@oh-my-pi/pi-catalog 18.12.2 → 18.12.4

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
+		}
+	}
+}