@linxiraos/pi-catalog 1.2.1 → 1.3.0

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

@@ -4,11 +4,27 @@ provider "openai-codex" {
 	default-model "gpt-5.5"
 	env "OPENAI_CODEX_OAUTH_TOKEN"
 	dynamic-models-authoritative #true
+	// Codex `/models` `display_name` is a title-cased slug (`GPT-6.1-Sol`,
+	// `GPT-5.6-Luna`) while the generator names bundled rows from the shared
+	// reference (`GPT-6.1 Sol`); without this the composer chip flips spelling
+	// once live discovery lands.
+	keep-bundled-names #true
 	kind-apis {
 		image "openai-codex-responses"
 	}
 	web-search "codex"
 
+	// Policy budgets: the Codex backend documents no request-size or image
+	// limit, and requests with several thousand images fail mid-stream at
+	// random. Inline images keep Anthropic's 24 MB; the official endpoint's
+	// image count stays well below those failures. A custom baseUrl keeps 200.
+	inline-image-byte-budget 24000000
+	image-count-budget 200
+	official-image-count-budget 2000
+	// The backend decodes zstd request bodies, as the official client sends.
+	// Its rejection body is undocumented, so any 400/415 re-checks uncompressed.
+	request-encodings "zstd"
+
 	seed api="openai-codex-responses" base-url="https://chatgpt.com/backend-api" bundle="always" {
 		// Label only: the Codex backend ignores `tool.model` and runs gpt-image-2-codex.
 		model "gpt-image-2" name="GPT Image 2" {
@@ -91,10 +107,12 @@ provider "openai-codex" {
 		}
 	}
 	// GPT-6.1 Sol keeps GPT-6 Sol's $2/$10 list price but halves cached input
-	// to $0.10 (OpenAI API docs, 2026-09-29). Like Astra, Codex reports a 272K
-	// default and an 872K maximum while OpenAI documents 922K max input (1.05M
-	// total); the curated maximum applies only with extended context on, and a
-	// higher live maximum still wins.
+	// to $0.10 (OpenAI API docs, 2026-09-29). OpenAI raised the Codex/Work
+	// default window from 272K to 500K at no extra charge (2026-10-11,
+	// https://x.com/thsottiaux/status/2109430343346581527). Codex reports an
+	// 872K maximum while OpenAI documents 922K max input (1.05M total); the
+	// curated maximum applies only with extended context on, and a higher live
+	// maximum still wins.
 	models "gpt-6.1-sol" "gpt-6.1-sol-wm" {
 		cost-patch {
 			input 2.0
@@ -103,7 +121,7 @@ provider "openai-codex" {
 			cache-write 0
 		}
 		limits-patch {
-			context-window 272000
+			context-window 500000
 		}
 		max-context-window 922000
 	}
@@ -209,3 +227,9 @@ provider "openai-codex" {
 		context-window-floor 1000000
 	}
 }
+
+// The Codex backend rejects every sampling control (`temperature`, `top_p`,
+// penalties, stop sequences), whichever model it serves (#3117).
+on-api "openai-codex-responses" {
+	supports-sampling-params #false
+}

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

@@ -3,6 +3,7 @@
 provider "openai" {
 	default-model "gpt-5.5"
 	env "OPENAI_API_KEY"
+	image-count-budget 200
 	kind-apis {
 		embedding "openai-embeddings"
 		image "openai-responses"

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

@@ -4,6 +4,10 @@ provider "opencode-go" {
 	default-model "kimi-k2.7-code"
 	env "OPENCODE_API_KEY"
 	dynamic-models-authoritative #true
+	// Replaces the provider check that disabled forced tool choice on Responses.
+	on-api "openai-responses" {
+		supports-forced-tool-choice #false
+	}
 
