@oh-my-pi/pi-catalog 18.5.0 → 18.6.0

package/CHANGELOG.md

@@ -2,6 +2,39 @@
 
 ## [Unreleased]
 
+## [18.6.0] - 2026-10-03
+
+### Fixed
+
+- Fixed DeepSeek V4 model IDs and the V4.1 Flash alias lacking version information in model identity and dashboards ([#14194](https://github.com/can1357/oh-my-pi/issues/14194)).
+- Fixed Antigravity models such as Claude Opus 5.5 and Sonnet 5.5 disappearing after `omp models refresh`. When the update check failed, omp reported an outdated Antigravity client version (2.8.0), so the server left the newer models out of the list. The fallback version is now 2.19.1.
+- DSML tool calls from DeepSeek models are now parsed on every host. This includes local servers (llama.cpp, LM Studio, vLLM), custom providers from `models.yml`, and gateways that weren't on the old list of supported hosts. Before, a complete `<｜DSML｜tool_calls>` envelope from these hosts showed up as plain text and the tool never ran ([#14202](https://github.com/can1357/oh-my-pi/pull/14202) by [@H4vC](https://github.com/H4vC)).
+
+## [18.5.1] - 2026-10-03
+
+### Added
+
+- Added support for listing models from Codex-compatible gateways through openaiCodexModelManagerOptions.baseUrl, with discovery caches isolated per endpoint.
+- Exposed the max reasoning-effort tier for Muse Spark 1.3 contributor models on Meta and Muse Code providers.
+
+### Changed
+
+- Added full configurable reasoning-effort ladders for ClinePass DeepSeek V4.1 Flash Free and Muse Spark 1.3 Contributor Free models.
+- Recognize bare and provider-qualified k3 and k3-256k selectors as Kimi K3, enabling K3-specific catalog policies and replace-edit fallback behavior while preserving explicit edit-mode overrides.
+- Updated Fireworks model pricing to use Fireworks-specific rates, correcting missing and inaccurate costs for supported models.
+- Updated Fireworks fast-model listings to replace retired models with glm-5.3-fast and kimi-k3-fast; GLM-5.3 Fast now exposes low, high, and max reasoning levels on Fireworks, Baseten, and Vercel AI Gateway.
+
+### Fixed
+
+- Fixed image input for OpenAI models used through custom Amazon Bedrock Converse providers when images are read through tools.
+- Fixed local Ollama thinking models continuing to reason when thinking was disabled, reducing delays such as slow session-title generation.
+- Fixed reasoning-level availability for OpenCode Go models so unsupported levels are no longer offered and reasoning can be disabled where supported.
+- Fixed Anthropic-compatible Claude Opus 5.5 sessions failing after tool use or system-prompt changes.
+- Fixed Cursor Grok 4.5 and 4.6 model selectors and overrides intermittently resolving to inconsistent model IDs.
+- Fixed Google Antigravity Gemini 3.1 Flash Image being recognized as an image-capable model for image roles and fallback chains.
+- Fixed GitHub Copilot base models reporting an incorrect long-context window when a separate -1m model is available.
+- Fixed newer OpenAI and Anthropic model families being incorrectly marked as accepting sampling parameters when accessed through compatible providers such as Amazon Bedrock, Google, Devin, or OpenRouter; explicit compatibility overrides continue to take precedence.
+
 ## [18.5.0] - 2026-10-03
 
 ### Added

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

@@ -102,6 +102,13 @@ export interface VariantCollapseTable {
  * Discovery uses this accessor to preserve injectable table seams.
  */
 export declare function reviewedCollapseTable(provider: string): VariantCollapseTable | undefined;
+/**
+ * Logical id of the reviewed `provider` family claiming `memberId` as a member
+ * or extra alias, or `undefined` when no reviewed family owns it. Cursor's rich
+ * discovery names lanes with it so they keep the id the legacy-slug collapse
+ * (and the bundled catalog) assigns.
+ */
+export declare function reviewedVariantFamilyId(provider: string, memberId: string): string | undefined;
 /**
  * The global automatic rule: derive an `X` + `X-thinking` family for every
  * pair where both ids are live in `specs` (trailing or infix token). Gates:

