@oh-my-pi/pi-catalog 18.2.0 → 18.2.2

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

@@ -39,4 +39,50 @@ provider "devin" {
 			supports-parallel-tool-calls #true
 		}
 	}
+
+
+	// Devin omits cost dimensions for plan-included models on lower tiers
+	// (PRO reports $0 for SWE-2, SWE-1.7, and GLM-5.2 High). `cost-fallback`
+	// seeds the enterprise-tier list price only when upstream reports no
+	// token price, so accounts whose discovery does carry dimensions keep
+	// their real rates.
+
+	// SWE-2: enterprise promo is 75% off list through 2026-12-31; list
+	// pricing ($3/$15, $0.30 cache read) applies from 2027-01-01.
+	models "swe-2*" {
+		cost-fallback {
+			input 0.75
+			output 3.75
+			cache-read 0.075
+			cache-write 0.75
+			effective-rates {
+				list-price {
+					effective-from "2027-01-01T00:00:00Z"
+					input 3
+					output 15
+					cache-read 0.3
+					cache-write 3
+				}
+			}
+		}
+	}
+
+	// SWE-1.7 and GLM-5.2 High: plan-included on PRO, no published promo
+	// schedule — enterprise list is the only published per-token price.
+	models "swe-1-7" "swe-1-7-medium" {
+		cost-fallback {
+			input 0.5
+			output 2.5
+			cache-read 0.2
+			cache-write 0.5
+		}
+	}
+	models "glm-5-2" {
+		cost-fallback {
+			input 1.4
+			output 4.4
+			cache-read 0.26
+			cache-write 1.4
+		}
+	}
 }

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/kimi-code.kdl

@@ -39,4 +39,52 @@ provider "kimi-code" {
 		thinking-efforts "minimal" "low" "medium" "high"
 		thinking-format "zai"
 	}
+
+	// /coding/v1/models carries no pricing. `cost-fallback` seeds the public
+	// Moonshot list price only when discovery reports $0, so a future priced
+	// envelope keeps its real rates. `kimi-for-coding` is K2.8 Preview — no
+	// public API SKU exists yet, so it uses the K2.7 Code line rate.
+	models "k3" "k3-256k" {
+		cost-fallback {
+			input 3.0
+			output 15.0
+			cache-read 0.3
+		}
+	}
+	models "kimi-for-coding" {
+		cost-fallback {
+			input 0.95
+			output 4.0
+			cache-read 0.19
+		}
+	}
+	models "kimi-for-coding-highspeed" {
+		cost-fallback {
+			input 1.9
+			output 8.0
+			cache-read 0.38
+		}
+	}
+	// Moonshot K2 list (`kimi-k2-0711-preview` / `kimi-k2-0905-preview`), not K2.6.
+	models "kimi-k2" {
+		cost-fallback {
+			input 0.6
+			output 2.5
+			cache-read 0.15
+		}
+	}
+	models "kimi-k2.5" {
+		cost-fallback {
+			input 0.6
+			output 3.0
+			cache-read 0.1
+		}
+	}
+	models "kimi-k2-turbo-preview" {
+		cost-fallback {
+			input 2.4
+			output 10.0
+			cache-read 0.6
+		}
+	}
 }

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*"