@oh-my-pi/pi-catalog 18.10.2 → 18.10.3

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

@@ -14,10 +14,22 @@
 // consecutive tool_use or tool_result blocks is one position. Warm-cache
 // tool-result pruning keeps its rewrites inside this window.
 // https://platform.claude.com/docs/en/build-with-claude/prompt-caching
+// A replayed signed thinking block bills the full reasoning its signature
+// carries (the server rebuilds it from the signature), at the size its turn
+// billed as output, not the signature's bytes. Thinking a request drops bills
+// nothing (earlier turns on lines that strip it: Opus before 4.5, Sonnet
+// before 4.6, Haiku before 5.5; turns before a compaction or prune on
+// prefix-bound lines; another model's blocks the reader cannot read), so
+// those turns count high.
+// https://platform.claude.com/docs/en/build-with-claude/context-windows#the-context-window-with-thinking
+// https://platform.claude.com/docs/en/build-with-claude/preserved-thinking#switching-models
 class "anthropic" {
 	prompt-cache-lookback 20
+	thinking-replay-tokens "output"
+	// verified: billing matched this formula on Opus 4.7+ and Fable 5.1.
 	image-tokenization {
 		regime "anthropic-patch"
+		verified #true
 		max-edge 2576
 		max-tokens 4784
 	}
@@ -37,21 +49,39 @@ class "anthropic" {
 			}
 		}
 	}
+	// Standard tier. verified: billing matched it on Sonnet 4.6 only; older
+	// Sonnet lines keep the floor until measured.
 	family "sonnet" {
-		revision "<4.7" {
+		revision "<4.6" {
 			image-tokenization {
 				regime "anthropic-patch"
 				max-edge 1568
 				max-tokens 1568
 			}
 		}
-		revision ">=10 <47" {
+		revision ">=4.6 <4.7" {
+			image-tokenization {
+				regime "anthropic-patch"
+				verified #true
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+		revision ">=10 <46" {
 			image-tokenization {
 				regime "anthropic-patch"
 				max-edge 1568
 				max-tokens 1568
 			}
 		}
+		revision ">=46 <47" {
+			image-tokenization {
+				regime "anthropic-patch"
+				verified #true
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
 	}
 	family "haiku" {
 		revision "<4.7" {
@@ -87,6 +117,14 @@ class "anthropic" {
 			supports-sampling-params #false
 		}
 	}
+	// Opus 5.5 made every first edit exactly right from `read` results it saw
+	// only as snapcompact frames (24/24); Haiku 5.5 lost leading whitespace
+	// (22/54), so other lines keep `read` as text until measured.
+	family "opus" {
+		revision ">=5.5 <10" {
+			snapcompact-image-reads #true
+		}
+	}
 	family "sonnet" {
 		revision ">=5 <10" {
 			supports-sampling-params #false

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

@@ -9,6 +9,11 @@ class "meta" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
 		thinking-mode "effort"
 	}
+	// Meta documents `max` for Muse Wonder (dev.meta.ai/docs/reasoning).
+	models "muse-wonder-1.0" {
+		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" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"

