     supportsNamedToolChoice: boolean;
@@ -649,6 +672,8 @@ export interface ResolvedOpenAISharedCompat {
     requiresAssistantContentForToolCalls: boolean;
     stripDeepseekSpecialTokens: boolean;
     streamMarkupHealingPattern?: OpenAIStreamMarkupHealingPattern;
+    /** See {@link OpenAICompat.streamRevision}. */
+    streamRevision?: OpenAICompat["streamRevision"];
     /** See {@link OpenAICompat.streamFirstEventTimeoutMs}. */
     streamFirstEventTimeoutMs?: number;
     reasoningDeltasMayBeCumulative: boolean;
@@ -689,7 +714,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.2.0",
+	"version": "18.2.1",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -35,12 +35,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.2.0",
-		"@oh-my-pi/pi-utils": "18.2.0"
+		"@oh-my-pi/omptype": "18.2.1",
+		"@oh-my-pi/pi-utils": "18.2.1"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.2.0",
+		"@oh-my-pi/pi-ai": "18.2.1",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -93,6 +93,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"clamp-output-to-model-max": wire("clampOutputToModelMax", OAI),
 	"disable-reasoning-on-forced-tool-choice": wire("disableReasoningOnForcedToolChoice", OAI),
 	"disable-reasoning-on-tool-choice": wire("disableReasoningOnToolChoice", OAI),
+	"disable-reasoning-with-tools": wire("disableReasoningWithTools", ["openai"]),
 	"drop-thinking-when-reasoning-effort": wire("dropThinkingWhenReasoningEffort", ["openai"]),
 	"empty-length-finish-is-context-error": wire("emptyLengthFinishIsContextError", OAI),
 	"extra-body": { ...wire("extraBody", ["openai"], "object"), verbatimKeys: true },
@@ -235,6 +236,19 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"supports-function-part-id": wire("supportsFunctionPartId", ["google"]),
 
 	// ── wire: shared across surfaces ──
+	/**
+	 * Whether this wire may revise text it has already streamed: bytes
+	 * reclassified out of the visible channel (a leaned-on thinking opener),
+	 * carved into a tool call, reordered by content-block index, or replaced
+	 * wholesale by an authoritative final payload. Unassigned means the wire
+	 * only appends, so the transcript may retire finished lines into native
+	 * scrollback while the turn is still streaming (see
+	 * `AssistantMessageComponent`). Declare `possible` only with a citable
+	 * mechanism: the renderer also verifies published rows every frame and stops
+	 * retiring the block on the first mismatch, so this axis decides where
+	 * mid-stream retirement is attempted, not whether it is safe.
+	 */
+	"stream-revision": wire("streamRevision", [...OAI, "bedrock"], "scalar", ["none", "possible"]),
 	"stream-first-event-timeout-ms": wire("streamFirstEventTimeoutMs", [...OAI, "google"]),
 	"stream-idle-timeout-ms": wire("streamIdleTimeoutMs", [...OAI, "anthropic", "bedrock", "google"]),
 	"strip-image-input": wire("stripImageInput", [...OAI, "anthropic", "google"]),

package/src/compat/behavior.ts

@@ -2,8 +2,8 @@
  * Typed accessors over the compiled runtime-behavior vocabulary
  * (`rules/runtime/behavior.kdl`): provider/model heuristics that run before
  * or outside exact bundled-model lookup — responses routing, API routing,
- * quota tiers, plan requirements, model limits, roster exclusions, hosted
- * defaults, and pricing peers.
+ * quota tiers, plan requirements, model limits, roster and discovery-mode
+ * exclusions, hosted defaults, and pricing peers.
  */
 import { globMatch } from "./cascade";
 import rules from "./rules.json";
@@ -103,6 +103,11 @@ export function quotaTierFor(provider: string, model: string): string | undefine
 	return undefined;
 }
 
+/** UTC offset for a provider's timezone-naive absolute retry-reset timestamps. */
+export function retryResetTimezoneOffsetFor(provider: string): string | undefined {
+	return behavior.retryResetTimezones.find(rule => rule.provider === provider)?.offset;
+}
+
 /** Whether a provider has catalog-authored model quota scopes. */
 export function hasQuotaTierPolicy(provider: string): boolean {
 	return behavior.quotaTiers.some(rule => rule.provider === provider);
@@ -176,6 +181,11 @@ export function isExcludedModel(provider: string, model: string): boolean {
 	return behavior.excludeModels.some(rule => rule.provider === provider && matchesList(rule.match, lower, lower));
 }
 
