@oh-my-pi/pi-catalog 18.8.8 → 18.9.0

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## [Unreleased]
 
+## [18.9.0] - 2026-10-11
+
+### Fixed
+
+- Fixed GLM-5.3-FlashX on Zhipu Coding Plan so bundled context, image, and reasoning-level capabilities are reported correctly.
+- Fixed OpenRouter Decisions endpoint derivation when the chat base URL has a trailing slash.
+- Fixed GPT-6.1 Sol Ultrafast usage pricing to apply OpenAI's published premium rates for the OpenAI API and Codex included usage.
+- Fixed GPT-6 Sol, GPT-6.1 Sol, and GPT-6 Luna so changing reasoning effort preserves prompt caching.
+
 ## [18.8.8] - 2026-10-10
 
 ### Added

package/dist/types/compat/reasoning-carry.d.ts

@@ -0,0 +1,21 @@
+import type { ModelIdentity } from "./types.js";
+/**
+ * The identity a deployment's reasoning is carried under, or `undefined` when
+ * the deployment's model family is not declared portable there. Class-only
+ * identities never qualify: they name a lineage, not a set of weights.
+ */
+export declare function portableReasoningIdentity(provider: string, api: string, modelId: string): ModelIdentity | undefined;
+/**
+ * Whether reasoning recorded on `source` may replay natively on `target`:
+ * both deployments declare the family portable and classify to the same
+ * class, family, and revision.
+ */
+export declare function carriesReasoning(source: {
+    provider: string;
+    api: string;
+    model: string;
+}, target: {
+    provider: string;
+    api: string;
+    id: string;
+}): boolean;

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

@@ -352,6 +352,11 @@ export interface OpenRouterModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * OpenRouter's Decisions API lives at `/api/alpha`, a sibling of the `/api/v1`
+ * chat root; derive it so a custom gateway base URL keeps both aligned.
+ */
+export declare function openrouterDecisionsBaseUrl(chatBaseUrl: string): string;
 export declare function openrouterModelManagerOptions(config?: OpenRouterModelManagerConfig): ModelManagerOptions<Api>;
 export interface ZenMuxModelManagerConfig {
     apiKey?: string;

