@oh-my-pi/pi-catalog 18.2.0 → 18.2.1

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

@@ -1,6 +1,12 @@
 // Model-lineage compat for "kimi"; regenerated from the frozen census using class taxonomy selectors.
 
 class "kimi" {
+	// Kimi chat templates can revise bytes they have already streamed: a leaned-on
+	// thinking opener reclassifies visible text as reasoning, and native
+	// section/tool-call tokens are carved out of the visible channel mid-stream.
+	// Declared revision-possible so finished prose stays in the live viewport
+	// instead of retiring into native scrollback ahead of the turn.
+	stream-revision "possible"
 	// Replaces the K2.6 reasoning stream-idle baseline on OpenAI-compatible routes.
 	on "baseten" "cline-pass" "cloudflare-ai-gateway" "coreweave" "deepinfra" "firepass" "fireworks" \
 		"huggingface" "moonshot" "novita" "nvidia" "opencode-go" "opencode-zen" "openrouter" \

package/src/compat/rules/providers/alibaba-token-plan.kdl

@@ -114,10 +114,17 @@ provider "alibaba-token-plan" {
 	models "deepseek-v4-pro" {
 		thinking-format "qwen"
 	}
-	// residue: taxonomy ranks and exact globs do not isolate these models.
-	models "qwen3.8-max" {
-		thinking-default-level "xhigh"
-		thinking-efforts "low" "medium" "xhigh"
+	// Max Preview rides only Alibaba's binary enable_thinking toggle: it stays on
+	// the Qwen dialect and must not emit an unsupported reasoning_effort field.
+	models "qwen3.8-max-preview" {
+		thinking-efforts "low" "high" "xhigh"
+		thinking-requires-effort #true
+	}
+	// Max and Flash steer thinking depth through the OpenAI reasoning_effort
+	// control and replay reasoning_content across turns (preserve_thinking
+	// defaults on for both).
+	models "qwen3.8-max" "qwen3.8-flash" {
+		replay-reasoning-content #true
 		when-thinking {
 			extra-body {
 				enable_thinking #true
@@ -125,9 +132,10 @@ provider "alibaba-token-plan" {
 			thinking-format "openai"
 		}
 	}
-	// residue: taxonomy ranks and exact globs do not isolate these models.
-	models "qwen3.8-max-preview" {
-		thinking-efforts "low" "high" "xhigh"
-		thinking-requires-effort #true
+	// Max exposes a distinct low/medium/xhigh ladder; Flash retains its
+	// provider-authored minimal/low/medium/high ladder.
+	models "qwen3.8-max" {
+		thinking-default-level "xhigh"
+		thinking-efforts "low" "medium" "xhigh"
 	}
 }

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

@@ -3,7 +3,15 @@
 provider "amazon-bedrock" {
 	default-model "us.anthropic.claude-opus-4-8"
 
+	// The Converse assembler positions content blocks by wire index
+	// (`amazon-bedrock.ts`), so a block can land above text already rendered.
+	stream-revision "possible"
 	class "anthropic" {
+		family "fable" {
+			revision ">=5.1 <5.2" {
+				thinking-efforts "low" "medium" "high" "xhigh" "max"
+			}
+		}
 		family "opus" {
 			revision ">=4.6 <4.7" {
 				thinking-mode "anthropic-adaptive"
@@ -42,4 +50,26 @@ provider "amazon-bedrock" {
 	models "moonshot.kimi-k2-thinking" {
 		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
+	// and gets a 400 back ("you requested 262000 output tokens"). AWS
+	// publishes the real caps on the model cards — 8K for Qwen3-Next 80B A3B
+	// and Qwen3 VL 235B A22B, 16K for Qwen3 Coder Next — and the context
+	// windows already match, so only the output side needs patching (#12089).
+	models "qwen.qwen3-next-80b-a3b" {
+		limits-patch {
+			max-tokens 8000
+		}
+	}
+	models "qwen.qwen3-vl-235b-a22b" {
+		limits-patch {
+			max-tokens 8000
+		}
+	}
+	models "qwen.qwen3-coder-next" {
+		limits-patch {
+			max-tokens 16000
+		}
+	}
 }