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

@@ -95,11 +95,11 @@ class "openai" {
 // Older lines carry no lineage rule and bill at their wire API's fallback
 // (providers/image-tokenization.kdl).
 class "openai" {
-	// gpt-5.6 sizing, shared by later lines. `<10` keeps separator-collapsed
-	// revisions (`openai-gpt-52-codex` parses as 52) off the 5.6+ sizing.
-	revision ">=5.6 <10" {
+	// gpt-5.6 sizing. verified: billing matched this formula on gpt-5.6 through 6.1.
+	revision ">=5.6 <6.2" {
 		image-tokenization {
 			regime "openai-patch"
+			verified #true
 			multiplier 1.2
 			auto "original"
 			low {
@@ -114,9 +114,30 @@ class "openai" {
 			}
 		}
 	}
+	// Later lines size like gpt-5.6, unverified until measured. `<10` keeps
+	// separator-collapsed revisions (`openai-gpt-52-codex` parses as 52) off it.
+	revision ">=6.2 <10" {
+		image-tokenization {
+			regime "openai-patch"
+			multiplier 1.2
+			auto "original"
+			low {
+				max-edge 512
+			}
+			high {
+				max-edge 2048
+				patch-budget 2500
+			}
+			original {
+				max-edge 65535
+			}
+		}
+	}
+	// verified: billing matched this formula on gpt-5.5.
 	revision "=5.5" {
 		image-tokenization {
 			regime "openai-patch"
+			verified #true
 			multiplier 1.2
 			auto "original"
 			low {
@@ -173,9 +194,11 @@ class "openai" {
 		}
 	}
 	// residue: astra's limits differ from the rest of its revision.
+	// verified: billing matched this formula on gpt-6-astra.
 	models "*gpt-6-astra*" {
 		image-tokenization {
 			regime "openai-patch"
+			verified #true
 			multiplier 1.2
 			auto "original"
 			low {

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

@@ -41,6 +41,32 @@ class "xai" {
 			}
 		}
 	}
+	// Grok bills 32px patches plus ~30–40 tokens per image, resizing past
+	// 2,401 patches at every detail level; the 1.017 multiplier carries the
+	// per-image tokens at full size (small images bill slightly more).
+	// verified: billing matched this rule on Grok 4.6, 4.7 and 4.20.
+	family "grok" {
+		revision ">=4.6 <=4.20" {
+			image-tokenization {
+				regime "openai-patch"
+				verified #true
+				multiplier 1.017
+				auto "original"
+				low {
+					max-edge 65535
+					patch-budget 2401
+				}
+				high {
+					max-edge 65535
+					patch-budget 2401
+				}
+				original {
+					max-edge 65535
+					patch-budget 2401
+				}
+			}
+		}
+	}
 	thinking-loop-guard "xai"
 }
 

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

@@ -2,6 +2,7 @@
 
 provider "amazon-bedrock" {
 	default-model "us.anthropic.claude-opus-5-5"
+	image-count-budget 90
 
 	// The Converse assembler positions content blocks by wire index
 	// (`amazon-bedrock.ts`), so a block can land above text already rendered.

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

@@ -85,7 +85,20 @@ provider "anthropic" {
 	on-api "anthropic-messages" {
 		inline-image-byte-budget 24000000
 	}
+	// The official endpoint carries at most 600 images and documents per
+	// request, 100 on models with a 200k-token window: every Claude line before
+	// 4.6, including separator-collapsed revisions (`claude-opus-45` parses as
+	// 45). A custom baseUrl keeps the conservative 90.
+	// https://platform.claude.com/docs/en/build-with-claude/vision
+	image-count-budget 90
+	official-image-count-budget 600
 	class "anthropic" {
+		revision "<4.6" {
+			official-image-count-budget 100
+		}
+		revision ">=10 <46" {
+			official-image-count-budget 100
+		}
 		revision ">=3.7 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		}

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

@@ -8,12 +8,27 @@ provider "cursor" {
 	}
 
 	thinking-mode "effort"
-	// Cursor's default route selects a differently priced backing model per request.
+	// Cursor relays its upstream's image count (Claude 600, GPT 1,500) and took
+	// a 64 MB body; 60 images at the 500 KiB resize target stay near 31 MB.
+	image-count-budget 60
+	// Cursor's default route selects a differently priced backing model per
+	// request, Claude and GPT included, so it keeps Claude's image count and
+	// their 1024px edge; Gemini and Composer read 1024px pages as well as
+	// larger ones.
 	models "default" {
 		pricing-status "variable"
+		image-count-budget 20
+		image-max-edge 1024
 	}
 
