@linxiraos/pi-catalog 1.3.2 → 1.4.1

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

@@ -163,34 +163,36 @@ class "anthropic" {
 	// Foundry / Claude Platform on AWS are first-party Anthropic routes whose
 	// effective URLs are checked at request time.
 	on "anthropic" "google-vertex" "amazon-bedrock" "bedrock-mantle" {
-		family "fable" {
-			revision ">=5" {
-				supports-server-compaction #true
+		on-api "anthropic-messages" {
+			family "fable" {
+				revision ">=5" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "mythos" {
-			revision ">=5" {
-				supports-server-compaction #true
+			family "mythos" {
+				revision ">=5" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "opus" {
-			revision ">=4.6" {
-				supports-server-compaction #true
+			family "opus" {
+				revision ">=4.6" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "sonnet" {
-			revision ">=4.6" {
-				supports-server-compaction #true
+			family "sonnet" {
+				revision ">=4.6" {
+					supports-server-compaction #true
+				}
 			}
-		}
-		family "haiku" {
-			revision ">=5.5 <6" {
+			family "haiku" {
+				revision ">=5.5 <6" {
+					supports-server-compaction #true
+				}
+			}
+			models "claude-mythos-preview" {
 				supports-server-compaction #true
 			}
 		}
-		models "claude-mythos-preview" {
-			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

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

@@ -108,4 +108,35 @@ class "deepseek" {
 	revision ">=4" {
 		portable-reasoning #true
 	}
+	// V4.1 Flash's vision config (patch 14, 3x3 merge, 1024-token cap, 544²
+	// floor). verified: OpenRouter billing matched it exactly on Fireworks and
+	// DeepInfra and within one token on Together, 256px to 4000px.
+	// Later Flash releases stay on their wire fallback until measured.
+	family "flash" {
+		revision ">=4.1 <4.2" {
+			image-tokenization {
+				regime "deepseek-patch"
+				verified #true
+				patch-size 14
+				downsample 3
+				max-tokens 1024
+				min-pixels 295936
+			}
+		}
+	}
+	// DeepSeek's API retired the V4.0 vision preview and serves its legacy id
+	// with the current Flash (api-docs.deepseek.com/guides/vision); hosts that
+	// still run the preview keep their fallback.
+	on "deepseek" {
+		models "deepseek-v4-flash-vision-exp" {
+			image-tokenization {
+				regime "deepseek-patch"
+				verified #true
+				patch-size 14
+				downsample 3
+				max-tokens 1024
+				min-pixels 295936
+			}
+		}
+	}
 }

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

@@ -11,4 +11,15 @@ class "gemma" {
 	on "google" "vercel-ai-gateway" {
 		thinking-mode "budget"
 	}
+	// Gemma 4 on the Gemini API toggles thinking only through `thinkingLevel`
+	// (`HIGH` on, `MINIMAL` off) and rejects any `thinkingBudget` with a 400
+	// (https://ai.google.dev/gemma/docs/core/gemma_on_gemini_api#thinking).
+	// Off must be sent explicitly as `MINIMAL` rather than omitting the config.
+	on "google" {
+		revision ">=4" {
+			thinking-mode "google-level"
+			thinking-efforts "high"
+			thinking-suppress-when-off #true
+		}
+	}
 }

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

@@ -81,15 +81,17 @@ class "minimax" {
 			thinking-efforts "low" "medium" "high"
 		}
 	}
-	// MiniMax-M3.1-Flash-Preview always thinks (`thinking.type: "disabled"` and
-	// effort `none` 400 with error 2013) and tunes depth through
-	// `output_config.effort` from `low` to `max` (the default when omitted).
-	// The identity map replaces the M3 family's adaptive-tag-only map so each
-	// level reaches the wire. Like the M2 family, thinking cannot be turned off:
-	// `thinking-requires-effort` hides the off level and clamps forced-off
-	// requests to the lowest effort instead of sending `disabled`.
+	// MiniMax-M3.1-Flash (and its bundled `-Preview` id) always thinks
+	// (`thinking.type: "disabled"` and effort `none` 400 with error 2013) and
+	// tunes depth through `output_config.effort` from `low` to `max` (the
+	// default when omitted). The identity map replaces the M3 family's
+	// adaptive-tag-only map so each level reaches the wire. Like the M2 family,
+	// thinking cannot be turned off: `thinking-requires-effort` hides the off
+	// level and clamps forced-off requests to the lowest effort instead of
+	// sending `disabled`. The launch id also rejects sampling params
+	// (`temperature`, `top_p`, `top_k`).
 	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
-		models "MiniMax-M3.1-Flash-Preview" {
+		models "MiniMax-M3.1-Flash" "MiniMax-M3.1-Flash-Preview" {
 			thinking-requires-effort #true
 			thinking-efforts "low" "medium" "high" "xhigh" "max"
 			thinking-effort-map {
@@ -98,6 +100,9 @@ class "minimax" {
 				high "high"
 			}
 		}
+		models "MiniMax-M3.1-Flash" {
+			supports-sampling-params #false
+		}
 	}
 	// Token Plan rows ride the Anthropic route; these OpenAI-only wire axes
 	// serve the legacy lightning rows that stay on the OpenAI-compatible one.