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

@@ -19,4 +19,10 @@ provider "azure" {
 	models "gpt-5.1-codex-mini" {
 		thinking-efforts "medium" "high"
 	}
+	// Azure Chat Completions rejects Astra reasoning whenever function tools are
+	// present; the documented escape hatch is an explicit `none` effort.
+	models "gpt-6-astra*" {
+		disable-reasoning-with-tools #true
+		reasoning-disable-mode "none-effort"
+	}
 }

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

@@ -16,6 +16,16 @@ provider "cerebras" {
 		input-modalities "text" "image"
 	}
 	thinking-mode "effort"
+	// Cerebras accepts OpenAI `reasoning_effort`, not Qwen-native thinking
+	// fields such as `enable_thinking`.
+	class "qwen" {
+		thinking-format "openai"
+	}
+	// Qwen 3.8 accepts low/medium/high; `none` is its wire-only off value.
+	models "qwen-3.8-27b" {
+		thinking-efforts "low" "medium" "high"
+		reasoning-disable-mode "none-effort"
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "zai-glm-4.7" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"

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

@@ -1,10 +1,13 @@
 // Command Code Provider API deployment contract. Model capabilities and effort
 // ladders follow command-code@1.44.0: the thinking-efforts below reproduce the
 // bundle's 36-id effort registry 1:1 (verified against its effort map), plus
-// the 5 Muse Spark lineage ids whose ladder OMP's classes/meta.kdl carries.
-// Every other served id exposes no effort dial upstream (`modelSupports-
-// ReasoningEffort` is false for them in the CLI), so a missing ladder here
-// is a deliberate absence, not a gap.
+// the 5 Muse Spark lineage ids whose ladder OMP's classes/meta.kdl carries and
+// `deepseek/deepseek-v4.1-flash`, which postdates that bundle (see its rule).
+// Every id the 1.44.0 census serves without an effort dial keeps a missing
+// ladder here as a deliberate absence, not a gap (`modelSupportsReasoning-
+// Effort` is false for those ids in that bundle). Later bundles (1.53.0) hand
+// dials to a few further ids — MiniMaxAI/MiniMax-M3 and meta/muse-spark-1.3's
+// `max` — which stay out of this change.
 provider "commandcode" {
 	default-model "claude-sonnet-4-6"
 	env "COMMAND_CODE_API_KEY" "COMMANDCODE_API_KEY"
@@ -54,6 +57,19 @@ provider "commandcode" {
 		supports-reasoning-effort #true
 		thinking-efforts "low" "high" "max"
 	}
+	// `deepseek/deepseek-v4.1-flash` postdates command-code@1.44.0, so it is
+	// absent from that bundle's effort registry; command-code@1.53.0 lists it
+	// in the same low/high/max group as Kimi K3 and GLM-5.3, and the live docs
+	// model table carries the same scale plus image input.
+	models "deepseek/deepseek-v4.1-flash" {
+		supports-reasoning-effort #true
+		thinking-efforts "low" "high" "max"
+		// The class's `token="vision"` exemption misses this id (its name
+		// carries no `vision` token) even though the deployment serves it with
+		// image input, so the opt-out is explicit here.
+		strip-image-input #false
+		input-modalities "text" "image"
+	}
 	models "google/gemini-3.1-flash-lite" "google/gemini-3.5-flash" \
 		"google/gemini-3.5-flash-lite" "google/gemini-3.6-flash" "google/gemini-3.7-flash" \
 		"google/gemini-3.8-flash" "gpt-5.4-mini" "xai/grok-4.5" {