 	// The Flash lanes accept images despite their IDs lacking a vision suffix;
 	// verified against the gateway (unguessable pixel-rendered text read back
@@ -15,6 +19,37 @@ provider "opencode-go" {
 		input-modalities "text" "image"
 		strip-image-input #false
 	}
+	// residue: the ids-only Go roster lists this bare V4.1 Flash alias with no
+	// price, limits or reasoning flag, and neither models.dev nor the bundle has
+	// a row to hydrate it from (#15357). Mirror the gateway's own
+	// `deepseek-v4.1-flash` row; `cost-fallback` yields to any price discovery
+	// reports later.
+	models "deepseek-flash" {
+		thinking-upgrade-neutral #true
+		cost-fallback {
+			input 0.15
+			output 0.6
+			cache-read 0.003
+			cache-write 0.0
+		}
+		limits-patch {
+			context-window 1000000
+			max-tokens 384000
+		}
+	}
+	// Live Go discovery lists Argon without capabilities; its native Gemini
+	// endpoint requires a thinking level and accepts LOW, MEDIUM, or HIGH.
+	models "gemini-4-argon-contributor" {
+		thinking-upgrade-neutral #true
+		thinking-mode "google-level"
+		thinking-efforts "low" "medium" "high"
+		thinking-requires-effort #true
+		input-modalities "text" "image"
+		limits-patch {
+			context-window 1048576
+			max-tokens 65536
+		}
+	}
 	// Replaces the OpenCode reasoning-model whenThinking fallback object.
 	when-thinking {
 		requires-reasoning-content-for-tool-calls #true

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

@@ -4,6 +4,10 @@ provider "opencode-zen" {
 	default-model "claude-opus-5"
 	env "OPENCODE_API_KEY"
 	dynamic-models-authoritative #true
+	// Replaces the provider check that disabled forced tool choice on Responses.
+	on-api "openai-responses" {
+		supports-forced-tool-choice #false
+	}
 
 	// Replaces the OpenCode reasoning-model whenThinking fallback object.
 	// Applies to every completions-routed reasoning model, discovered ids

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

@@ -13,6 +13,7 @@ provider "openrouter" {
 		stt "openai-transcriptions"
 	}
 	web-search "openrouter"
+	image-count-budget 90
 
 	// TypeSafe's Jev answers only through the Decisions API (`/api/alpha/decisions`,
 	// System One wire shape); the default `/models` roster omits `text->decisions`
@@ -118,6 +119,17 @@ provider "openrouter" {
 	models "~typesafe/*" "typesafe/*" {
 		kind "judge"
 	}
+	models "~typesafe/jev-latest" "typesafe/jev-1.13" "typesafe/jev-1.14" {
+		input-modalities "text" "image"
+		judgment-image-input {
+			mime-types "image/png,image/jpeg"
+			max-bytes 16777216
+			min-pixels 78400
+			max-pixels 2458624
+			min-side 280
+			max-side 1568
+		}
+	}
 	models "cohere/rerank-v3.5" {
 		kind "rerank"
 	}
@@ -150,6 +162,13 @@ provider "openrouter" {
 			allows-synthetic-reasoning-content-for-tool-calls #false
 		}
 	}
+	// OpenRouter forwards these hosted Muse SKUs to Meta's backend, which
+	// accepts only tool_choice "auto" and 400s on "none" (see providers/meta.kdl).
+	// Other Muse models, including muse-spark-open, retain discovered tool_choice.
+	models "meta/muse-spark-1.1" "meta/muse-spark-1.2" "meta/muse-spark-1.2-contributor" \
+		"meta/muse-spark-1.3" "meta/muse-spark-1.3-contributor" "meta/muse-wonder" {
+		supports-tool-choice #false
+	}
 	class "minimax" {
 		family "m3" {
 			thinking-efforts "minimal" "low" "medium" "high"