+/** Whether an exact upstream discovery mode is excluded from a provider's coding-model roster. */
+export function isExcludedDiscoveryMode(provider: string, mode: string): boolean {
+	return behavior.excludeDiscoveryModes.some(rule => rule.provider === provider && rule.modes.includes(mode));
+}
+
 /**
  * Whether a provider is retired: its entire roster is excluded from the
  * generated bundle and must never be resurrected from the previous snapshot.
@@ -184,6 +194,16 @@ export function isRetiredProvider(provider: string): boolean {
 	return behavior.retiredProviders.includes(provider);
 }
 
+/**
+ * Whether a provider's bundled rows may seed cross-provider bare-id enrichment
+ * references. Gateway-namespaced providers (ClinePass) are isolated so their
+ * limits, pricing, and reasoning controls never contaminate an unrelated proxy
+ * model that happens to advertise the same bare id.
+ */
+export function isBareIdReferenceProvider(provider: string): boolean {
+	return !behavior.referenceIsolatedProviders.includes(provider);
+}
+
 /** The declared subscription tier required to use a provider model id, if any. */
 export function planRequirementFor(provider: string, model: string): string | undefined {
 	const rule = behavior.planRequirements.find(candidate => candidate.provider === provider);

package/src/compat/cascade.ts

@@ -293,6 +293,7 @@ function cloneAxes(axes: ResolvedAxes): ResolvedAxes {
 		wire: cloneAxisRecord(axes.wire),
 		thinking: cloneAxisRecord(axes.thinking),
 		catalog: cloneAxisRecord(axes.catalog),
+		reasoning: axes.reasoning,
 	};
 }
 
@@ -326,7 +327,7 @@ export function resolveCascadeRules(cascade: CompiledCascade, target: ResolveTar
 
 function resolveOverIndex(index: RuleIndex, target: ResolveTarget): ResolvedAxes {
 	const ranked = rankRelevantRules(index, prepareTarget(target));
-	let reasoning = target.reasoning;
+	let reasoning = target.reasoning === true;
 	if (!reasoning) {
 		for (const { rule, rank } of ranked) {
 			if (rule.hasExactEffortsRule && rank[0] === 2) {
@@ -347,5 +348,6 @@ function resolveOverIndex(index: RuleIndex, target: ResolveTarget): ResolvedAxes
 		wire: collect(wire, rule => rule.wire),
 		thinking: collect(thinking, rule => rule.thinking),
 		catalog: collect(catalog, rule => rule.catalog),
+		reasoning,
 	};
 }

package/src/compat/context-window.ts

@@ -6,8 +6,11 @@ import { resolveModelPolicy } from "./resolve";
  * Rule-owned maxima by provider/id/api. Resolve once per process rather than
  * walking the static policy cascade on every catalog rebuild. Null caches the
  * absence of a curated maximum; undefined means the key has not been resolved.
+ * Bounded: one entry per distinct model; the wire-id set is bounded.
  */
 const ruleMaximumCache = new Map<string, number | null>();
+const clampOverrideCache = new Map<string, boolean>();
+const RULE_POLICY_CACHE_MAX = 8192;
 
 /**
  * Extended-context capacity. Curated maxima correct stale lower discovery
@@ -20,6 +23,7 @@ export function resolveMaxContextWindow(model: Model): number | undefined {
 	if (curated === undefined) {
 		const maximum = resolveModelPolicy(toModelSpec(model)).catalog.maxContextWindow;
 		curated = typeof maximum === "number" && Number.isFinite(maximum) && maximum > 0 ? maximum : null;
+		if (ruleMaximumCache.size >= RULE_POLICY_CACHE_MAX) ruleMaximumCache.clear();
 		ruleMaximumCache.set(key, curated);
 	}
 
@@ -49,7 +53,13 @@ export function codexOverrideCeiling(model: Model): number | undefined {
  * it here keeps provider deployment contracts out of TypeScript.
  */
 export function clampsContextOverride(model: Model): boolean {
-	return resolveModelPolicy(toModelSpec(model)).catalog.clampContextOverride === true;
+	const key = `${model.provider} ${model.id} ${model.api}`;
+	const cached = clampOverrideCache.get(key);
+	if (cached !== undefined) return cached;
+	const clamps = resolveModelPolicy(toModelSpec(model)).catalog.clampContextOverride === true;
+	if (clampOverrideCache.size >= RULE_POLICY_CACHE_MAX) clampOverrideCache.clear();
+	clampOverrideCache.set(key, clamps);
+	return clamps;
 }
 