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

@@ -11,6 +11,15 @@ provider "cursor" {
 	class "anthropic" {
 		family "fable" {
 			requires-cursor-tool-schema-projection #true
+			// Cursor advertised Context window 300k (Max 1M recovered elsewhere).
+			context-window-floor 300000
+		}
+		family "opus" {
+			revision ">=5 <6" {
+				// Cursor advertised Claude Opus 5 Context window 300k; Max 1M
+				// stays recovered on labeled/maxMode rows via Math.max.
+				context-window-floor 300000
+			}
 		}
 	}
 	// GetUsableModels advertises no input modalities or context windows.
@@ -21,11 +30,34 @@ provider "cursor" {
 			input-modalities "text" "image"
 			context-window-floor 1000000
 		}
+		family "k2.7-code" {
+			// Cursor advertised Kimi K2.7 Code Context window 262k.
+			context-window-floor 262000
+		}
+	}
+	class "xai" {
+		family "grok" {
+			revision ">=4.5 <4.7" {
+				// Cursor advertised Grok 4.5/4.6 Context window 256k (product
+				// cap; not xAI native 500k). Max context is unmarked.
+				context-window-floor 256000
+			}
+		}
+	}
+	class "openai" {
+		revision ">=5.6 <5.7" {
+			// Cursor advertised GPT-5.6 Sol/Terra/Luna Context window 272k;
+			// raises unlabeled -fast variants from the 200k default.
+			context-window-floor 272000
+		}
 	}
 	models "k3" "*/k3" {
 		input-modalities "text" "image"
 		context-window-floor 1000000
 	}
+	models "default" {
+		context-window-floor 256000
+	}
 	// Cursor-only families verified to accept selectedImages even though the
 	// roster carries no modality metadata.
 	models "cursor-grok-4" "cursor-grok-4.*" "cursor-grok-4-*" "cursor-grok-4:*" "cursor-grok-4_*" {

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

@@ -17,8 +17,8 @@ provider "deepseek" {
 	// First-party peak USD / 1M tokens; all other UTC hours receive 50% off.
 	// Source: https://api-docs.deepseek.com/quick_start/pricing (2026-09-10).
 	// residue: the pricing page's Flash-priced names, not a taxonomy family: the
-	// bare alias carries no family, and the retired `v4-flash`/`-vision-exp` ids
-	// are still accepted and billed at the Flash card.
+	// retired `v4-flash`/`-vision-exp` ids are still accepted and billed at the
+	// Flash card.
 	models "deepseek-flash" "deepseek-v4-flash" "deepseek-v4-flash-vision-exp" {
 		clamp-output-to-model-max #true
 		time-based-cost {
@@ -43,9 +43,9 @@ provider "deepseek" {
 			cache-write 0
 		}
 	}
-	// residue: upstream discovery has not seeded the bare alias's limits; the
-	// pricing page documents this SKU (DeepSeek-V4.1-Flash) as 1M context / 384K
-	// output, and without them the agent cannot enforce its context budget.
+	// residue: the pricing page documents this SKU (DeepSeek-V4.1-Flash) as 1M
+	// context / 384K output; pinned as a reviewed correction so the documented
+	// window survives an upstream metadata gap.
 	models "deepseek-flash" {
 		limits-patch {
 			context-window 1000000

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

@@ -11,6 +11,21 @@ provider "google-vertex" {
 	// Vertex rejects `output_config.effort` and its effort beta header;
 	// replaces the provider checks in anthropic.ts request shaping.
 	supports-output-effort #false
+
+	// Vertex rejects maxOutputTokens=65536 for the Gemini 2.5 Flash-Lite line
+	// with 400 "supported range is from 1 (inclusive) to 65536 (exclusive)";
+	// the 2.5 Flash/Pro siblings accept 65536 on the same endpoint (Pro does
+	// not validate the field at all), so only the Lite line is clamped here.
+	// Other hosts are unverified and keep the documented value.
+	class "gemini" {
+		family "lite" {
+			revision ">=2.5 <2.6" {
+				limits-patch {
+					max-tokens 65535
+				}
+			}
+		}
+	}
 	class "anthropic" {
 		revision ">=4.5 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"

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

@@ -69,6 +69,9 @@ provider "meta" {
 	include-encrypted-reasoning #true
 	supports-reasoning-effort #true
 	clamp-output-to-model-max #true
+	// 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
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"