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

@@ -24,6 +24,12 @@ export interface OpenAICodexModelManagerConfig {
      * keeps the previous/bundled catalog instead.
      */
     resolveAccounts?: () => Promise<readonly OpenAICodexAccount[] | null>;
+    /**
+     * Codex backend base URL (e.g. a Codex-compatible gateway from `models.yml`).
+     * Defaults to the official ChatGPT backend. Also scopes the discovery cache,
+     * so a gateway roster never serves the official endpoint and vice versa.
+     */
+    baseUrl?: string;
     clientVersion?: string;
     fetch?: FetchImpl;
 }

package/dist/types/types.d.ts

@@ -651,6 +651,12 @@ export interface BedrockCompat {
      * `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS` alias, then the 300s default.
      */
     streamIdleTimeoutMs?: number;
+    /**
+     * Whether the served model accepts explicit sampling parameters
+     * (`temperature`, `topP`, …). Class rules set `false` for model lines that
+     * reject them on every host. Unassigned: accepted.
+     */
+    supportsSamplingParams?: boolean;
     /**
      * Whether the model accepts a forced `toolChoice` (`any` / `tool`). Claude
      * Opus/Sonnet 5.5 reject it outright; the request builder downgrades forced
@@ -666,6 +672,8 @@ export interface ResolvedBedrockCompat {
     supportsLongPromptCacheRetention: boolean;
     promptCacheMinimumTokens: number;
     promptCacheMaximumCheckpoints: number;
+    /** See {@link BedrockCompat.supportsSamplingParams}. */
+    supportsSamplingParams?: boolean;
     supportsForcedToolChoice: boolean;
     /**
      * Stream-watchdog idle-timeout fallback in ms for hosts with no keepalive
@@ -918,9 +926,11 @@ export interface DevinCompat {
     modelRouter?: boolean;
     /** Whether the upstream model supports native parallel tool calls. */
     supportsParallelToolCalls?: boolean;
+    /** See {@link BedrockCompat.supportsSamplingParams}. */
+    supportsSamplingParams?: boolean;
 }
 /** Fully-resolved devin-agent compat view. */
-export type ResolvedDevinCompat = Required<DevinCompat>;
+export type ResolvedDevinCompat = Required<Omit<DevinCompat, "supportsSamplingParams">> & Pick<DevinCompat, "supportsSamplingParams">;
 /**
  * Compatibility settings for the Google API family (google-generative-ai,
  * google-vertex, google-gemini-cli). Class-driven defaults come from the
@@ -955,13 +965,16 @@ export interface GoogleCompat {
     stripImageInput?: boolean;
     /** 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">> & {
+export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel" | "supportsSamplingParams">> & {
     streamFirstEventTimeoutMs?: number;
     streamIdleTimeoutMs?: number;
     thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
     antigravityUsageLabel?: string;
+    supportsSamplingParams?: boolean;
 };
 /** Sparse, user-authored compat overrides for a given API (models.json / config vocabulary). */
 export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "bedrock-converse-stream" ? BedrockCompat : TApi extends "devin-agent" ? DevinCompat : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? GoogleCompat : undefined;

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

@@ -24,7 +24,7 @@ export declare const getGeminiCliHeaders: (modelId?: string) => {
  * client the version and manifest are captured from, independent of the host
  * platform. Overrides: PI_AI_ANTIGRAVITY_VERSION / _CL / _OS / _ARCH.
  */
-export declare const DEFAULT_ANTIGRAVITY_VERSION = "2.8.0";
+export declare const DEFAULT_ANTIGRAVITY_VERSION = "2.19.1";
 /** Current Antigravity client version: env override → manifest-discovered → pinned fallback. */
 export declare function getAntigravityVersion(): string;
 /**
@@ -35,8 +35,11 @@ export declare function parseAntigravityManifestVersion(yamlText: string): strin
 /**
  * Resolves the latest Antigravity release from the official update manifest.
  * Success is cached for the process lifetime; failures are silent (the pinned
- * fallback stays valid) and clear the in-flight cache so a later call retries.
- * Skipped entirely when PI_AI_ANTIGRAVITY_VERSION is set.
+ * fallback stays valid) and suppress further lookups for
+ * `ANTIGRAVITY_VERSION_RETRY_MS`. Skipped entirely when PI_AI_ANTIGRAVITY_VERSION is set.
+ *
+ * The lookup is shared by concurrent callers, so it is bounded only by its own
+ * timeout; `signal` ends this caller's wait without cancelling the lookup.
  */
 export declare function ensureAntigravityVersion(fetcher?: FetchImpl, signal?: AbortSignal): Promise<void>;
 /** Antigravity `User-Agent` header value; rebuilt when the discovered version changes. */