 /**

package/src/compat/resolve.ts

@@ -305,7 +305,11 @@ interface OpenAIDetection {
 	isOpenRouter: boolean;
 }
 
-function detectOpenAI(spec: ModelSpec<"openai-completions" | "openrouter">, facts: IdentityFacts): OpenAIDetection {
+function detectOpenAI(
+	spec: ModelSpec<"openai-completions" | "openrouter">,
+	facts: IdentityFacts,
+	reasoningCapable: boolean,
+): OpenAIDetection {
 	const provider = spec.provider;
 	const baseUrl = spec.baseUrl;
 	const hostModel = { provider, baseUrl };
@@ -315,7 +319,7 @@ function detectOpenAI(spec: ModelSpec<"openai-completions" | "openrouter">, fact
 	const isXiaomiHost = modelMatchesHost(hostModel, "xiaomi");
 	const isDirectDeepseekApi = modelMatchesHost(hostModel, "deepseekDirect");
 	const isDeepseekFamily = modelMatchesHost(hostModel, "deepseekFamily") || facts.is("deepseek");
-	const isDeepseekReasoning = isDeepseekFamily && Boolean(spec.reasoning);
+	const isDeepseekReasoning = isDeepseekFamily && reasoningCapable;
 	const isLocalOpenAICompatBackend =
 		PROXY_OPENAI_COMPAT_PROVIDERS[provider] !== true &&
 		(LOCAL_OPENAI_COMPAT_PROVIDERS[provider] === true || hasLocalLoopbackBaseUrl(baseUrl));
@@ -348,6 +352,7 @@ function detectOpenAI(spec: ModelSpec<"openai-completions" | "openrouter">, fact
 function detectOpenAICompat(
 	spec: ModelSpec<"openai-completions" | "openrouter">,
 	d: OpenAIDetection,
+	reasoningCapable: boolean,
 ): ResolvedOpenAICompat {
 	const provider = spec.provider;
 	const baseUrl = spec.baseUrl;
@@ -435,12 +440,12 @@ function detectOpenAICompat(
 			? GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
 			: facts.is("mimo") && hostMatchesUrl(baseUrl, "xiaomi")
 				? 300_000
-				: spec.reasoning &&
+				: reasoningCapable &&
 					  facts.is("kimi") &&
 					  (facts.family("k3") || facts.family("k2.7-code")) &&
 					  hostMatchesUrl(baseUrl, "moonshotNative")
 					? 300_000
-					: spec.reasoning && facts.is("deepseek") && hostMatchesUrl(baseUrl, "deepseekDirect")
+					: reasoningCapable && facts.is("deepseek") && hostMatchesUrl(baseUrl, "deepseekDirect")
 						? 300_000
 						: d.isLocalServingBackend
 							? LOCAL_OPENAI_COMPAT_STREAM_IDLE_TIMEOUT_MS
@@ -476,13 +481,14 @@ function detectOpenAICompat(
 		// provider rule without changing Copilot Responses rows.
 		supportsReasoningParams: provider !== "github-copilot",
 		supportsSamplingParams: !(facts.is("openai") && (facts.family("o-series") || facts.revGte("5"))),
-		supportsPenaltyAndStopParams: !(isGrok && Boolean(spec.reasoning)),
+		supportsPenaltyAndStopParams: !(isGrok && reasoningCapable),
 		reasoningEffortMap: {},
 		supportsUsageInStreaming: !isCerebrasHost,
 		alwaysSendMaxTokens: facts.is("kimi"),
 		disableReasoningOnForcedToolChoice:
 			!d.isClinePass && ((facts.is("kimi") && !isMoonshotKimiK3) || isAnthropicModel),
-		disableReasoningOnToolChoice: !d.isClinePass && isDeepseekFamily && Boolean(spec.reasoning) && !d.isOpenRouter,
+		disableReasoningOnToolChoice: !d.isClinePass && isDeepseekFamily && reasoningCapable && !d.isOpenRouter,
+		disableReasoningWithTools: false,
 		supportsToolChoice: d.isClinePass || !d.isDirectDeepseekReasoning,
 		supportsForcedToolChoice:
 			!d.requiresEnabledThinking && !(d.isOpenCodeHost && d.isDeepseekReasoning) && !(d.isClinePass && isQwen),
@@ -506,12 +512,12 @@ function detectOpenAICompat(
 		reasoningContentField: d.isClinePass ? "reasoning" : "reasoning_content",
 		requiresReasoningContentForToolCalls:
 			(facts.is("kimi") && !d.isOpenCodeProvider) ||
-			(isDeepseekFamily && Boolean(spec.reasoning)) ||
+			(isDeepseekFamily && reasoningCapable) ||
 			d.isXiaomiMimo ||
-			(d.isOpenRouter && Boolean(spec.reasoning)),
+			(d.isOpenRouter && reasoningCapable),
 		requiresReasoningContentForAllAssistantTurns:
-			((isDeepseekFamily && Boolean(spec.reasoning)) || d.isXiaomiMimo) && !d.isOpenRouter,
-		allowsSyntheticReasoningContentForToolCalls: (!isDeepseekFamily || !spec.reasoning) && !d.isXiaomiMimo,
+			((isDeepseekFamily && reasoningCapable) || d.isXiaomiMimo) && !d.isOpenRouter,
+		allowsSyntheticReasoningContentForToolCalls: (!isDeepseekFamily || !reasoningCapable) && !d.isXiaomiMimo,
 		replayReasoningContent: d.isLocalOpenAICompatBackend,
 		qwenPreserveThinking:
 			(thinkingFormat === "qwen" || thinkingFormat === "qwen-chat-template") && d.isLocalOpenAICompatBackend,
@@ -577,6 +583,11 @@ const DSML_HEALING_PROVIDERS: Record<string, true> = {
 	nanogpt: true,
 	"opencode-go": true,
 	openrouter: true,
+	// Transparent gateways / user-configured hosts forward the upstream model's
+	// native chat template unchanged, so a deepseek-classed model behind them
+	// still emits DSML tool-call envelopes and needs the DSML healer.
+	litellm: true,
+	nous: true,
 };
 
