@oh-my-pi/pi-catalog 18.3.0 → 18.3.2

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [18.3.1] - 2026-09-25
+
+### Added
+
+- Added compatibility flags for models that stop output at the context-window limit and models that support steering, including defaults for Claude 4.5+ and GPT-6+ models.
+
+### Fixed
+
+- Fixed forced-tool requests for Claude Opus 5.5 so tool selection falls back gracefully when necessary.
+- Added the provider-advertised `xhigh` reasoning level for Yolo-Auto Qwen3.8 models.
+- Fixed Devin Fusion pairing requests failing with `no API providers are available` by routing them through the lead model with the lead's limits and pricing; pairings without an available lead are no longer listed ([#13000](https://github.com/can1357/oh-my-pi/pull/13000) by [@DarkPhilosophy](https://github.com/DarkPhilosophy)).
+
 ## [18.3.0] - 2026-09-24
 
 ### Added

package/THIRD-PARTY-NOTICES.txt

@@ -78,8 +78,8 @@ below.
 uutils coreutils (https://github.com/uutils/coreutils), MIT
 --------------------------------------------------------------------------------
 Covers: base32, base64, basename, cat, cksum (shared checksum machinery),
-b2sum, md5sum, sha1sum, sha224sum, sha256sum, sha384sum, sha512sum, comm, cut,
-date, dirname, head, hostname, ln, ls, mkdir, mktemp, mv, nproc, paste,
+b2sum, md5sum, sha1sum, sha224sum, sha256sum, sha384sum, sha512sum, comm, cp,
+cut, date, dirname, head, hostname, ln, ls, mkdir, mktemp, mv, nproc, paste,
 printenv, readlink, realpath, rm, seq, sort, stat, tac, tail, tee, touch, tr,
 truncate, uname, uniq, wc, whoami, yes.
 

package/dist/types/compat/fallback-credit.d.ts

@@ -0,0 +1,7 @@
+import type { Model } from "../types.js";
+/**
+ * Model ids on which a refusal's Anthropic fallback-credit token may be
+ * redeemed, authored on the `fallback-credit-targets` catalog axis. Empty when
+ * no rule assigns targets (the model cannot redeem credit on any fallback).
+ */
+export declare function fallbackCreditTargets(model: Model): readonly string[];

package/dist/types/compat/output-limits.d.ts

@@ -1,3 +1,8 @@
 import type { Model } from "../types.js";
+/**
+ * Whether the host truncates generation at the context window rather than
+ * rejecting prompt + output cap overflow; the output fitter skips these models.
+ */
+export declare function stopsOutputAtContextWindow(model: Model): boolean;
 /** Whether discovery and transport policy allow preserving a caller's output cap. */
 export declare function supportsOutputTokenLimit(model: Model): boolean;

package/dist/types/compat/server-side-fallback.d.ts

@@ -0,0 +1,7 @@
+import type { Model } from "../types.js";
+/**
+ * Ordered model ids for Anthropic's server-side refusal `fallbacks` chain,
+ * authored on the `server-side-fallback-models` catalog axis. Empty when no
+ * rule assigns a chain (the model or host does not support the feature).
+ */
+export declare function serverSideFallbackModels(model: Model): readonly string[];