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

@@ -51,10 +51,66 @@ provider "amazon-bedrock" {
 		thinking-efforts "minimal" "low" "medium" "high"
 		thinking-mode "budget"
 	}
+	// AWS documents compact-2026-09-04 for native Runtime APIs, not Converse.
+	// Expose this only as a side-compaction capability; ordinary turns
+	// continue to use the model's bedrock-converse-stream API. These revisions
+	// are limited to the Bedrock model cards that name the new beta:
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-compaction.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-4-5.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-5-5.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-4-7.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-4-8.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5.html
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5-5.html
+	// The haiku/sonnet/opus families are Anthropic-only (taxonomy/anthropic.kdl).
+	on-api "bedrock-converse-stream" {
+		family "haiku" {
+			revision ">=4.5 <4.6" {
+				supports-server-compaction #true
+			}
+			revision ">=5.5 <5.6" {
+				supports-server-compaction #true
+			}
+		}
+		family "sonnet" {
+			revision ">=5 <5.1" {
+				supports-server-compaction #true
+			}
+			revision ">=5.5 <5.6" {
+				supports-server-compaction #true
+			}
+		}
+		family "opus" {
+			revision ">=4.7 <4.9" {
+				supports-server-compaction #true
+			}
+			revision ">=5 <5.1" {
+				supports-server-compaction #true
+			}
+			revision ">=5.5 <5.6" {
+				supports-server-compaction #true
+			}
+		}
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "moonshot.kimi-k2-thinking" {
 		thinking-requires-effort #true
 	}
+	// residue: the inference-profile prefixes leave the K3 rows in the
+	// "unknown" class, so an exact selector is the only match. K3 reasons via
+	// OpenAI-style reasoning_effort (low/high/max, default max), matching
+	// Moonshot's official API and every other provider's K3 entry. Verified
+	// live via Converse on global.moonshotai.kimi-k3: identical prompt yields
+	// ~1K output tokens at "high" vs ~2.4K at "max" (invalid values are
+	// accepted silently and behave like the max default).
+	models "*.moonshotai.kimi-k3" {
+		thinking-mode "effort"
+		thinking-efforts "low" "high" "max"
+		thinking-default-level "max"
+		thinking-requires-effort #true
+	}
 	// The Qwen rows arrive with maxTokens copied from the upstream context
 	// window (262000 on a 262144 window), so a request that does not set an
 	// output budget of its own asks Bedrock for more than the model accepts

package/src/compat/rules/providers/factory-droid.kdl

@@ -12,6 +12,18 @@ provider "factory-droid" {
 	dynamic-models-authoritative #true
 	// Native subscription limits are not direct-host price-tier or capacity heuristics.
 	context-window-authoritative #true
+	// Factory's hosted HTTP routes refuse request bodies over 4,500,000 bytes.
+	// The native CLI keeps inline media under 3,500,000 base64 bytes, leaving
+	// ~1 MB for system, tools and text, but never evicts the newest user
+	// message's or newest tool result's media. Applied only on Factory's own
+	// hosts; a custom gateway baseUrl keeps its own limit.
+	inline-image-byte-budget 3500000
+	request-body-byte-limit 4500000
+	// Factory's own hosts per residency (wire/factory-droid.ts factoryDroidApiBaseUrl);
+	// seed rows are never bundled, so they are listed here for isCatalogEndpointUrl,
+	// which gates the budgets above (Factory's compat names no official endpoint).
+	endpoint-hosts "api.factory.ai" "api.eu.factory.ai"
+	catalog-endpoint-budgets #true
 
 	// The registry roster. Factory has no model-listing endpoint, so first-party
 	// clients ship this list and narrow it per account with feature flags, org