 /**
@@ -635,7 +646,7 @@ function fixupOpenAICompat(
 		compat.omitReasoningEffort = true;
 	}
 
-	const axisWhenThinking = spec.reasoning ? objectPayload(axes.wire.whenThinking) : undefined;
+	const axisWhenThinking = compatReasoning(spec, axes) ? objectPayload(axes.wire.whenThinking) : undefined;
 	const whenThinkingPolicy =
 		spec.compat?.whenThinking ??
 		axisWhenThinking ??
@@ -662,8 +673,9 @@ function resolveOpenAICompletionsPolicy(
 	facts: IdentityFacts,
 	axes: ResolvedAxes,
 ): ResolvedOpenAICompat {
-	const d = detectOpenAI(spec, facts);
-	const compat = detectOpenAICompat(spec, d);
+	const reasoningCapable = compatReasoning(spec, axes);
+	const d = detectOpenAI(spec, facts, reasoningCapable);
+	const compat = detectOpenAICompat(spec, d, reasoningCapable);
 	applyWireAxes(compat, axes.wire, "openai-completions");
 	applyCompatOverrides(compat, spec.compat);
 	overlayEffortMapAxis(compat, axes, spec.compat);
@@ -689,7 +701,7 @@ function resolveOpenAIResponsesPolicy(
 	const supportsPromptCacheBreakpoints =
 		isOfficialOpenAIEndpoint(provider, baseUrl) && facts.is("openai") && facts.revGte("5.6");
 	const thinkingFormat: ResolvedOpenAISharedCompat["thinkingFormat"] = isOpenRouter ? "openrouter" : "openai";
-	const reasoningCapable = Boolean(spec.reasoning);
+	const reasoningCapable = compatReasoning(spec, axes);
 	const isLocalServingBackend =
 		(PROXY_OPENAI_COMPAT_PROVIDERS[provider] !== true && LOCAL_OPENAI_COMPAT_PROVIDERS[provider] === true) ||
 		hasLocalLoopbackBaseUrl(baseUrl);
@@ -729,6 +741,7 @@ function resolveOpenAIResponsesPolicy(
 		filterReasoningHistory: isOpenRouter && isAnthropicModel,
 		disableReasoningOnForcedToolChoice: facts.is("kimi"),
 		disableReasoningOnToolChoice: isDeepseekFamily && reasoningCapable && !isOpenRouter,
+		disableReasoningWithTools: false,
 		supportsToolChoice: true,
 		supportsForcedToolChoice: provider !== "opencode-go" && provider !== "opencode-zen",
 		supportsNamedToolChoice: true,
@@ -860,7 +873,8 @@ function resolveAnthropicPolicy(
 		supportsSamplingParams: !facts.anthropicAdaptiveGenAtLeast("4.7"),
 		requiresToolResultId: false,
 		requiresThinkingEnabled,
-		replayUnsignedThinking: !signingEndpoint && (Boolean(spec.reasoning) || modelMatchesHost(spec, "deepseekFamily")),
+		replayUnsignedThinking:
+			!signingEndpoint && (compatReasoning(spec, axes) || modelMatchesHost(spec, "deepseekFamily")),
 		escapeBuiltinToolNames: false,
 		injectClaudeCodeInstruction: true,
 		stripImageInput: false,
@@ -884,7 +898,7 @@ function resolveBedrockPolicy(spec: ModelSpec<"bedrock-converse-stream">, axes:
 		promptCacheMaximumCheckpoints: 0,
 	};
 	// Reasoning capability is a mechanism gate; adaptive-lineage duration is rule-owned.
-	compat.streamIdleTimeoutMs = spec.reasoning ? BEDROCK_REASONING_STREAM_IDLE_TIMEOUT_MS : undefined;
+	compat.streamIdleTimeoutMs = compatReasoning(spec, axes) ? BEDROCK_REASONING_STREAM_IDLE_TIMEOUT_MS : undefined;
 	applyWireAxes(compat, axes.wire, "bedrock-converse-stream");
 	applyCompatOverrides(compat, spec.compat);
 	return compat;
@@ -1047,6 +1061,18 @@ function readRuleThinking(axes: ResolvedAxes): RuleThinking {
 	return out;
 }
 