package/dist/types/types.d.ts

@@ -340,6 +340,8 @@ export interface OpenAICompat {
     extraBody?: Record<string, unknown>;
     /** Request-session header that should mirror the normalized prompt-cache key. Default: unset. */
     promptCacheSessionHeader?: "x-grok-conv-id";
+    /** Request header carrying the conversation session id, independent of prompt caching. Default: unset. */
+    sessionHeader?: "x-litellm-session-id";
     /** Whether chat-completions payloads should include provider-specific prompt-cache markers. */
     cacheControlFormat?: "anthropic" | undefined;
     /**
@@ -423,8 +425,8 @@ export interface OpenAICompat {
     /**
      * 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).
+     * effort stays pinned for prompt caching (GPT-6 family). Default:
+     * rule-detected (`true` for GPT-6 models 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`.
@@ -621,12 +623,31 @@ export interface AnthropicCompat {
     injectClaudeCodeInstruction?: boolean;
     /** Strip image inputs before encoding (text-only serving of a multimodal id). */
     stripImageInput?: boolean;
+    /**
+     * Largest width or height a single image block may carry. The canonical
+     * Anthropic API rejects anything larger with `At least one of the image
+     * dimensions exceed max allowed size: 8000 pixels`, and the rejection
+     * poisons every retry and model fallback because the block stays in
+     * context, so the request builder downscales to this value first. Hosts
+     * with a different image contract override it. Default: 8000.
+     */
+    maxImageDimension?: number;
+    /**
+     * Largest base64 payload a single image block may carry, in bytes. The
+     * canonical Anthropic API measures the encoded string, not the decoded
+     * bytes (`image exceeds 10 MB maximum: 14012300 bytes > 10485760 bytes`),
+     * and an image inside {@link maxImageDimension} can still cross it, so
+     * payload size is clamped on its own. Hosts with a different image
+     * 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
-     * budget models). Default: false.
+     * budget models), and replay history without thinking blocks on every turn
+     * that sends `thinking: { type: "disabled" }`. Default: false.
      */
     stripThinkingHistory?: boolean;
     /**
@@ -636,10 +657,15 @@ export interface AnthropicCompat {
     effortBeta?: boolean;
     /**
      * Wire form of a disabled-thinking turn: omit the field, send
-     * `{ type: "disabled" }`, or keep adaptive thinking. Unset keeps the
-     * direct-provider behavior.
+     * `{ type: "disabled" }`, keep adaptive thinking, or send
+     * `{ type: "between_tools" }`. Unset keeps the direct-provider behavior.
+     */
+    disabledThinking?: "omit" | "disabled" | "adaptive" | "between-tools";
+    /**
+     * `output_config.effort` pinned on a `between_tools` turn. Unset leaves
+     * the turn's effort untouched.
      */
-    disabledThinking?: "omit" | "disabled" | "adaptive";
+    betweenToolsEffort?: "low" | "medium" | "high";
     /** The model is a fast-mode SKU: send `speed: "fast"` with the fast-mode beta. Default: false. */
     fastMode?: boolean;
 }
@@ -787,6 +813,7 @@ export interface ResolvedOpenAISharedCompat {
     emptyLengthFinishIsContextError: boolean;
     usesOpenAIToolCallIdLimit: boolean;
     promptCacheSessionHeader?: OpenAICompat["promptCacheSessionHeader"];
+    sessionHeader?: OpenAICompat["sessionHeader"];
     /**
      * Whether this model accepts explicit OpenAI prompt-cache breakpoints.
      * Built catalog models always materialize this false-by-default value;
@@ -821,7 +848,7 @@ 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" | "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" | "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" | "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">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -851,8 +878,8 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     /**
      * Whether a `configuration_update` input item may change `reasoning.effort`
      * mid-conversation while the request-level effort stays byte-stable for
-     * prompt caching. Rule-owned: GPT-6 Astra only; every other model rejects
-     * the item type with 400.
+     * prompt caching. Rule-owned: the GPT-6 family; earlier models do not
+     * support the item type.
      */
     supportsConfigurationUpdate: boolean;
     /**
@@ -908,15 +935,28 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
  */
 export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
-export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard" | "bedrockMessagesApi" | "effortBeta" | "disabledThinking" | "stripThinkingHistory" | "fastMode">> & {
+export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard" | "bedrockMessagesApi" | "effortBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
     /** Effort-beta override; undefined keeps the transport's legacy heuristic. */
     effortBeta?: AnthropicCompat["effortBeta"];
     /** Disabled-thinking wire form; undefined keeps the direct-provider behavior. */
     disabledThinking?: AnthropicCompat["disabledThinking"];
-    /** Strip thinking history on non-thinking-led turns; undefined behaves as false. */
+    /** Effort pinned on `between_tools` turns; undefined leaves effort untouched. */
+    betweenToolsEffort?: AnthropicCompat["betweenToolsEffort"];
+    /** Strip thinking history on non-thinking-led and disabled-thinking turns; undefined behaves as false. */
     stripThinkingHistory?: AnthropicCompat["stripThinkingHistory"];
     /** Fast-mode SKU; undefined behaves as false. */
     fastMode?: AnthropicCompat["fastMode"];
+    /**
+     * Largest width or height a single image block may carry. Undefined defers
+     * to the canonical Anthropic API limit applied by the request builder.
+     */
+    maxImageDimension?: number;
+    /**
+     * Largest base64 image payload a single block may carry, in bytes.
+     * Undefined defers to the canonical Anthropic API limit applied by the
+     * request builder.
+     */
+    maxImagePayloadBytes?: number;
     /** Thinking-loop watchdog guard family applied to streamed reasoning. */
     thinkingLoopGuard?: AnthropicCompat["thinkingLoopGuard"];
     /**

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

@@ -3,7 +3,7 @@
  * native request corpus with `packages/ai/scripts/capture-factory-droid-native.ts`
  * against the matching CLI binary and make `factory-droid-native-parity.test.ts` pass.
  */
-export declare const FACTORY_DROID_CLIENT_VERSION = "0.230.0";
+export declare const FACTORY_DROID_CLIENT_VERSION = "0.237.0";
 /** Wire protocols the Factory proxy multiplexes; `api-routes provider="factory-droid"` picks one per model. */
 export declare const FACTORY_DROID_WIRES: readonly ["openai-completions", "openai-responses", "anthropic-messages", "google-generate"];
 /** Wire protocol used by the Factory proxy for each model. */