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

@@ -0,0 +1,61 @@
+// https://docs.inceptionlabs.ai/get-started/models (2026-10-20).
+provider "inception" {
+	default-model "mercury-3"
+	env "INCEPTION_API_KEY"
+	dynamic-models-authoritative #true
+	discovery label="Inception"
+
+	seed api="openai-completions" base-url="https://api.inceptionlabs.ai/v1" bundle="fallback" {
+		model "mercury-3" name="Mercury 3" {
+			reasoning #true
+			input "text"
+			cost input=0.1 output=0.35 cache-read=0.01 cache-write=0
+			limits context=1000000 max-tokens=131072
+		}
+		model "mercury-2.5" name="Mercury 2.5" {
+			reasoning #true
+			input "text"
+			cost input=0.2 output=0.75 cache-read=0.02 cache-write=0
+			limits context=260000 max-tokens=65536
+		}
+		model "mercury-2" name="Mercury 2" {
+			reasoning #true
+			input "text"
+			cost input=0.25 output=0.75 cache-read=0.025 cache-write=0
+			limits context=128000 max-tokens=50000
+		}
+	}
+
+	supports-developer-role #false
+	supports-store #false
+	max-tokens-field "max_completion_tokens"
+	clamp-output-to-model-max #true
+
+	// No Mercury taxonomy exists; these reviewed chat ids expose reasoning
+	// controls omitted by /models. Other endpoints must not inherit them.
+	models "mercury-2" "mercury-2.5" "mercury-3" {
+		thinking-upgrade-neutral #true
+		thinking-mode "effort"
+		thinking-efforts "minimal" "low" "medium" "high"
+		thinking-effort-map { minimal "instant" }
+		thinking-default-level "medium"
+		supports-reasoning-effort #true
+		reasoning-disable-mode "lowest-effort"
+	}
+	models "mercury-3" priority=1 {
+		thinking-default-level "low"
+		// Mercury streams nothing while it reasons. High effort defaults to a
+		// 131,072-token budget; at the measured ~400 reasoning tokens/s that
+		// can take ~330s before the first chunk, past the 300s default.
+		stream-first-event-timeout-ms 600000
+		// /models reports $0. `cost-fallback` seeds the published launch price
+		// ($0.10/$0.35, 50% off the $0.20/$0.70 list price, no announced end) only
+		// while it does, so a real upstream tariff takes over once published.
+		cost-fallback {
+			input 0.1
+			output 0.35
+			cache-read 0.01
+			cache-write 0
+		}
+	}
+}

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

@@ -33,15 +33,15 @@ provider "moonshot" {
 			input-modalities "text" "image"
 		}
 	}
-	// stencil.so reports a $3 K3 cache write; Moonshot bills only cache hits
-	// ($0.30) and misses (the $3 input rate), with no cache-write charge.
+	// Official K3 rates: cache write is billed separately by TTL ($3 for the
+	// default 5min tier, $6 for 1h). omp sends no TTL, so the 5min rate applies.
 	// https://platform.moonshot.ai/docs/pricing/chat-k3
 	models "kimi-k3" {
 		cost-patch {
 			input 3.0
 			output 15.0
 			cache-read 0.3
-			cache-write 0.0
+			cache-write 3.0
 		}
 	}
 	// The /models roster carries no modality metadata for the named vision/vl

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

@@ -24,6 +24,9 @@ provider "openai-codex" {
 	// 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"
+	// Logged-in Codex accounts keep prompt-cache entries with no TTL, so
+	// idle compaction would only discard a prefix that stays warm.
+	prompt-cache-persistent #true
 
 	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.

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

@@ -1,37 +1,27 @@
 // Provider-wire compat for "yolo-auto"; wire constraints apply to every discovered model.
 
 provider "yolo-auto" {
-	default-model "qwen3.8-flash"
+	default-model "qwen3.8-27b"
 	env "YOLO_AUTO_API_KEY"
 	dynamic-models-authoritative #true
 	discovery label="Yolo-Auto"
 