+/**
+ * Compat-time reasoning capability. `axes.reasoning` also promotes targets on
+ * any exact `thinking-efforts` rule (the cascade's thinking-axis gate), but
+ * compat may only be repaired where the provider contract opted in with
+ * `thinking-upgrade-neutral`; everywhere else a spec that reports no reasoning
+ * stays the authoritative capability surface.
+ */
+function compatReasoning<TApi extends Api>(spec: ModelSpec<TApi>, axes: ResolvedAxes): boolean {
+	if (spec.reasoning) return true;
+	return axes.reasoning && readRuleThinking(axes).upgradeNeutral === true;
+}
+
 /** Identity-derived `requiresEffort` default (mandatory-reasoning lineages). */
 function impliesMandatoryReasoning(facts: IdentityFacts, modelId: string): boolean {
 	if (facts.identity.thinkingVariant) return true;

package/src/compat/rules/README.md

@@ -288,15 +288,17 @@ behavior {
         route "openai-completions" prefix="openai/" strip-prefix=#true
     }
     model-limits provider="github-copilot" { limits "gpt-5.6" context=272000 max-tokens=128000 }
+    exclude-discovery-modes "embedding" "moderation" provider="litellm"
     exclude-models provider="nanogpt" substring="embed" substring="tts"
     plan-requirement provider="openai-codex" { tier "pro" substring="-spark" }
+    retry-reset-timezone provider="zai" offset="+08:00"
     pricing-peer provider="google-antigravity" peers="google" "google-vertex" {
         alias "gemini-3-pro" peer-id="gemini-3-pro-preview"
     }
 }
 ```
 
-Matcher properties on `route` / `exclude-models` / `tier` nodes are `exact=` / `prefix=` / `substring=` / `glob=`, repeatable. `strip-prefix=#true` on a prefix route strips the matched prefix off the wire id. Values are copied verbatim from the TS constants they replaced; runtime accessors live in `src/compat/behavior.ts`.
+`exclude-discovery-modes` takes one or more exact, case-sensitive upstream mode strings plus `provider=`; discovery mappers preserve missing, malformed, and unknown modes unless the provider policy explicitly lists them. Matcher properties on `route` / `exclude-models` / `tier` nodes are `exact=` / `prefix=` / `substring=` / `glob=`, repeatable. `strip-prefix=#true` on a prefix route strips the matched prefix off the wire id. Values are copied verbatim from the TS constants they replaced; runtime accessors live in `src/compat/behavior.ts`.
 
 ## Auth grammar
 

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

@@ -84,7 +84,15 @@ class "deepseek" {
 	models "*v4.1-flash*" priority=10 {
 		strip-image-input #false
 	}
-	on "ollama-cloud" "nvidia" "deepseek" "fireworks" "nanogpt" "opencode-go" "openrouter" {
+	// V4.1 Flash is natively multimodal (vision encoder + projector, image-text
+	// pre-training) but its release id carries no `vision` token to key on.
+	// Same priority as the V4.1 Flash carve-out: the OpenCode Go provider rule
+	// declares the same value for its `deepseek-flash` lane and would otherwise
+	// tie at equal rank.
+	models "deepseek-flash" priority=10 {
+		strip-image-input #false
+	}
+	on "ollama-cloud" "nvidia" "deepseek" "fireworks" "nanogpt" "opencode-go" "openrouter" "litellm" "nous" {
 		stream-markup-healing-pattern "dsml"
 	}
 	thinking-loop-guard "deepseek"