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

@@ -0,0 +1,60 @@
+// Provider-wire compat for "sambanova". SambaNova Cloud is an OpenAI-compatible
+// gateway whose roster changes often (preview models come and go on short
+// notice), so the live `/v1/models` roster is authoritative. Its rows carry
+// pricing and limits but no capability metadata; the selectors below own
+// reasoning, thinking wire format, and image input.
+
+provider "sambanova" {
+	default-model "gpt-oss-120b"
+	env "SAMBANOVA_API_KEY"
+	discovery label="SambaNova"
+
+	dynamic-models-authoritative #true
+	skip-cross-provider-reference-fills #true
+
+	// `/v1/models` requires a key. This row keeps the provider visible in
+	// `/model` before login; the first live roster replaces it. Rate card from
+	// https://cloud.sambanova.ai/pricing.
+	seed api="openai-completions" base-url="https://api.sambanova.ai/v1" bundle="fallback" {
+		model "gpt-oss-120b" name="GPT OSS 120B" {
+			reasoning #true
+			input "text"
+			cost input=0.22 output=0.59 cache-read=0 cache-write=0
+			limits context=131072
+		}
+	}
+
+	// `/chat/completions` documents `reasoning_effort` as low/medium/high
+	// (https://docs.sambanova.ai/docs/api-reference/chat-completions/create-chat-based-completion).
+	thinking-mode "effort"
+	thinking-format "openai"
+	class "gpt-oss" {
+		thinking-efforts "low" "medium" "high"
+		thinking-upgrade-neutral #true
+	}
+	class "minimax" {
+		family "m3" {
+			thinking-efforts "low" "medium" "high"
+			thinking-upgrade-neutral #true
+		}
+	}
+	// DeepSeek V3.x and Gemma 4 think only when the request sets
+	// `chat_template_kwargs.enable_thinking`; it is off by default and no
+	// effort reaches the template, so the ladder is one on-rung beside off
+	// (https://docs.sambanova.ai/docs/en/features/text-generation).
+	class "deepseek" {
+		revision ">=3.1 <4" {
+			thinking-format "qwen-chat-template"
+			thinking-efforts "high"
+			thinking-upgrade-neutral #true
+		}
+	}
+	class "gemma" {
+		revision ">=4" {
+			thinking-format "qwen-chat-template"
+			thinking-efforts "high"
+			thinking-upgrade-neutral #true
+			input-modalities "text" "image"
+		}
+	}
+}

package/src/compat/rules/providers/tool-call-names.kdl

@@ -0,0 +1,37 @@
+// Tool names in replayed history must meet the target wire's declaration limits.
+// Anthropic tool definitions: https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools
+on-api "anthropic-messages" "openai-responses" "azure-openai-responses" "openai-codex-responses" {
+	tool-call-name-charset "ascii"
+	tool-call-name-max-length 128
+}
+
+// Bedrock ToolUseBlockStart.name: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ToolUseBlockStart.html
+on-api "bedrock-converse-stream" {
+	tool-call-name-charset "ascii"
+	tool-call-name-max-length 64
+}
+
+// Chat Completions uses the shorter tool-name bound; OpenRouter can select it
+// at request time. Factory Droid encodes caller names into this shape.
+on-api "openai-completions" "openrouter" "factory-droid-agent" {
+	tool-call-name-charset "ascii"
+	tool-call-name-max-length 64
+}
+
+// Google FunctionDeclaration.name: https://googleapis.github.io/js-genai/release_docs/interfaces/types.FunctionDeclaration.html
+on-api "google-generative-ai" {
+	tool-call-name-charset "google"
+	tool-call-name-max-length 128
+}
+
+// Vertex uses the Google FunctionDeclaration shape, independently of the public Gemini API.
+on-api "google-vertex" {
+	tool-call-name-charset "google"
+	tool-call-name-max-length 128
+}
+
+// Code Assist's Gemini CLI proxy gets the conservative 64-character wire bound.
+on-api "google-gemini-cli" {
+	tool-call-name-charset "google"
+	tool-call-name-max-length 64
+}

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

@@ -14,6 +14,17 @@ provider "typesafe" {
 			supports-tools #false
 		}
 	}
+	models "jev-latest" "jev-preview" "jev-1.13.0" "jev-1.14.0" {
+		input-modalities "text" "image"
+		judgment-image-input {
+			mime-types "image/png,image/jpeg"
+			max-bytes 16777216
+			min-pixels 78400
+			max-pixels 2458624
+			min-side 280
+			max-side 1568
+		}
+	}
 
 	models "*" {
 		kind "judge"

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

@@ -5,9 +5,16 @@ provider "umans" {
 	env "UMANS_AI_CODING_PLAN_API_KEY"
 	dynamic-models-authoritative #true
 	discovery label="Umans AI Coding Plan" allow-unauthenticated=#true
+	image-count-budget 10
 
 	// Replaces resolve.ts's Umans provider baseline for Anthropic built-in tool names.
 	escape-builtin-tool-names #true
+	// Replaces the GLM 5.2+ Anthropic budget-plus-effort default.
+	class "glm" {
+		revision ">=5.2" {
+			thinking-mode "anthropic-budget-effort"
+		}
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "umans-coder" "umans-kimi-k2.7" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"