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.8.8",
+	"version": "18.9.0",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -93,12 +93,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.8.8",
-		"@oh-my-pi/pi-utils": "18.8.8"
+		"@oh-my-pi/omptype": "18.9.0",
+		"@oh-my-pi/pi-utils": "18.9.0"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.8.8",
+		"@oh-my-pi/pi-ai": "18.9.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -140,6 +140,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"qwen-preserve-thinking": wire("qwenPreserveThinking", ["openai"]),
 	"reject-root-object-union": wire("rejectRootObjectUnion", OAI),
 	"retry-without-strict-on-grammar-error": wire("retryWithoutStrictOnGrammarError", OAI),
+	"session-header": wire("sessionHeader", OAI, "scalar", ["x-litellm-session-id"]),
 	"reasoning-content-field": wire("reasoningContentField", OAI, "scalar", [
 		"reasoning_content",
 		"reasoning",
@@ -230,12 +231,20 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"allow-anthropic-header-overrides": wire("allowAnthropicHeaderOverrides", ["anthropic"]),
 	"disable-adaptive-thinking": wire("disableAdaptiveThinking", ["anthropic"]),
 	"disable-strict-tools": wire("disableStrictTools", ["anthropic"]),
-	"disabled-thinking": wire("disabledThinking", ["anthropic", "bedrock"], "scalar", ["omit", "disabled", "adaptive"]),
+	"between-tools-effort": wire("betweenToolsEffort", ["anthropic"], "scalar", ["low", "medium", "high"]),
+	"disabled-thinking": wire("disabledThinking", ["anthropic", "bedrock"], "scalar", [
+		"omit",
+		"disabled",
+		"adaptive",
+		"between-tools",
+	]),
 	"effort-beta": wire("effortBeta", ["anthropic"]),
 	"escape-builtin-tool-names": wire("escapeBuiltinToolNames", ["anthropic"]),
 	"fast-mode": wire("fastMode", ["anthropic"]),
 	"first-party-provider": wire("firstPartyProvider", ["anthropic"]),
 	"inject-claude-code-instruction": wire("injectClaudeCodeInstruction", ["anthropic"]),
+	"max-image-dimension": wire("maxImageDimension", ["anthropic"]),
+	"max-image-payload-bytes": wire("maxImagePayloadBytes", ["anthropic"]),
 	"official-endpoint": wire("officialEndpoint", ["anthropic", "openai-responses"]),
 	"replay-unsigned-thinking": wire("replayUnsignedThinking", ["anthropic"]),
 	"requires-thinking-enabled": wire("requiresThinkingEnabled", ["anthropic"]),
@@ -408,6 +417,18 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"prompt-cache-lookback": { key: "promptCacheLookback", set: "catalog", shape: "scalar" },
 	"long-usage-limit-fallback": { key: "longUsageLimitFallback", set: "catalog", shape: "scalar" },
 	"max-context-window": { key: "maxContextWindow", set: "catalog", shape: "scalar" },
+	/**
+	 * Every host of the model returns its full plaintext trace wherever the
+	 * wire does not mark reasoning as a summary (`ThinkingContent.summary ===
+	 * false` vouches for nothing more), and its identity (class + family +
+	 * revision) pins one set of weights, so a turn's reasoning may replay
+	 * natively on another host serving the same model. Marked summaries never
+	 * carry; do not declare it where a host serves summaries unmarked.
+	 * Declared per family in `classes/*.kdl`; a host whose reasoning must not be
+	 * carried in or out sets `#false` in its `providers/*.kdl`. Read through
+	 * `compat/reasoning-carry.ts`.
+	 */
+	"portable-reasoning": { key: "portableReasoning", set: "catalog", shape: "scalar", values: [true, false] },
 	"pricing-status": {
 		key: "pricingStatus",
 		set: "catalog",