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

@@ -12,6 +12,9 @@ provider "muse-code" {
 	include-encrypted-reasoning #true
 	supports-reasoning-effort #true
 	clamp-output-to-model-max #true
+	// 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
 	// Subscription tier bills per request; the compact edit prompt cuts the
 	// recurring tool-schema bytes. Verified 2026-09-05 against api.meta.ai/v1:
 	// `custom` tools are rejected with 400 "`custom` tools are not supported

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

@@ -17,6 +17,12 @@ provider "openrouter" {
 		// isOpenRouterAnthropicModel checks in the OpenAI request builders.
 		retry-without-strict-on-grammar-error #true
 	}
+	class "meta" {
+		family "muse-spark" {
+			filter-reasoning-history #true
+			allows-synthetic-reasoning-content-for-tool-calls #false
+		}
+	}
 	class "minimax" {
 		family "m3" {
 			thinking-efforts "minimal" "low" "medium" "high"

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

@@ -20,6 +20,12 @@ behavior {
 	// `opencode` provider in the picker.
 	retired-providers "wafer-pass" "wandb" "opencode"
 
+	// ClinePass exposes gateway-specific limits, pricing, and reasoning
+	// controls; matching them by bare model id contaminates unrelated proxy
+	// models (a LiteLLM proxy serving `glm-5.3-flash`), so its rows are barred
+	// from seeding cross-provider bare-id enrichment references (#10932).
+	reference-isolated-providers "cline-pass"
+
 	// Discovered Responses/Codex models receive these additional operations.
 	model-operations provider="openai" {
 		exact "o3"
@@ -83,6 +89,11 @@ behavior {
 			glob="*/gpt-5.6-sol-pro" glob="*/gpt-5.6-luna" glob="*/gpt-5.6-luna-pro"
 	}
 
+	// Z.AI and Zhipu emit timezone-naive account-reset wall clocks in Beijing
+	// time. Provider-aware retry paths append this offset before parsing.
+	retry-reset-timezone provider="zai" offset="+08:00"
+	retry-reset-timezone provider="zhipu-coding-plan" offset="+08:00"
+
 	// Cloudflare AI Gateway namespace routing. The Workers AI prefix remains
 	// on the request id; the upstream namespaces are stripped.
 	api-routes provider="cloudflare-ai-gateway" {
@@ -102,6 +113,12 @@ behavior {
 
 	// Provider-default wire model for model-less hosted search operations.
 	hosted-default provider="kimi-search" model="kimi-for-coding"
+	// LiteLLM publishes task-specific endpoint models beside conversational
+	// models. Exact mode metadata avoids brittle model-name heuristics while
+	// unknown and malformed future values remain selectable.
+	exclude-discovery-modes "audio_speech" "audio_transcription" "batch" "embedding" "guardrail" \
+		"image_edit" "image_generation" "moderation" "ocr" "rerank" "search" "vector_store" \
+		"video_generation" provider="litellm"
 	// models.dev rows OMP cannot route successfully: AWS does not document
 	// the jp. Opus 5 profile, and the openai.gpt-5.x rows are Mantle-only
 	// ids that Bedrock rejects or misroutes.

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

@@ -5,6 +5,11 @@ class "deepseek" {
     family "r1" glob="*deepseek-r1*"
     family "reasoner" glob="*deepseek-reasoner*"
     family "flash" glob="*deepseek*v4*flash*"
+    // V4.1 Flash dropped the generation digit from its release id: the native
+    // roster serves it as bare `deepseek-flash`, which the `v4` glob cannot
+    // match. Without a family the descriptor's V4-generation filter drops the
+    // row, so it survived only as a stale previous-snapshot entry.
+    family "flash" glob="*deepseek-flash"
     family "pro" glob="*deepseek*v4*pro*"
     // Bare V4 chat/reasoner ids without a flash/pro segment.
     family "v4" glob="*deepseek-v4*"