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.5.0",
+	"version": "18.6.0",
 	"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": "18.5.0",
-		"@oh-my-pi/pi-utils": "18.5.0"
+		"@oh-my-pi/omptype": "18.6.0",
+		"@oh-my-pi/pi-utils": "18.6.0"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.5.0",
+		"@oh-my-pi/pi-ai": "18.6.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -293,7 +293,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"stream-idle-timeout-ms": wire("streamIdleTimeoutMs", [...OAI, "anthropic", "bedrock", "google"]),
 	"strip-image-input": wire("stripImageInput", [...OAI, "anthropic", "google"]),
 	"supports-forced-tool-choice": wire("supportsForcedToolChoice", [...OAI, "anthropic", "bedrock"]),
-	"supports-sampling-params": wire("supportsSamplingParams", [...OAI, "anthropic"]),
+	"supports-sampling-params": wire("supportsSamplingParams", [...OAI, "anthropic", "bedrock", "devin", "google"]),
 	"thinking-loop-guard": wire("thinkingLoopGuard", [...OAI, "anthropic", "google"], "scalar", [
 		"gemini",
 		"deepseek",
@@ -304,7 +304,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"thinking-default-level": { key: "defaultLevel", set: "thinking", shape: "scalar", values: EFFORTS },
 	"thinking-effort-budgets": { key: "effortBudgets", set: "thinking", shape: "object" },
 	"thinking-effort-map": { key: "effortMap", set: "thinking", shape: "object" },
-	"thinking-efforts": { key: "efforts", set: "thinking", shape: "array", values: EFFORTS },
+	"thinking-efforts": { key: "efforts", set: "thinking", shape: "array", values: EFFORTS, emptyArray: true },
 	"thinking-mode": {
 		key: "mode",
 		set: "thinking",

package/src/compat/collapse.ts

@@ -490,6 +490,17 @@ export function reviewedCollapseTable(provider: string): VariantCollapseTable |
 	return REVIEWED_COLLAPSE_TABLES[provider] ?? REVIEWED_COLLAPSE_TABLES[provider.toLowerCase()];
 }
 
+/**
+ * Logical id of the reviewed `provider` family claiming `memberId` as a member
+ * or extra alias, or `undefined` when no reviewed family owns it. Cursor's rich
+ * discovery names lanes with it so they keep the id the legacy-slug collapse
+ * (and the bundled catalog) assigns.
+ */
+export function reviewedVariantFamilyId(provider: string, memberId: string): string | undefined {
+	const table = reviewedCollapseTable(provider);
+	return table ? getAliasIndex(table).resolve(memberId.trim()) : undefined;
+}
+
 /**
  * The global automatic rule: derive an `X` + `X-thinking` family for every
  * pair where both ids are live in `specs` (trailing or infix token). Gates:

package/src/compat/resolve.ts

@@ -179,7 +179,7 @@ function overlayEffortMapAxis(
 }
 
 function effortList(value: unknown): readonly Effort[] | undefined {
-	if (!Array.isArray(value) || value.length === 0) return undefined;
+	if (!Array.isArray(value)) return undefined;
 	const out: Effort[] = [];
 	for (const entry of value) {
 		const effort = THINKING_EFFORTS.find(candidate => candidate === entry);
@@ -505,7 +505,7 @@ function detectOpenAICompat(
 		// API-conditional: this completions-only Copilot exclusion cannot be a
 		// provider rule without changing Copilot Responses rows.
 		supportsReasoningParams: provider !== "github-copilot",
-		supportsSamplingParams: !(facts.is("openai") && (facts.family("o-series") || facts.revGte("5"))),
+		supportsSamplingParams: true,
 		supportsPenaltyAndStopParams: !(isGrok && reasoningCapable),
 		reasoningEffortMap: {},
 		supportsUsageInStreaming: !isCerebrasHost,
@@ -598,26 +598,12 @@ function detectOpenAICompat(
 	};
 }
 