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

@@ -47,4 +47,11 @@ provider "venice" {
 	models "qwen3-235b" {
 		thinking-format "qwen"
 	}
+	// Kimi K2.7 Code's documented recommended output budget; Venice discovery
+	// reports the context-sized `max_completion_tokens` instead.
+	class "kimi" {
+		family "k2.7-code" {
+			max-tokens-ceiling 32768
+		}
+	}
 }

package/src/compat/rules/providers/xai-oauth.kdl

@@ -9,11 +9,19 @@ provider "xai-oauth" {
 		tts "xai-tts"
 	}
 	web-search "xai"
+	// api.x.ai rejects a body over 50 MiB and counts images only against the
+	// context window; 60 images at the 500 KiB resize target stay near 41 MB.
+	image-count-budget 60
+	// Public xAI prices are API-equivalent estimates for stats; SuperGrok
+	// requests stay subscription-backed, so the estimated long-context tier
+	// must not cap the runtime context window.
+	long-context-window-uncapped #true
 
 	// Source of truth for the xai-oauth chat picker. Declaration order is headline order.
 	// Context windows from hermes-agent/agent/model_metadata.py:205-220
-	// ("Values sourced from models.dev (2026-04)"). grok-build is xAI's
-	// coding-fine-tuned chat model; 512K context per user spec (2026-05-17).
+	// ("Values sourced from models.dev (2026-04)") unless a row says otherwise.
+	// grok-build is xAI's coding-fine-tuned chat model; 512K context per user
+	// spec (2026-05-17).
 	//
 	// xai-oauth's /v1/models exposes no per-request output limit on the OAuth
 	// (Grok Build / SuperGrok) surface, so maxTokens mirrors contextWindow. The
@@ -65,25 +73,27 @@ provider "xai-oauth" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=500000 max-tokens=500000
 		}
-		// Text-only per the bundled catalog.
+		// grok-4.20-*-0309: api.x.ai rejects prompts over 1M tokens on this surface
+		// ([input_too_large] "... > 1000000 tokens", all three ids, 2026-10-12).
+		// Multi-agent is text-only per the bundled catalog.
 		model "grok-4.20-multi-agent-0309" name="Grok 4.20 (Multi-Agent)" {
 			reasoning #true
 			input "text"
 			cost input=0 output=0 cache-read=0 cache-write=0
-			limits context=2000000 max-tokens=2000000
+			limits context=1000000 max-tokens=1000000
 		}
 		model "grok-4.20-0309-reasoning" name="Grok 4.20 (Reasoning)" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
-			limits context=2000000 max-tokens=2000000
+			limits context=1000000 max-tokens=1000000
 			supports-reasoning-effort #false
 		}
 		model "grok-4.20-0309-non-reasoning" name="Grok 4.20 (Non-Reasoning)" {
 			reasoning #false
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
-			limits context=2000000 max-tokens=2000000
+			limits context=1000000 max-tokens=1000000
 		}
 		// Cursor's "Composer 2.5 Fast" exposed via SuperGrok: non-reasoning,
 		// text-only, 200K context (mirrors Cursor's composer-* catalog entries).

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

@@ -8,6 +8,9 @@ provider "xai" {
 		tts "xai-tts"
 	}
 	web-search "xai"
+	// api.x.ai rejects a body over 50 MiB and counts images only against the
+	// context window; 60 images at the 500 KiB resize target stay near 41 MB.
+	image-count-budget 60
 
 	seed api="xai-tts" base-url="https://api.x.ai/v1" bundle="always" {
 		model "grok-tts" name="Grok TTS" {