package/dist/types/types.d.ts

@@ -401,6 +401,13 @@ export interface OpenAICompat {
      * sent as the top-level `reasoning.effort`.
      */
     supportsConfigurationUpdate?: boolean;
+    /**
+     * Whether the Responses WebSocket accepts `response.steer`, which queues user
+     * input into the in-flight response (GPT-6 family). Default: rule-detected.
+     * Set `false` for proxies that reject the event; steering then waits for the
+     * next request boundary.
+     */
+    supportsSteering?: 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. */
@@ -727,7 +734,7 @@ 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" | "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">> & {
+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" | "supportsSteering" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -761,6 +768,11 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
      * the item type with 400.
      */
     supportsConfigurationUpdate: boolean;
+    /**
+     * Whether the WebSocket transport may send `response.steer` to deliver user
+     * input into the in-flight response. Rule-owned: GPT-6 family.
+     */
+    supportsSteering: boolean;
     /** Inject the `# Juice: 0 !important` developer item when reasoning is forced off (gpt-5.6+). */
     requiresReasoningOffJuiceInstruction: boolean;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.3.0",
+	"version": "18.3.2",
 	"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": "18.3.0",
-		"@oh-my-pi/pi-utils": "18.3.0"
+		"@oh-my-pi/omptype": "18.3.2",
+		"@oh-my-pi/pi-utils": "18.3.2"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.3.0",
+		"@oh-my-pi/pi-ai": "18.3.2",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -139,6 +139,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"requires-reasoning-off-juice-instruction": wire("requiresReasoningOffJuiceInstruction", ["openai-responses"]),
 	"supports-all-turns-reasoning-context": wire("supportsAllTurnsReasoningContext", ["openai-responses"]),
 	"supports-configuration-update": wire("supportsConfigurationUpdate", ["openai-responses"]),
+	"supports-steering": wire("supportsSteering", ["openai-responses"]),
 	"strip-deepseek-special-tokens": wire("stripDeepseekSpecialTokens", OAI),
 	"stream-markup-healing-pattern": wire("streamMarkupHealingPattern", OAI, "scalar", [
 		"kimi",
@@ -303,6 +304,18 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 		shape: "scalar",
 		values: [true, false],
 	},
+	/**
+	 * The host accepts a prompt plus `max_tokens` beyond the context window and
+	 * ends generation at the window (Anthropic `model_context_window_exceeded`)
+	 * instead of rejecting the request, so callers must not lower the output cap
+	 * to fit the window.
+	 */
+	"stops-output-at-context-window": {
+		key: "stopsOutputAtContextWindow",
+		set: "catalog",
+		shape: "scalar",
+		values: [true, false],
+	},
 	"clamp-context-override": { key: "clampContextOverride", set: "catalog", shape: "scalar" },
 	"context-promotion-target": { key: "contextPromotionTarget", set: "catalog", shape: "scalar" },
 	"context-window-floor": { key: "contextWindowFloor", set: "catalog", shape: "scalar" },
@@ -335,6 +348,18 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 		shape: "scalar",
 	},
 	"supports-assistant-prefill": { key: "supportsAssistantPrefill", set: "catalog", shape: "scalar" },
+	/**
+	 * Ordered Anthropic model ids forwarded as the server-side `fallbacks`
+	 * chain when the user opts in. Each must appear in the requested model's
+	 * `allowed_fallback_models` (GET /v1/models/{id}); anything else is a 400.
+	 */
+	"server-side-fallback-models": { key: "serverSideFallbackModels", set: "catalog", shape: "array" },
+	/**
+	 * Anthropic model ids a refusal's `fallback_credit_token` may be redeemed
+	 * on (the refused model's permitted fallback targets). Unordered; a retry
+	 * on any other model cannot redeem the credit.
+	 */
+	"fallback-credit-targets": { key: "fallbackCreditTargets", set: "catalog", shape: "array" },
 	priority: { key: "priority", set: "catalog", shape: "scalar" },
 	"service-tier-cost": { key: "serviceTierCost", set: "catalog", shape: "object" },
 	"time-based-cost": { key: "timeBased", set: "catalog", shape: "object" },

package/src/compat/fallback-credit.ts

@@ -0,0 +1,13 @@
+import type { Model } from "../types";
+import { resolveCatalogPolicy } from "./catalog-policy";
+
+/**
+ * Model ids on which a refusal's Anthropic fallback-credit token may be
+ * redeemed, authored on the `fallback-credit-targets` catalog axis. Empty when
+ * no rule assigns targets (the model cannot redeem credit on any fallback).
+ */
+export function fallbackCreditTargets(model: Model): readonly string[] {
+	const value = resolveCatalogPolicy(model).fallbackCreditTargets;
+	if (!Array.isArray(value)) return [];
+	return value.filter((entry): entry is string => typeof entry === "string");
+}

package/src/compat/output-limits.ts

@@ -1,6 +1,14 @@
 import type { Model } from "../types";
 import { resolveCatalogPolicy } from "./catalog-policy";
 
+/**
+ * Whether the host truncates generation at the context window rather than
+ * rejecting prompt + output cap overflow; the output fitter skips these models.
+ */
+export function stopsOutputAtContextWindow(model: Model): boolean {
+	return resolveCatalogPolicy(model).stopsOutputAtContextWindow === true;
+}
+
 /** Whether discovery and transport policy allow preserving a caller's output cap. */
 export function supportsOutputTokenLimit(model: Model): boolean {
 	const policy = resolveCatalogPolicy(model);

package/src/compat/resolve.ts

@@ -726,6 +726,7 @@ function resolveOpenAIResponsesPolicy(
 		supportsReasoningSummary: !isXaiHost,
 		supportsAllTurnsReasoningContext: false,
 		supportsConfigurationUpdate: false,
+		supportsSteering: false,
 		requiresReasoningOffJuiceInstruction: false,
 		stripImageInput: false,
 		thinkingLoopGuard: undefined,
@@ -832,6 +833,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
 		supportsAllTurnsReasoningContext: compat.supportsAllTurnsReasoningContext,
 		supportsConfigurationUpdate: compat.supportsConfigurationUpdate,
+		supportsSteering: compat.supportsSteering,
 		officialEndpoint: compat.officialEndpoint,
 		harmonyLeakMitigation: compat.harmonyLeakMitigation,
 		cacheControlFormat: compat.cacheControlFormat,