-const DSML_HEALING_PROVIDERS: Record<string, true> = {
-	ollama: true,
-	"ollama-cloud": true,
-	nvidia: true,
-	deepseek: true,
-	fireworks: 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,
-};
-
 /**
  * Default leaked-markup healer. Kimi/DeepSeek dedicated grammars are keyed on
  * identity class; official OpenAI heals nothing; everything else defaults to
- * the generic `thinking` healer.
+ * the generic `thinking` healer. DSML is DeepSeek's own tool-call grammar, so
+ * every host serving a DeepSeek model gets it — gateways, local backends, and
+ * custom providers alike.
  */
 function detectStreamMarkupHealing(
 	provider: string,
@@ -628,7 +614,7 @@ function detectStreamMarkupHealing(
 	// Kimi ids keep the generic healer, matching the census.
 	const isKimiK2 = facts.is("kimi") && facts.identity.family?.startsWith("k2") === true;
 	if (provider === "kimi-code" || provider === "moonshot" || isKimiK2) return "kimi";
-	if (facts.is("deepseek") && DSML_HEALING_PROVIDERS[provider] === true) return "dsml";
+	if (facts.is("deepseek")) return "dsml";
 	if (isOfficialOpenAIEndpoint(provider, baseUrl)) return undefined;
 	return "thinking";
 }
@@ -759,7 +745,7 @@ function resolveOpenAIResponsesPolicy(
 		thinkingLoopGuard: undefined,
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
-		supportsSamplingParams: !(facts.is("openai") && (facts.family("o-series") || facts.revGte("5"))),
+		supportsSamplingParams: true,
 		supportsPenaltyAndStopParams: !isXaiHost,
 		thinkingFormat,
 		reasoningDisableMode: resolveReasoningDisableMode(thinkingFormat),
@@ -906,7 +892,7 @@ function resolveAnthropicPolicy(
 		supportsThinkingBindingControls: false,
 		supportsBetweenToolsThinking: false,
 		supportsForcedToolChoice: !requiresThinkingEnabled && !facts.family("fable", "mythos"),
-		supportsSamplingParams: !facts.anthropicAdaptiveGenAtLeast("4.7"),
+		supportsSamplingParams: true,
 		requiresToolResultId: false,
 		requiresThinkingEnabled,
 		replayUnsignedThinking:
@@ -1162,6 +1148,7 @@ function resolveThinkingPolicy<TApi extends Api>(
 	if (compat !== undefined && "trustExplicitThinkingOnly" in compat && compat.trustExplicitThinkingOnly === true) {
 		return undefined;
 	}
+	if (rule.efforts?.length === 0) return undefined;
 	const config: ThinkingConfig = {
 		mode: rule.mode ?? defaultThinkingMode(spec, facts),
 		efforts: rule.efforts ?? fallbackEfforts(spec, compat),

package/src/compat/rules/README.md

@@ -209,7 +209,7 @@ The directive vocabulary is closed and lives in **`src/compat/axes.ts`** — one
 The three value shapes are:
 
 - **Scalar**: exactly one KDL boolean, integer, float, or string argument and no children. `#null` is rejected.
-- **Array**: one or more scalar arguments and no children; it resolves to a JSON array. Axes marked `emptyArray` in `axes.ts` also accept a bare directive, which assigns an explicit empty list (`region-upstreams-eu` with no arguments: the region never serves the model).
+- **Array**: one or more scalar arguments and no children; it resolves to a JSON array. Axes marked `emptyArray` in `axes.ts` also accept a bare directive, which assigns an explicit empty list (`region-upstreams-eu` with no arguments: the region never serves the model). Bare `thinking-efforts` keeps a reasoning-capable deployment without selectable effort tiers off the effort dial.
 - **Object**: no arguments and a child block, including an empty block. Child names are kebab-case: an axis-directive spelling compiles to its resolved axis key (`template-reasoning-effort` → `qwenTemplateReasoningEffort`), anything else converts mechanically (`input-threshold` → `inputThreshold`); camelCase names are a compile error. `extra-body` payloads (top-level or nested) are the exception — their child names are literal wire JSON keys copied verbatim (`enable_thinking`). Each child is either one scalar or another object; arrays are not representable inside an object payload.
 