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

@@ -36,6 +36,11 @@ provider "zai" {
 	// Replaces resolve.ts's ZAI provider baseline for Anthropic tool-result IDs.
 	requires-tool-result-id #true
 	class "glm" {
+		// Replaces the GLM 5.2+ Anthropic budget-plus-effort default; the lower
+		// priority yields to the family modes below on equal-rank overlaps.
+		revision ">=5.2" priority=-1 {
+			thinking-mode "anthropic-budget-effort"
+		}
 		family "flash" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 			thinking-mode "budget"

package/src/compat/rules/runtime/behavior.kdl

@@ -173,14 +173,21 @@ behavior {
 	// serve; the dedicated tool surfaces route them instead.
 	exclude-models provider="xai-oauth" exact="grok-imagine-image-quality" exact="grok-imagine-image-2.0" \
 		prefix="grok-imagine-video" prefix="grok-stt-" prefix="grok-voice-"
+	// Antigravity still advertises the Gemini 3.5 Flash wire ids, but every
+	// request to them answers with a retirement notice instead of a reply.
+	// `gemini-3-flash` is the family's recycled alias and routes to the same ids.
+	exclude-models provider="google-antigravity" exact="gemini-3.5-flash" exact="gemini-3-flash"
 	// Meta's /v1/models lists image generation and transcription SKUs beside
-	// the Muse Spark chat models.
-	exclude-models provider="meta" prefix="muse-image-" prefix="muse-voice-"
-	exclude-models provider="muse-code" prefix="muse-image-" prefix="muse-voice-"
+	// the Muse Spark chat models, and SAM segmentation beside them.
+	exclude-models provider="meta" prefix="muse-image-" prefix="muse-voice-" prefix="sam-"
+	exclude-models provider="muse-code" prefix="muse-image-" prefix="muse-voice-" prefix="sam-"
 	// NanoGPT's roster interleaves non-text SKUs by name token.
 	exclude-models provider="nanogpt" substring="embedding" substring="image" substring="vision" \
 		substring="audio" substring="speech" substring="transcribe" substring="moderation" \
 		substring="realtime" substring="whisper" substring="tts"
+	// SambaNova's /v1/models may include embedding and audio SKUs.
+	exclude-models provider="sambanova" substring="embedding" substring="whisper" substring="audio" \
+		substring="e5-mistral" substring="clip"
 	// AIML API's roster interleaves media/embedding SKUs; bounded tokens keep
 	// chat ids from false-matching on plain substrings.
 	exclude-models provider="aimlapi" token="audio" token="embed" token="embedding" token="embeddings" \
@@ -219,9 +226,10 @@ behavior {
 	// effort ladder that lists `/responses` therefore rides Responses, so
 	// thinking-off actually disables reasoning instead of clamping to the
 	// lowest effort. Ladder ids without `/responses` (deepseek-v4-flash-fast,
-	// gemini-3.7-flash, Qwen3.8-Flash, Qwen3.8-Max-0902, hy4-preview,
-	// space-bunny-alpha), the Qwen3.8 ids (chat already disables them through
-	// `enable_thinking: false`), and non-reasoning ids stay on chat completions.
+	// gemini-3.7-flash, glm-5.3-flash since 2026-10-14, Qwen3.8-Flash,
+	// Qwen3.8-Max-0902, hy4-preview, space-bunny-alpha), the Qwen3.8 ids (chat
+	// already disables them through `enable_thinking: false`), and
+	// non-reasoning ids stay on chat completions.
 	api-routes provider="commandcode" default="openai-completions" {
 		route "anthropic-messages" prefix="claude-"
 		route "openai-responses" exact="gpt-5.3-codex" exact="gpt-5.4" exact="gpt-5.4-mini" exact="gpt-5.5" \
@@ -230,7 +238,7 @@ behavior {
 			exact="deepseek/deepseek-v4-flash" exact="deepseek/deepseek-v4-flash-vision-exp" \
 			exact="deepseek/deepseek-v4-pro" exact="deepseek/deepseek-v4.1-flash" \
 			exact="deepseek/deepseek-v4.1-flash-fast" exact="moonshotai/Kimi-K3" \
-			exact="z-ai/glm-5.3-flash" exact="z-ai/glm-5.3-flashx" exact="zai-org/GLM-5.2" \
+			exact="z-ai/glm-5.3-flashx" exact="zai-org/GLM-5.2" \
 			exact="zai-org/GLM-5.3" exact="google/gemini-3.1-flash-lite" exact="google/gemini-3.5-flash" \
 			exact="google/gemini-3.5-flash-lite" exact="google/gemini-3.6-flash" \
 			exact="google/gemini-3.8-flash" exact="MiniMaxAI/MiniMax-M3" exact="stepfun/Step-5-Preview" \
@@ -255,6 +263,8 @@ behavior {
 	// at /responses (#10610, #12030). The Muse prefix covers gateway-first
 	// revisions; exact ids stay as cache-migration identity.
 	api-routes provider="opencode-zen" {
+		// Paid Argon is served only on Zen's native Google endpoint (#15977).
