@oh-my-pi/pi-catalog 18.4.4 → 18.4.6

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

@@ -0,0 +1,120 @@
+// Provider-wire compat for "helmcode": Helmcode is an OpenAI-compatible gateway
+// for private EU inference (helmcode.com/docs). One base URL serves three
+// plan-covered open-weight chat models, the per-key GLM 5.3 add-on, and nine
+// resold frontier models (Claude/GPT/Gemini). Non-chat rows on `/v1/models`
+// (embeddings, rerank, kokoro TTS, whisper STT) are excluded via
+// runtime/behavior.kdl.
+
+provider "helmcode" {
+	default-model "deepseek-v4-flash"
+	env "HELMCODE_API_KEY"
+	dynamic-models-authoritative #true
+	// Helmcode's own endpoint and docs are the deployment truth: same-id rows on
+	// other hosts carry different output ceilings and effort ladders, so the
+	// generator must not backfill names, limits, or capabilities from them.
+	// Runtime discovery still resolves resold frontier ids (Claude, GPT,
+	// Gemini) against the first-party vendor's bundled row for context, output
+	// cap, modalities, and list price; Helmcode passes that list price through
+	// with no markup, converted to EUR (helmcode.com/docs/models, "Frontier
+	// models"), so the surfaced USD cost is the vendor's rate card.
+	skip-cross-provider-reference-fills #true
+	discovery label="Helmcode"
+
+	// Bundled fallback so the provider is usable before the first authoritative
+	// discovery call. The open-weight chat models are flat-rate under the plan
+	// and GLM 5.3 is a per-key monthly add-on, so neither has a per-token price
+	// (surfaced as 0). Resold frontier models are priced per token in the
+	// console only; they arrive through runtime discovery.
+	seed api="openai-completions" base-url="https://api.helmcode.com/v1" bundle="fallback" {
+		model "deepseek-v4-flash" name="DeepSeek V4 Flash" {
+			reasoning #true
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1000000
+		}
+		model "qwen3.6" name="Qwen3.6" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=262144
+		}
+		model "gemma4" name="Gemma 4" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=262144
+		}
+		model "glm5.3" name="GLM 5.3" {
+			reasoning #true
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1000000
+		}
+	}
+
+	// `reasoning_effort` ladders per helmcode.com/docs/models ("Controlling
+	// reasoning"). Provider-scoped rather than seed-only so rows discovered at
+	// runtime resolve the same ladder, and no class-level host ladder (which
+	// sends `xhigh`, a value Helmcode does not document) applies. `/v1/models`
+	// carries no capability data, so a discovered id with no bundled row, no
+	// vendor row, and no ladder below arrives as non-reasoning.
+	thinking-mode "effort"
+	// residue: taxonomy extracts no revision from Helmcode's dotless ids.
+	// `minimal` through `max` only steer this adaptive model; `none` is the
+	// one value that skips reasoning, so thinking-off sends it.
+	models "deepseek-v4-flash" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "minimal" "low" "medium" "high" "max"
+		reasoning-disable-mode "none-effort"
+	}
+	// residue: taxonomy extracts no revision from Helmcode's dotless ids.
+	// Helmcode serves Qwen3.6 on the OpenAI `reasoning_effort` field, not the
+	// Qwen class's `enable_thinking` toggle, which would drop the chosen tier.
+	// `none` and `minimal` both skip reasoning, so `minimal` is not a tier and
+	// thinking-off sends `none`. With no field both default to a 16,384 token
+	// budget, which is the `high` tier.
+	models "qwen3.6" "gemma4" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "high" "max"
+		thinking-default-level "high"
+		thinking-format "openai"
+		reasoning-disable-mode "none-effort"
+	}
+	// residue: taxonomy extracts no revision from Helmcode's dotless ids.
+	models "glm5.3" "glm5.3-flash" "glm5.2" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "high" "max"
+		thinking-default-level "max"
+	}
+
+	// The resold frontier models run on their vendors behind this same
+	// chat-completions surface, and Helmcode documents no `reasoning_effort`
+	// values for them. Keep them to the core low/medium/high tiers of that
+	// field so the inherited vendor ladders never send `minimal`, `xhigh`, or
+	// `max`, nor a vendor-native thinking shape, to this host. priority=1
+	// outranks the equally scoped lineage ladders in classes/*.kdl; Gemini's
+	// ladders are family-scoped there, so the override is too.
+	class "anthropic" priority=1 {
+		thinking-mode "effort"
+		thinking-efforts "low" "medium" "high"
+	}
+	// Helmcode bills cache writes only for Anthropic models; OpenAI and Google
+	// cache creation is free there, so the inherited OpenAI write rate is zeroed.
+	class "openai" priority=1 {
+		thinking-mode "effort"
+		thinking-efforts "low" "medium" "high"
+		cost-patch {
+			cache-write 0.0
+		}
+	}
+	class "gemini" {
+		family "flash" priority=1 {
+			thinking-mode "effort"
+			thinking-efforts "low" "medium" "high"
+		}
+		family "lite" priority=1 {
+			thinking-mode "effort"
+			thinking-efforts "low" "medium" "high"
+		}
+	}
+}