+	// Cursor re-encodes every image for its Claude and GPT upstreams to fit
+	// 1024×1024 before the model reads it; Gemini and Composer read larger
+	// images as sent.
 	class "anthropic" {
+		image-max-edge 1024
+		// Past 20 images Claude rejects any image edge over 3,000 px, checked
+		// before Cursor downsizes.
+		image-count-budget 20
 		family "fable" {
 			requires-cursor-tool-schema-projection #true
 			// Cursor advertised Context window 300k (Max 1M recovered elsewhere).
@@ -56,6 +71,8 @@ provider "cursor" {
 		}
 	}
 	class "openai" {
+		// GPT images are fitted to 1024×1024 like Claude's (see above).
+		image-max-edge 1024
 		revision ">=5.6 <5.7" {
 			// Cursor advertised GPT-5.6 Sol/Terra/Luna Context window 272k;
 			// raises unlabeled -fast variants from the 200k default.
@@ -711,3 +728,15 @@ provider "cursor" {
 		}
 	}
 }
+
+// The same 1024px edge for targets that name only Cursor's wire (`{ api, id }`,
+// no host). priority=-1: a fallback, so the equally specific provider rule
+// above wins when both match.
+on-api "cursor-agent" {
+	class "anthropic" priority=-1 {
+		image-max-edge 1024
+	}
+	class "openai" priority=-1 {
+		image-max-edge 1024
+	}
+}

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

@@ -9,6 +9,9 @@ provider "google-antigravity" {
 		image "google-gemini-cli"
 	}
 	web-search "gemini"
+	// No size rejection up to a 64 MB body; 60 images at the 500 KiB resize
+	// target stay near 41 MB. Gemini counts images only against its context.
+	image-count-budget 60
 
 	seed api="google-gemini-cli" base-url="https://daily-cloudcode-pa.googleapis.com" bundle="always" {
 		model "gemini-3-pro-image" name="Gemini 3 Pro Image" {
@@ -26,6 +29,18 @@ provider "google-antigravity" {
 	preserves-max-output-tokens #false
 
 	class "anthropic" {
+		// Claude here is Vertex Anthropic: one image whose base64 passes
+		// 5,242,880 bytes is a 400 "image exceeds 5 MB maximum" (half the
+		// direct API's 10 MiB), and 8001 px is a 400 like on the direct API.
+		// Declaring the contract makes the Cloud Code Assist wire clamp images
+		// for Claude the way `anthropic-messages` does; Gemini on the same
+		// gateway has no such limits and keeps them unset.
+		max-image-dimension 8000
+		max-image-payload-bytes 5242880
+		// The Claude upstream rejects a message over 30 MB and, past 20 images,
+		// any image edge over 3,000 px. The clamp above caps each image; with no
+		// byte budget on this wire, the count bounds the body.
+		image-count-budget 20
 		family "opus" {
 			thinking-mode "budget"
 		}
@@ -59,6 +74,17 @@ provider "google-antigravity" {
 			}
 		}
 	}
+	// Cloud Code Assist bills a Gemini 2.5 or 3.x image a fixed 1,089 tokens
+	// from 32px to 2048px (extreme strips bill less); priority=1 outranks the
+	// lineage's 1,120 at equal specificity. verified: billing matched this rule
+	// on Gemini 2.5 Flash Lite, 3.1 Flash Lite and 3.8 Flash.
+	class "gemini" priority=1 {
+		image-tokenization {
+			regime "fixed"
+			verified #true
+			tokens 1089
+		}
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "claude-opus-4-6" {
 		thinking-efforts "minimal" "low" "medium" "high"

package/src/compat/rules/providers/google-gemini-cli.kdl

@@ -2,6 +2,7 @@
 
 provider "google-gemini-cli" {
 	default-model "gemini-3.1-pro-preview"
+	image-count-budget 200
 
 	class "gemini" {
 		// Cloud Code Assist rejects an unsigned first function call; the bypass