-	// Documented Yolo-Auto catalog (yolo-auto.com/models, yolo-auto.com/docs):
-	// `qwen3.8-flash` is the default model on Free and paid plans at a 256K
-	// (262144) context window, and the paid `yolo` route is initially backed by
-	// the same Flash deployment. The flat-rate `/v1/models` response carries
-	// bare ids with no limit fields and is authoritative once discovery runs.
-	// Without provider-local seeds the bare `qwen3.8-flash` slug resolves
-	// through the global reference index, which prefers the largest bundled
-	// window (1M) and misreports the deployment cap. The compat block mirrors
-	// the provider's documented wire surface: the API speaks the Qwen
-	// chat-template dialect with `reasoning_effort` support and rejects the
-	// `developer` role and `store` param.
+	// Documented Yolo-Auto catalog (yolo-auto.com/models, yolo-auto.com/docs,
+	// 2026-10): `qwen3.8-27b`, `yolo`, `yolo-small` (256K) and `yolo-128k`
+	// (128K), all paid-only. `qwen3.8-27b` is the recommended default; the
+	// `yolo` route is currently backed by the same Qwen3.8-27B deployment
+	// (its server-side target can change). The flat-rate `/v1/models`
+	// response is authoritative once discovery runs. Without provider-local
+	// seeds the bare `qwen3.8-27b` slug resolves through the global reference
+	// index, which prefers the largest bundled window (1M) and misreports the
+	// deployment cap. The compat block mirrors the provider's documented wire
+	// surface: the API speaks the Qwen chat-template dialect with
+	// `reasoning_effort` support and rejects the `developer` role and `store`
+	// param. `yolo-small` and `yolo-128k` are documented only by id and
+	// context, and their `/v1/models` rows advertise no thinking ladder, so
+	// they seed as plain text routes until the provider documents more.
 	seed api="openai-completions" base-url="https://yolo-auto.com/v1" bundle="fallback" {
-		model "deepseek-flash-v4" name="DeepSeek Flash V4" {
-			reasoning #true
-			input "text" "image"
-			cost input=0 output=0 cache-read=0 cache-write=0
-			limits context=131072
-			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
-			thinking-effort-map { minimal "low"; low "low"; medium "high"; high "high"; xhigh "max"; max "max" }
-			supports-developer-role #false
-			supports-store #false
-			supports-reasoning-effort #true
-			thinking-format "chat-template"
-		}
-		model "qwen3.8-flash" name="Qwen3.8 Flash" {
+		model "qwen3.8-27b" name="Qwen3.8-27B" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
@@ -57,6 +47,22 @@ provider "yolo-auto" {
 			thinking-format "qwen-chat-template"
 			template-reasoning-effort #true
 		}
+		model "yolo-small" name="Yolo Small" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=262144
+			supports-developer-role #false
+			supports-store #false
+		}
+		model "yolo-128k" name="Yolo 128K" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=131072
+			supports-developer-role #false
+			supports-store #false
+		}
 	}
 
 	supports-developer-role #false
@@ -64,17 +70,7 @@ provider "yolo-auto" {
 	// 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" {
+	models "qwen3.8-27b" "yolo" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
-	models "deepseek-flash-v4" {
-		disable-reasoning-on-tool-choice #true
-		max-tokens-field "max_completion_tokens"
-		requires-reasoning-content-for-all-assistant-turns #true
-		requires-reasoning-content-for-tool-calls #true
-		supports-reasoning-effort #true
-		thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
-		thinking-format "chat-template"
-		thinking-mode "effort"
-	}
 }

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

@@ -12,7 +12,7 @@ class "qwen" {
 
     revision prefix="qwen"
 
-    override id="yolo-auto-yolo-flash-identity" provider="yolo-auto" model="yolo" class="qwen" revision="3.8" rationale="Yolo-Auto's paid yolo route is initially backed by the same Qwen3.8 Flash deployment as qwen3.8-flash; the opaque alias carries no lineage tokens" provenance="yolo-auto.com/models (2026-09)"
+    override id="yolo-auto-yolo-identity" provider="yolo-auto" model="yolo" class="qwen" revision="3.8" rationale="Yolo-Auto's paid yolo route is currently backed by the same Qwen3.8-27B deployment as qwen3.8-27b; the opaque alias carries no lineage tokens" provenance="yolo-auto.com/models (2026-10)"
     override id="prismml-bonsai-27b-qwen-3-6" glob="*bonsai-27b*" class="qwen" revision="3.6" rationale="Reviewed PrismML Bonsai and Ternary-Bonsai 27B lineage is Qwen3.6-27B; the wire basename may carry a prefix or GGUF quantization suffix" provenance="packages/coding-agent/src/config/model-discovery.ts; original reviewed alias e3aa6594e5"
     override id="prismml-bonsai-2-27b-qwen-3-8" glob="*bonsai-2-27b*" class="qwen" revision="3.8" rationale="Reviewed PrismML Bonsai 2 27B lineage is Qwen3.8-27B; the wire basename may carry a prefix or GGUF quantization suffix" provenance="packages/coding-agent/src/config/model-discovery.ts"
     override id="kilo-qwq-32b-family" provider="kilo" model="qwq-32b" logical="qwen/qwq-32b" class="qwen" rationale="The reviewed Kilo QwQ deployment belongs to the Qwen family despite its opaque product spelling" provenance="fixtures/llm-oracle/catalog/models.normalized.json"