 A rule cannot assign the same resolved axis twice in one block.

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

@@ -7,6 +7,30 @@ class "anthropic" {
 			thinking-prefix-binding #true
 		}
 	}
+	// Adaptive Claude (Opus 4.7+, Sonnet/Fable/Mythos 5+) rejects explicit
+	// sampling params (`temperature`, `top_p`, `top_k`) with a 400 on every
+	// serving host. The `<10` bound keeps whole-number revisions such as
+	// `claude-opus-45` (Opus 4.5) out of the range.
+	family "opus" {
+		revision ">=4.7 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "sonnet" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "fable" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "mythos" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
 	// On-demand compaction (compact-2026-09-04) works on the Claude API, Vertex,
 	// and Amazon Bedrock's `/anthropic` routes, not Bedrock Converse.
 	// Foundry / Claude Platform on AWS are first-party Anthropic routes whose
@@ -145,6 +169,12 @@ class "anthropic" {
 			revision ">=5" {
 				supports-per-message-effort #true
 			}
+			// Opus 5.5 is prefix-bound like Sonnet 5.5; without `drop_block` a
+			// changed system prompt or tool set fails the replay with a 400.
+			// Bounded for the same reason as the forced-tool-choice rule above.
+			revision ">=5.5 <6" {
+				supports-thinking-binding-controls #true
+			}
 		}
 	}
 	on "aimlapi" "gitlab-duo" "kilo" "nanogpt" "venice" {

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

@@ -92,8 +92,10 @@ class "deepseek" {
 	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"
-	}
+	// DSML is the model's own tool-call grammar, not a host quirk: any server
+	// running DeepSeek's chat template without (or with a broken) tool parser
+	// 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"
 }

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

@@ -107,7 +107,7 @@ class "glm" {
 	// 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:*" \
-		"*glm-5.3-flash-uncensored" "*glm-5.3-free" "*glm-5.3-highspeed" "*glm-5.3-flash-lab" priority=10 {
+		"*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

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

@@ -115,5 +115,12 @@ class "kimi" {
 			native-kimi-k3-reasoning #true
 			clamp-output-to-model-max #true
 		}
+		// Native K3 accepts forced tools on the Anthropic surface too;
+		// omitting thinking leaves its mandatory provider default intact.
+		on-api "anthropic-messages" {
+			family "k3" {
+				supports-forced-tool-choice #true
+			}
+		}
 	}
 }

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

@@ -3,13 +3,13 @@
 class "meta" {
 	// Exact ids on purpose: only an exact selector upgrades hosts whose
 	// discovery misreports a Muse Spark revision as non-reasoning.
-	// muse-spark-1.3 exposes the max effort tier; older/contributor revisions
-	// keep the 5-tier ladder.
-	models "muse-spark-1.3" {
+	// muse-spark-1.3 standard and contributor expose the max effort tier;
+	// older revisions keep the 5-tier ladder.
+	models "muse-spark-1.3" "muse-spark-1.3-contributor" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
 		thinking-mode "effort"
 	}
-	models "muse-spark-1.1" "muse-spark-1.2" "muse-spark-1.2-contributor" "muse-spark-1.3-contributor" {
+	models "muse-spark-1.1" "muse-spark-1.2" "muse-spark-1.2-contributor" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

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

@@ -1,9 +1,15 @@
 // Model-lineage compat for "openai"; regenerated from the frozen census using class taxonomy selectors.
 