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

@@ -3,6 +3,7 @@
 provider "google-vertex" {
 	default-model "gemini-3.1-pro-preview"
 	allow-unauthenticated #true
+	image-count-budget 200
 
 	// Replaces google-shared.ts stripping functionCall/functionResponse ids for Vertex.
 	supports-function-part-id #false

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

@@ -7,6 +7,7 @@ provider "google" {
 		image "google-generative-ai"
 	}
 	web-search "gemini"
+	image-count-budget 200
 
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"

package/src/compat/rules/providers/image-tokenization.kdl

@@ -1,6 +1,7 @@
 // Image billing for readers whose lineage has no `image-tokenization` rule in
 // classes/*.kdl, by the wire API carrying the request. priority=-1: these are
-// fallbacks, so a class-wide lineage rule (same specificity) always wins.
+// fallbacks, so a class-wide lineage rule (same specificity) always wins. None
+// is `verified`: nobody measured the billing of the readers they price.
 
 // OpenAI-compatible wires bill like GPT-5.5: 32px patches × 1.2, `auto` sized
 // like `original`.

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

@@ -54,6 +54,19 @@ provider "meta" {
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}
+		// Muse Wonder: dev.meta.ai/docs/models + pricing-rate-limits (2026-10-12).
+		// `max` documented; no contributor tier on Model API. Meta publishes no
+		// output cap, so it takes the Muse Spark ceiling.
+		model "muse-wonder-1.0" name="Muse Wonder" {
+			reasoning #true
+			input "text" "image"
+			cost input=4 output=20 cache-read=0.2 cache-write=0
+			limits context=1048576 max-tokens=131072
+			thinking-mode "effort"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
+			supports-reasoning-effort #true
+			include-encrypted-reasoning #true
+		}
 		model "muse-spark-1.3-contributor" name="Muse Spark 1.3 (C)" {
 			reasoning #true
 			input "text" "image"
@@ -72,7 +85,10 @@ provider "meta" {
 	// api.meta.ai/v1 accepts only tool_choice "auto"; "none", "required" and named
 	// choices are rejected with 400 (verified 2026-09-10), so omit the field.
 	supports-tool-choice #false
+	// `/v1/models` carries no capability metadata: an id no seed or taxonomy
+	// knows is offered this ladder for the user to opt into.
 	class "unknown" {
+		thinking-unverified #true
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

package/src/compat/rules/providers/muse-code.kdl

@@ -27,7 +27,16 @@ provider "muse-code" {
 	// `custom` tools are rejected with 400 "`custom` tools are not supported
 	// on this endpoint", so apply-patch must stay a function tool here.
 	edit-prompt-variant "compact"
+	on-api "openai-responses" {
+		// Spark 1.3 Contributor warmed siblings hit 10/10 vs 0/10 with per-child keys (2026-10-12).
+		// Meta documents shared-prefix prompt_cache_key routing; keep the opt-in on these exact Muse Code IDs.
+		models "muse-spark-1.3" "muse-spark-1.3-contributor" "muse-wonder-1.0" "muse-wonder-1.0-contributor" "muse-wonder-full-1.0" "muse-wonder-full-1.0-contributor" {
+			task-sibling-prompt-cache #true
+		}
+	}
+	// Same capability-less `/v1/models` as Meta: unknown ids are opt-in.
 	class "unknown" {
+		thinking-unverified #true
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

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

@@ -9,6 +9,14 @@ provider "openai-codex" {
 	}
 	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
+
 	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" {

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/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`

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

@@ -5,6 +5,7 @@ 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

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

@@ -9,6 +9,9 @@ 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
 
 	// 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

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/runtime/behavior.kdl

@@ -174,9 +174,9 @@ behavior {
 	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-"
 	// 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" \

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

@@ -1,10 +1,13 @@
 class "meta" {
 	bounded "muse-spark"
+	bounded "muse-wonder"
 	bounded "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-"
 }