package/src/compat/reasoning-carry.ts

@@ -0,0 +1,64 @@
+/**
+ * Reasoning carry: whether reasoning a model produced on one host may replay
+ * natively on another host serving the same model.
+ *
+ * The model-side half is declared with the `portable-reasoning` catalog axis
+ * per family in `rules/classes/*.kdl`: the family's stored reasoning is its own
+ * full plaintext trace, and its identity pins one set of weights. A host whose
+ * reasoning must neither be carried in nor out overrides it to `#false` in its
+ * `rules/providers/*.kdl`. Whether the target host has a native slot for
+ * earlier reasoning is derived by the request encoders in `@oh-my-pi/pi-ai`,
+ * not declared here.
+ *
+ * Resolved through the cascade on demand (like delegation bias) because
+ * bundled rows are frozen by the generator; results are memoized per
+ * deployment.
+ */
+import { resolveCascade } from "./cascade";
+import { classifyModel } from "./taxonomy";
+import type { ModelIdentity } from "./types";
+
+/** `null` records a deployment that does not carry reasoning. */
+const portableIdentityMemo = new Map<string, ModelIdentity | null>();
+
+/**
+ * The identity a deployment's reasoning is carried under, or `undefined` when
+ * the deployment's model family is not declared portable there. Class-only
+ * identities never qualify: they name a lineage, not a set of weights.
+ */
+export function portableReasoningIdentity(provider: string, api: string, modelId: string): ModelIdentity | undefined {
+	const key = `${provider}\0${api}\0${modelId}`;
+	const memo = portableIdentityMemo.get(key);
+	if (memo !== undefined) return memo ?? undefined;
+	const identity = classifyModel(provider, modelId, { lenient: true });
+	let portable: ModelIdentity | null = null;
+	if (identity.family !== undefined || identity.revision !== undefined) {
+		const declared = resolveCascade({
+			provider,
+			api,
+			class: identity.class,
+			model: modelId,
+			reasoning: true,
+			...(identity.family !== undefined && { family: identity.family }),
+			...(identity.revision !== undefined && { revision: identity.revision }),
+		}).catalog.portableReasoning;
+		if (declared === true) portable = identity;
+	}
+	portableIdentityMemo.set(key, portable);
+	return portable ?? undefined;
+}
+
+/**
+ * Whether reasoning recorded on `source` may replay natively on `target`:
+ * both deployments declare the family portable and classify to the same
+ * class, family, and revision.
+ */
+export function carriesReasoning(
+	source: { provider: string; api: string; model: string },
+	target: { provider: string; api: string; id: string },
+): boolean {
+	const to = portableReasoningIdentity(target.provider, target.api, target.id);
+	if (!to) return false;
+	const from = portableReasoningIdentity(source.provider, source.api, source.model);
+	return from !== undefined && from.class === to.class && from.family === to.family && from.revision === to.revision;
+}

package/src/compat/resolve.ts

@@ -915,6 +915,11 @@ function resolveAnthropicPolicy(
 		escapeBuiltinToolNames: false,
 		injectClaudeCodeInstruction: true,
 		stripImageInput: false,
+		// Unset means "canonical Anthropic image contract"; the request builder
+		// owns those numbers. The keys must exist for a KDL rule or a
+		// 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,

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

@@ -98,4 +98,10 @@ class "deepseek" {
 	// providers alike. On hosts that parse it the healer sees no markup.
 	stream-markup-healing-pattern "dsml"
 	thinking-loop-guard "deepseek"
+	// 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.
+	revision ">=4" {
+		portable-reasoning #true
+	}
 }