 class "openai" {
-	// Replaces the o-series mandatory-reasoning fallback.
+	// Replaces the o-series mandatory-reasoning fallback. o-series and GPT-5+
+	// reject explicit sampling params (`temperature`, `top_p`) with a 400 on
+	// every serving host (#5606), whichever API fronts them.
 	family "o-series" {
 		thinking-requires-effort #true
+		supports-sampling-params #false
+	}
+	revision ">=5" {
+		supports-sampling-params #false
 	}
 	// GPT wire ladders are model generations, not host contracts: <5.2 keeps
 	// the classic four-tier scale, 5.2+ drops minimal for xhigh, and 5.6+

package/src/compat/rules/providers/amazon-bedrock.kdl

@@ -36,8 +36,6 @@ provider "amazon-bedrock" {
 	}
 	class "openai" {
 		thinking-mode "effort"
-		// OpenAI models reject Converse image blocks nested inside toolResult content.
-		requires-tool-result-image-hoisting #true
 	}
 	class "deepseek" {
 		requires-reasoning-content-for-all-assistant-turns #true
@@ -51,11 +49,6 @@ provider "amazon-bedrock" {
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"
 		thinking-mode "budget"
-		// Opaque application-inference-profile ARNs carry no vendor signal, so an
-		// OpenAI-backed profile classifies here and would otherwise nest images in
-		// toolResult content and hit the same rejection. Hoisting is accepted by
-		// every Converse model measured, so it is the safe default for unknowns.
-		requires-tool-result-image-hoisting #true
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "moonshot.kimi-k2-thinking" {

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

@@ -28,12 +28,14 @@ provider "cursor" {
 		}
 	}
 	// GetUsableModels advertises no input modalities or context windows.
-	// K3 (including the bare `k3` alias Cursor serves) is natively
-	// multimodal and 1M-context.
+	// Reviewed full-context K3 aliases are multimodal and 1M-context.
+	// K3 family identity alone does not verify other Cursor SKUs such as k3-256k.
 	class "kimi" {
 		family "k3" {
-			input-modalities "text" "image"
-			context-window-floor 1000000
+			models "*kimi-k3*" "k3" "*/k3" {
+				input-modalities "text" "image"
+				context-window-floor 1000000
+			}
 		}
 		family "k2.7-code" {
 			// Cursor advertised Kimi K2.7 Code Context window 262k.
@@ -56,10 +58,6 @@ provider "cursor" {
 			context-window-floor 272000
 		}
 	}
-	models "k3" "*/k3" {
-		input-modalities "text" "image"
-		context-window-floor 1000000
-	}
 	models "default" {
 		context-window-floor 256000
 	}

package/src/compat/rules/providers/google-antigravity.kdl

@@ -64,7 +64,11 @@ provider "google-antigravity" {
 	models "gemini-3-pro" {
 		thinking-mode "google-level"
 	}
-	models "gemini-3-pro-image" {
+	// Antigravity discovery lists image SKUs beside chat models with no output
+	// modality, so they would build as chat. Accounts advertise either SKU in
+	// `imageGenerationModelIds` (gemini-3.1-flash-image replaced
+	// gemini-3-pro-image on some accounts, #11106).
+	models "gemini-3-pro-image" "gemini-3.1-flash-image" {
 		kind "image"
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.

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

@@ -42,8 +42,8 @@ provider "meta" {
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}
-		// Meta documents the `max` effort tier for Muse Spark 1.3 (standard) only;
-		// contributor tiers and other revisions stay on the 5-tier ladder.
+		// Meta documents the `max` effort tier for Muse Spark 1.3 standard and
+		// contributor tiers; other revisions stay on the 5-tier ladder.
 		model "muse-spark-1.3" name="Muse Spark 1.3" {
 			reasoning #true
 			input "text" "image"
@@ -60,7 +60,7 @@ provider "meta" {
 			cost input=0.1 output=0.2 cache-read=0.002 cache-write=0
 			limits context=1048576 max-tokens=131072
 			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}