+		route "google-generative-ai" exact="gemini-4-argon"
 		route "anthropic-messages" exact="union-alpha"
 		route "openai-responses" prefix="muse-spark-" exact="muse-spark-1.3-contributor-free" exact="gpt-6-astra"
 		route "openai-completions" exact="minimax-m3" exact="minimax-m3-free"
@@ -263,6 +273,7 @@ behavior {
 	// responses pins are gateway-verified (deepseek-v4-flash 2026-08-08;
 	// muse-spark per opencode.ai/docs/go/#endpoints, #8957, #10610).
 	api-routes provider="opencode-go" {
+		route "google-generative-ai" exact="gemini-4-argon-contributor"
 		route "anthropic-messages" exact="union-alpha"
 		route "openai-responses" prefix="muse-spark-" exact="deepseek-v4-flash" exact="muse-spark-1.2" \
 			exact="muse-spark-1.2-contributor" exact="muse-spark-1.3" exact="muse-spark-1.3-contributor"
@@ -345,7 +356,6 @@ behavior {
 	// Google); price rows at their first-party equivalents, Vertex before
 	// Anthropic for Claude ids. Aliases map ids whose peer ships differently.
 	pricing-peer provider="google-antigravity" peers="google" "google-vertex" "anthropic" {
-		alias "gemini-3-flash" peer-id="gemini-3-flash-preview"
 		alias "gemini-3-pro" peer-id="gemini-3-pro-preview"
 		alias "gemini-3.1-pro" peer-id="gemini-3.1-pro-preview"
 		alias "claude-opus-4-5" peer-id="claude-opus-4-5@20251101"
@@ -372,4 +382,114 @@ behavior {
 	}
 	hosted-default provider="zai-search" model="glm-4.7"
 	hosted-default provider="synthetic-search" model="auto"
+	// Built-in default model selector chains per agent role, most preferred
+	// first; the coding agent reads them through `roleDefaultModels()`.
+	role-defaults "smol" {
+		model "claude-haiku-5-5"
+		model "claude-haiku-5.5"
+		model "haiku-5-5"
+		model "haiku-5.5"
+		model "google-antigravity/gemini-3.8-flash"
+		model "gemini-3.8-flash"
+		model "gemini-3-8-flash"
+		model "gemini-3.7-flash"
+		model "gemini-3-7-flash"
+		model "gemini-3.6-flash"
+		model "gemini-3-6-flash"
+		model "gemini-3.5-flash"
+		model "gemini-3-5-flash"
+		model "openai-codex/gpt-5.3-codex-spark"
+		model "gpt-5.3-codex-spark"
+		model "5.3-spark"
+		model "spark"
+		model "zai/glm-5.3-flash"
+		model "glm-5.3-flash"
+		model "gemini-3.1-flash-lite"
+		model "gemini-3-1-flash-lite"
+		model "flash-lite"
+		model "cerebras/gpt-oss-120b"
+		model "gpt-oss-120b"
+		model "oss-120b"
+		model "cerebras/zai-glm-4.7"
+		model "cerebras/zai-glm-4.6"
+		model "cerebras/zai-glm"
+		model "claude-haiku-4-5"
+		model "claude-haiku-4.5"
+		model "haiku-4-5"
+		model "haiku-4.5"
+		model "haiku"
+		model "openai-codex/gpt-5.6-luna"
+		model "gpt-5.6-luna"
+		model "openai-codex/gpt-5.4-mini"
+		model "gpt-5.4-mini"
+		model "-mini"
+	}
+	role-defaults "slow" {
+		model "openai-codex/gpt-5.6-sol"
+		model "gpt-5.6-sol"
+		model "anthropic/claude-fable-5-1"
+		model "claude-fable-5-1"
+		model "claude-fable-5.1"
+		model "fable-5.1"
+		model "fable-5-1"
+		model "anthropic/claude-fable-5"
+		model "claude-fable-5"
+		model "fable-5"
+		model "kimi-code/k3"
+		model "kimi-k3"
+		model "moonshotai/kimi-k3"
+		model "k3"
+		model "zai/glm-5.3"
+		model "glm-5.3"
+		model "anthropic/claude-opus-5-5"
+		model "claude-opus-5-5"
+		model "opus-5-5"
+		model "anthropic/claude-opus-5"
+		model "claude-opus-5"
+		model "opus-5"
+		model "openai-codex/gpt-5.5"
+		model "gpt-5.5"
+		model "claude-opus-4-8"
+		model "claude-opus-4.8"
+		model "opus-4-8"
+		model "opus-4.8"
+	}
+	role-defaults "image" {
+		model "openai/gpt-image-2"
+		model "openai-codex/gpt-image-2"
+		model "google-antigravity/gemini-3-pro-image"
+		model "xai/grok-imagine-image"
+		model "xai-oauth/grok-imagine-image"
+		model "openrouter/google/gemini-3-pro-image"
+		model "google/gemini-3-pro-image"
+		model "deepinfra/black-forest-labs/FLUX-2-pro"
+	}
+	role-defaults "web" {
+		model "web/parallel"
+		model "web/hosted"
+		model "web/exa"
+		model "web/firecrawl"
+		model "web/searxng"
+		model "web/startpage"
+		model "web/duckduckgo"
+		model "web/ecosia"
+		model "web/google"
+		model "web/mojeek"
+		model "web/public"
+	}
+	role-defaults "speech" {
+		model "local/kokoro"
+		model "xai/grok-tts"
+		model "xai-oauth/grok-tts"
+	}
+	role-defaults "dictation" {
+		model "local/parakeet-tdt-0.6b-v3"
+	}
+	role-defaults "judge" {
+		model "typesafe/jev-latest"
+		model "openrouter/~typesafe/jev-latest"
+		model "@tiny"
+		model "@smol"
+		model "@default"
+	}
 }