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

@@ -36,6 +36,31 @@ class "anthropic" {
 			supports-server-compaction #true
 		}
 	}
+	// Sonnet 4.5 and newer end generation with `model_context_window_exceeded`
+	// when prompt + max_tokens overruns the window, instead of a 400, so the
+	// full output cap can be requested without sizing the prompt. Verified for
+	// the Claude API only:
+	// https://platform.claude.com/docs/en/build-with-claude/handling-stop-reasons#model-context-window-exceeded
+	on "anthropic" {
+		revision ">=4.5" {
+			stops-output-at-context-window #true
+		}
+	}
+	// Opus 5.5 rejects forced tool selection the way Fable/Mythos do — a
+	// `tool_choice` of `tool` or `any` fails the request with 400
+	// `invalid_request_error`: `tool_choice: type "tool" and "any" are not
+	// supported for this model`. The transport downgrades the selector to
+	// `auto` so the tool stays offered instead of the turn dying.
+	//
+	// Upper-bounded on purpose: separator-collapsed ids parse their revision
+	// as a whole number (`claude-opus-45` is Opus 4.5, revision 45), so an
+	// open `>=5.5` would silently catch them, and a later Opus generation has
+	// to be re-verified against the API before it inherits the restriction.
+	family "opus" {
+		revision ">=5.5 <6" {
+			supports-forced-tool-choice #false
+		}
+	}
 	on "anthropic" "cloudflare-ai-gateway" "google-vertex" {
 		family "fable" {
 			revision ">=5" {

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

@@ -37,8 +37,11 @@ class "openai" {
 	revision "=5.6" {
 		delegation-bias "gated"
 	}
+	// GPT-6 accepts `response.steer` over the Responses WebSocket: user input
+	// joins the in-flight response instead of waiting for the next request.
 	revision ">=6" {
 		delegation-bias "restrained"
+		supports-steering #true
 	}
 	// GPT-6 Astra changes reasoning effort mid-conversation through a
 	// `configuration_update` input item so the request-level effort (and the

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

@@ -58,6 +58,21 @@ provider "anthropic" {
 		revision ">=3.7 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		}
+		// Server-side refusal fallback chain (opt-in via
+		// `providers.anthropic.serverSideFallback`). Targets must be in the
+		// model's `allowed_fallback_models` on GET /v1/models/{id}; Fable 5 and
+		// 5.1 publish ["claude-opus-4-8", "claude-opus-5"] (2026-09-24).
+		// `claude-opus-5-5` is rejected with a 400 (#13059).
+		family "fable" {
+			server-side-fallback-models "claude-opus-5"
+			// Fallback-credit redemption targets: the refused model's permitted
+			// fallback targets. Fable 5 and 5.1 → Opus 4.8 and Opus 5
+			// (https://platform.claude.com/docs/en/build-with-claude/fallback-credit).
+			fallback-credit-targets "claude-opus-4-8" "claude-opus-5"
+		}
+		family "mythos" {
+			server-side-fallback-models "claude-opus-5"
+		}
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "*20240229" "claude-3-7-sonnet-20250219" {

package/src/compat/rules/providers/yolo-auto.kdl

@@ -37,7 +37,7 @@ provider "yolo-auto" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=262144 max-tokens=131072
 			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 			supports-developer-role #false
 			supports-store #false
 			supports-reasoning-effort #true
@@ -50,7 +50,7 @@ provider "yolo-auto" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=262144 max-tokens=131072
 			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 			supports-developer-role #false
 			supports-store #false
 			supports-reasoning-effort #true
@@ -61,6 +61,12 @@ provider "yolo-auto" {
 
 	supports-developer-role #false
 	supports-store #false
+	// Live /v1/models advertises `thinking: minimal..xhigh` for the Qwen3.8
+	// rows; the cascade rule carries it into the regenerated bundle, whose
+	// thinking re-bake ignores seed-level ladders.
+	models "qwen3.8-flash" "yolo" "qwen3.8-27b" {
+		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
+	}
 	models "deepseek-flash-v4" {
 		disable-reasoning-on-tool-choice #true
 		max-tokens-field "max_completion_tokens"