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

@@ -13,6 +13,9 @@ provider "ollama" priority=-1 {
 	// Replaces both Ollama empty-length finish baselines.
 	empty-length-finish-is-context-error #true
 	thinking-efforts "low" "medium" "high" "max"
+	// Ollama's OpenAI-compatible endpoints accept effort "none" to turn thinking
+	// off; the lowest ladder effort ("low") still thinks on e.g. Gemma 4.
+	reasoning-disable-mode "none-effort"
 	// Ollama continues a trailing assistant turn verbatim on /api/chat and
 	// /v1/responses alike; callers use it to commit an output prefix.
 	supports-assistant-prefill #true

package/src/compat/rules/providers/opencode-go.kdl

@@ -78,6 +78,17 @@ provider "opencode-go" {
 			thinking-mode "effort"
 		}
 	}
+	// LongCat exposes a reasoning toggle, not selectable effort tiers.
+	// Off uses the gateway's boolean reasoning switch, never a fake lowest tier.
+	models "longcat-2.5-preview-free" {
+		thinking-efforts
+		supports-reasoning-effort #false
+		reasoning-disable-mode "openrouter-enabled-false"
+	}
+	// Space Bunny publishes low through max, without the generic minimal tier.
+	models "space-bunny-free" {
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "glm-5" "glm-5.1" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"

package/src/compat/rules/providers/tool-result-images.kdl

@@ -0,0 +1,16 @@
+// Converse rejects image blocks nested inside toolResult content for OpenAI models,
+// so their tool-result images move into sibling user blocks. The rule is scoped by
+// adapter, not provider name, so a custom provider on bedrock-converse-stream gets
+// it too.
+on-api "bedrock-converse-stream" {
+	class "openai" {
+		requires-tool-result-image-hoisting #true
+	}
+	// Opaque application-inference-profile ARNs carry no vendor signal, so an
+	// OpenAI-backed profile classifies here and would otherwise nest images in
+	// toolResult content and hit the same rejection. Hoisting is accepted by
+	// every Converse model measured, so it is the safe default for unknowns.
+	class "unknown" {
+		requires-tool-result-image-hoisting #true
+	}
+}