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

@@ -46,6 +46,17 @@ collapse {
 		efforts "minimal" "low" "medium" "high"
 		requires-effort #true
 	}
+	// Gemini 4 Argon ships as `-low`/`-medium`/`-high` thinking-level SKUs,
+	// matching the three levels Google documents for the model (no minimal).
+	variant-family "google-antigravity" "gemini-4-argon" name="Gemini 4 Argon" {
+		members "gemini-4-argon-low" "gemini-4-argon-medium" "gemini-4-argon-high"
+		route "low" "gemini-4-argon-low"
+		route "medium" "gemini-4-argon-medium"
+		route "high" "gemini-4-argon-high"
+		mode "google-level"
+		efforts "low" "medium" "high"
+		requires-effort #true
+	}
 	variant-family "google-antigravity" "gemini-3.5-flash" name="Gemini 3.5 Flash" {
 		members "gemini-3.5-flash-extra-low" "gemini-3.5-flash-low" "gemini-3-flash-agent"
 		route "off" "gemini-3.5-flash-extra-low"

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

@@ -1,3 +1,4 @@
 class "gemma" {
     bounded "gemma"
+    revision prefix="gemma-"
 }

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

@@ -1,10 +1,17 @@
 class "meta" {
 	bounded "muse-spark"
+	bounded "muse-wonder"
 	bounded "llama"
+	// SambaNova serves the bare `Meta-Llama-*` spelling with no namespace
+	// segment, which `bounded "llama"` never sees (prefix-anchored) and
+	// `namespace "meta-llama"` cannot match.
+	bounded "meta-llama"
 	namespace "meta-llama"
 
 	family "llama" glob="*llama*"
 	family "muse-spark" glob="*muse-spark*"
+	family "muse-wonder" glob="*muse-wonder*"
 
 	revision prefix="muse-spark-"
+	revision prefix="muse-wonder-"
 }

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

@@ -16,6 +16,8 @@ class "openai" {
 
 	family "gpt" glob="gpt-*"
 	family "chatgpt" glob="chatgpt-*"
+	family "gpt-image" glob="gpt-image-*"
+	family "gpt-image" glob="chatgpt-image-*"
 	family "o-series" glob="o1"
 	family "o-series" glob="o1-*"
 	family "o-series" glob="o1.*"