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

@@ -65,7 +65,7 @@ class "glm" {
 	// GLM Coding Plan: the selectable 1M-context served ids; pin them so
 	// endpoint discovery or older bundled fallbacks cannot regress to 200k.
 	on "zai" "zhipu-coding-plan" {
-		models "glm-5.2" "glm-5.3" "glm-5.3-flash" {
+		models "glm-5.2" "glm-5.3" "glm-5.3-flash" "glm-5.3-flashx" {
 			limits-patch {
 				context-window 1000000
 				max-tokens 131072
@@ -84,10 +84,9 @@ class "glm" {
 			}
 		}
 	}
-	// glm-5.3-flash ships native image input that upstream metadata still
-	// reports as text-only.
+	// GLM-5.3 Flash and FlashX accept native image input on their coding routes.
 	on "zai" "zhipu-coding-plan" {
-		models "glm-5.3-flash" {
+		models "glm-5.3-flash" "glm-5.3-flashx" {
 			input-modalities "text" "image"
 		}
 	}
@@ -99,17 +98,27 @@ class "glm" {
 			thinking-efforts "minimal" "low" "medium" "high"
 		}
 	}
-	// GLM-5.3+ coding SKUs (base/air/turbo/flash, non-vision) expose a uniform
+	// GLM-5.3+ coding SKUs (base/air/turbo/flash/flashx) expose a uniform
 	// wire-exact low/high/max reasoning-effort ladder on every host, thinking
 	// can no longer be disabled, and the default effort is max.
 	// residue: taxonomy ranks cannot separate these ids from the vision (`…v`)
-	// and flashx/preview shapes that keep host dialects; the explicit priority resolves
+	// and preview shapes that keep host dialects; the explicit priority resolves
 	// the intentional overlap with census host-dialect residues such as
 	// baseten's `zai-org/GLM-5*`.
-	models "*glm-5.3" "*glm-5.3-air" "*glm-5.3-turbo" "*glm-5.3-flash" "*glm-5.3:*" \
+	models "*glm-5.3" "*glm-5.3-air" "*glm-5.3-turbo" "*glm-5.3-flash" "*glm-5.3-flashx" "*glm-5.3:*" \
 		"*glm-5.3-flash-uncensored" "*glm-5.3-free" "*glm-5.3-highspeed" "*glm-5.3-fast" "*glm-5.3-flash-lab" priority=10 {
 		thinking-default-level "max"
 		thinking-efforts "low" "high" "max"
 		thinking-requires-effort #true
 	}
+	// GLM 5+ reasoning reaches omp as its plaintext trace on every host, and
+	// family plus revision pin the weights (glm-5.2, glm-5.3-flash, ...). GLM 4
+	// classifies too coarsely to carry: 4.0 also matches glm-4.5v and glm-4.6v.
+	revision ">=5" {
+		portable-reasoning #true
+		// residue: fine-tuned weights that classify as the base GLM revision.
+		models "*-uncensored" "*-cybersecurity" {
+			portable-reasoning #false
+		}
+	}
 }

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

@@ -123,4 +123,12 @@ class "kimi" {
 			}
 		}
 	}
+	// K3 reasoning reaches omp as its plaintext trace on every host, and K3 was
+	// trained to read all of its earlier reasoning back (preserved thinking).
+	// A recall probe showed the reasoning slot of moonshot, factory-droid,
+	// openrouter and devin reaches the model, so K3 reasoning moves between
+	// those hosts natively instead of as visible `<think>` text.
+	family "k3" {
+		portable-reasoning #true
+	}
 }

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

@@ -98,4 +98,14 @@ class "minimax" {
 	on "minimax-code" "minimax-code-cn" "wafer-serverless" {
 		reasoning-content-field "reasoning_content"
 	}
+	// M3 reasoning reaches omp as its plaintext trace on every host, and
+	// MiniMax asks agents to pass all earlier reasoning back.
+	family "m3" {
+		portable-reasoning #true
+		// residue: M3.1 Flash Preview classifies as family m3 (no revision is
+		// extracted), but it is a different model.
+		models "*m3.1*" {
+			portable-reasoning #false
+		}
+	}
 }

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

@@ -49,10 +49,12 @@ class "openai" {
 		delegation-bias "restrained"
 		supports-steering #true
 	}
-	// GPT-6 Astra changes reasoning effort mid-conversation through a
+	// The GPT-6 family changes reasoning effort mid-conversation through a
 	// `configuration_update` input item so the request-level effort (and the
-	// cached prompt prefix) stays put; every other SKU 400s on the item type.
-	models "gpt-6-astra" {
+	// cached prompt prefix) stays put. OpenAI's reasoning guide ("Change
+	// reasoning mid-conversation") documents it for the family, and Codex's
+	// `/models` reports `supports_reasoning_effort_updates` on every GPT-6 SKU.
+	revision ">=6 <7" {
 		supports-configuration-update #true
 	}
 	// Codex SKUs: the advertised 400K figure includes the output budget; the

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

@@ -37,4 +37,11 @@ provider "abliteration" {
 
 	include-encrypted-reasoning #false
 	stream-idle-timeout-ms 0
+	// Abliterated weights classify as the GLM revision they derive from, but
+	// they are modified weights: their reasoning is not carried to or from GLM.
+	class "glm" {
+		revision ">=5" {
+			portable-reasoning #false
+		}
+	}
 }

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

@@ -36,6 +36,10 @@ provider "cursor" {
 				input-modalities "text" "image"
 				context-window-floor 1000000
 			}
+			// Cursor's agent protocol relays thinking opaquely, so it is not
+			// verified as K3's full trace: Cursor K3 turns are not carried to
+			// other K3 hosts (Cursor itself refuses foreign history).
+			portable-reasoning #false
 		}
 		family "k2.7-code" {
 			// Cursor advertised Kimi K2.7 Code Context window 262k.