package/src/compat/rules/taxonomy/_collapse.kdl

@@ -639,6 +639,12 @@ collapse {
 	// Retired id: the Codex image model was relabeled to what the backend runs
 	// (gpt-image-2-codex); keep saved `openai-codex/gpt-image-1` roles resolving.
 	provider-alias "openai-codex" "gpt-image-1" "gpt-image-2"
+	// Retired rich-lane ids: AvailableModels discovery named Grok 4.5/4.6 lanes
+	// after the unprefixed model name before they adopted these family ids
+	// (#14164). Family `aliases` (not `provider-alias`) so saved selectors and
+	// models.yml `modelOverrides` keyed by the old ids both migrate. Grok 4.6
+	// is spelled out instead of riding the `cursor-grok-{rev}` template: a
+	// templated `grok-{rev}` alias would also claim the live Grok 4.7+ ids.
 	variant-family "cursor" "cursor-grok-4.5" name="Grok 4.5" {
 		members "cursor-grok-4.5-low" "cursor-grok-4.5-medium" "cursor-grok-4.5-high"
 		route "low" "cursor-grok-4.5-low"
@@ -648,6 +654,7 @@ collapse {
 		efforts "low" "medium" "high"
 		requires-effort #true
 		default-member "cursor-grok-4.5-medium"
+		aliases "grok-4.5"
 	}
 	variant-family "cursor" "cursor-grok-4.5-fast" name="Grok 4.5 Fast" {
 		members "cursor-grok-4.5-low-fast" "cursor-grok-4.5-medium-fast" "cursor-grok-4.5-high-fast"
@@ -658,8 +665,33 @@ collapse {
 		efforts "low" "medium" "high"
 		requires-effort #true
 		default-member "cursor-grok-4.5-medium-fast"
+		aliases "grok-4.5-fast"
 	}
-	variant-family "cursor" "cursor-grok-{rev}" name="Grok {rev}" revision=">=4.6" {
+	variant-family "cursor" "cursor-grok-4.6" name="Grok 4.6" {
+		members "cursor-grok-4.6-low" "cursor-grok-4.6-medium" "cursor-grok-4.6-high" "cursor-grok-4.6-xhigh"
+		route "low" "cursor-grok-4.6-low"
+		route "medium" "cursor-grok-4.6-medium"
+		route "high" "cursor-grok-4.6-high"
+		route "xhigh" "cursor-grok-4.6-xhigh"
+		mode "effort"
+		efforts "low" "medium" "high" "xhigh"
+		requires-effort #true
+		default-member "cursor-grok-4.6-medium"
+		aliases "grok-4.6"
+	}
+	variant-family "cursor" "cursor-grok-4.6-fast" name="Grok 4.6 Fast" {
+		members "cursor-grok-4.6-low-fast" "cursor-grok-4.6-medium-fast" "cursor-grok-4.6-high-fast" "cursor-grok-4.6-xhigh-fast"
+		route "low" "cursor-grok-4.6-low-fast"
+		route "medium" "cursor-grok-4.6-medium-fast"
+		route "high" "cursor-grok-4.6-high-fast"
+		route "xhigh" "cursor-grok-4.6-xhigh-fast"
+		mode "effort"
+		efforts "low" "medium" "high" "xhigh"
+		requires-effort #true
+		default-member "cursor-grok-4.6-medium-fast"
+		aliases "grok-4.6-fast"
+	}
+	variant-family "cursor" "cursor-grok-{rev}" name="Grok {rev}" revision=">=4.7" {
 		members "cursor-grok-{rev}-low" "cursor-grok-{rev}-medium" "cursor-grok-{rev}-high" "cursor-grok-{rev}-xhigh"
 		route "low" "cursor-grok-{rev}-low"
 		route "medium" "cursor-grok-{rev}-medium"
@@ -670,7 +702,7 @@ collapse {
 		requires-effort #true
 		default-member "cursor-grok-{rev}-medium"
 	}
-	variant-family "cursor" "cursor-grok-{rev}-fast" name="Grok {rev} Fast" revision=">=4.6" {
+	variant-family "cursor" "cursor-grok-{rev}-fast" name="Grok {rev} Fast" revision=">=4.7" {
 		members "cursor-grok-{rev}-low-fast" "cursor-grok-{rev}-medium-fast" "cursor-grok-{rev}-high-fast" "cursor-grok-{rev}-xhigh-fast"
 		route "low" "cursor-grok-{rev}-low-fast"
 		route "medium" "cursor-grok-{rev}-medium-fast"

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

@@ -15,6 +15,10 @@ class "deepseek" {
     family "v4" glob="*deepseek-v4*"
     family "v3" glob="*deepseek-v3*"
 
+    revision prefix="deepseek-v" anywhere=#true
+
+    override id="deepseek-bare-flash-v4-1" provider="deepseek" model="deepseek-flash" class="deepseek" family="flash" revision="4.1" rationale="The first-party bare Flash alias serves DeepSeek V4.1 Flash, not V4 Flash" provenance="providers/deepseek.kdl"
+    override id="opencode-go-bare-flash-v4-1" provider="opencode-go" model="deepseek-flash" class="deepseek" family="flash" revision="4.1" rationale="OpenCode Go's bare Flash lane shares the reviewed V4.1 Flash deployment contract" provenance="providers/opencode-go.kdl"
     override id="vercel-deepseek-v3-2-thinking-product" provider="vercel-ai-gateway" model="deepseek-v3.2-thinking" logical="deepseek/deepseek-v3.2-thinking" class="deepseek" rationale="Oracle exposes this independently priced product rather than an effort route" provenance="fixtures/llm-oracle/catalog/models.normalized.json"
     override id="opencode-zen-big-pickle-family" provider="opencode-zen" model="big-pickle" class="deepseek" rationale="OpenCode Zen's big-pickle is a DeepSeek reasoning alias whose upstream 400s require exact reasoning_content replay" provenance="models.json census"
     override id="umans-deepseek-v4-flash-family" provider="umans" model="umans-deepseek-v4-flash-0731" logical="umans-deepseek-v4-flash-0731" class="deepseek" rationale="The opaque Umans product identifier is backed by the reviewed DeepSeek V4 Flash family" provenance="fixtures/llm-oracle/catalog-policy/compat-profiles.json:compat-04"