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

@@ -89,6 +89,23 @@ behavior {
 		fallback "spark" substring="-spark"
 		fallback "chat" substring="gpt-"
 	}
+	// Factory Droid bills each model from one subscription pool: Droid Core
+	// (flat rate) or Standard Credits. Unknown ids have no inferred entitlement.
+	quota-tiers provider="factory-droid" {
+		tier "core" "inkling" "mistral-medium-3.5" "glm-5.3-flash" "glm-5.3" "glm-5.2" "glm-5.2-fast" \
+			"kimi-k3" "qwen3.8-max" "nemotron-3-ultra" "deepseek-v4.1-flash" "deepseek-v4-flash-0731" \
+			"deepseek-v4-pro" "minimax-m3" "minimax-m2.7"
+		tier "standard" "claude-fable-5.1" "claude-fable-5" "claude-opus-5-5" "claude-opus-5-5-fast" \
+			"claude-opus-5" "claude-opus-5-fast" "claude-opus-4-8" "claude-opus-4-8-fast" "claude-opus-4-7" \
+			"claude-opus-4-6" "claude-opus-4-5-20251101" "claude-sonnet-5-5" "claude-sonnet-5" \
+			"claude-sonnet-4-6" "claude-sonnet-4-5-20250929" "claude-haiku-4-5-20251001" "gpt-6-astra" \
+			"gpt-6.1-sol" "gpt-6-sol" "gpt-6-luna" "gpt-5.6-sol" "gpt-5.6-sol-fast" "gpt-5.6-terra" \
+			"gpt-5.6-luna" "gpt-5.5" "gpt-5.5-fast" "gpt-5.5-pro" "gpt-5.4" "gpt-5.4-fast" "gpt-5.4-mini" \
+			"gpt-5.4-mini-fast" "gpt-5.3-codex" "gpt-5.3-codex-fast" "gpt-5.2" "garnet-07-15" \
+			"gemini-3.1-pro-preview" "gemini-3.8-flash" "gemini-3.7-flash" "gemini-3.6-flash" \
+			"gemini-3.5-flash" "gemini-3-flash-preview" "grok-4.7" "grok-4.6" "grok-4.5" "atlas-07-21" \
+			"aster-07-15"
+	}
 	// Subscription eligibility for model-aware Codex credential routing.
 	plan-requirement provider="openai-codex" {
 		tier "pro" substring="-spark"
@@ -149,6 +166,9 @@ behavior {
 	// Text-chat transports cannot serve StepFun's audio-only and image-editing
 	// SKUs, which its /v1/models roster interleaves with the chat models.
 	exclude-models provider="stepfun" prefix="stepaudio-" prefix="step-image-" prefix="step-tts-" prefix="step-2x-large"
+	// Helmcode's /v1/models roster lists embedding, rerank, TTS, and STT models
+	// beside the chat models (helmcode.com/docs/models).
+	exclude-models provider="helmcode" exact="qwen3-embedding" exact="rerank" exact="kokoro" exact="whisper"
 	// SuperGrok's roster interleaves media models the chat picker cannot
 	// serve; the dedicated tool surfaces route them instead.
 	exclude-models provider="xai-oauth" exact="grok-imagine-image-quality" exact="grok-imagine-image-2.0" \
@@ -192,14 +212,31 @@ behavior {
 		route "openai-responses" prefix="grok-4." prefix="gpt-5" prefix="gpt-6" prefix="oswe" prefix="mai-"
 	}
 	// Command Code serves Claude ids through its Anthropic-compatible Messages
-	// endpoint, its ten GPT ids (the same set command-code@1.67.0 marks
-	// `spec: "responses"`) through its Responses endpoint, and every other id
-	// through chat completions.
+	// endpoint and everything else through chat completions or Responses
+	// (`supported_endpoints` on /provider/v1/models). Chat completions rejects
+	// `reasoning_effort: "none"` (low|medium|high|xhigh|max only), while
+	// Responses accepts `reasoning.effort: "none"`. Every id with a reviewed
+	// effort ladder that lists `/responses` therefore rides Responses, so
+	// thinking-off actually disables reasoning instead of clamping to the
+	// lowest effort. Ladder ids without `/responses` (deepseek-v4-flash-fast,
+	// gemini-3.7-flash, Qwen3.8-Flash, Qwen3.8-Max-0902, hy4-preview,
+	// space-bunny-alpha), the Qwen3.8 ids (chat already disables them through
+	// `enable_thinking: false`), and non-reasoning ids stay on chat completions.
 	api-routes provider="commandcode" default="openai-completions" {
 		route "anthropic-messages" prefix="claude-"
 		route "openai-responses" exact="gpt-5.3-codex" exact="gpt-5.4" exact="gpt-5.4-mini" exact="gpt-5.5" \
 			exact="gpt-5.6-luna" exact="gpt-5.6-sol" exact="gpt-5.6-terra" exact="gpt-6-astra" \
-			exact="gpt-6-luna" exact="gpt-6-sol"
+			exact="gpt-6-luna" exact="gpt-6-sol" \
+			exact="deepseek/deepseek-v4-flash" exact="deepseek/deepseek-v4-flash-vision-exp" \
+			exact="deepseek/deepseek-v4-pro" exact="deepseek/deepseek-v4.1-flash" \
+			exact="deepseek/deepseek-v4.1-flash-fast" exact="moonshotai/Kimi-K3" \
+			exact="z-ai/glm-5.3-flash" exact="z-ai/glm-5.3-flashx" exact="zai-org/GLM-5.2" \
+			exact="zai-org/GLM-5.3" exact="google/gemini-3.1-flash-lite" exact="google/gemini-3.5-flash" \
+			exact="google/gemini-3.5-flash-lite" exact="google/gemini-3.6-flash" \
+			exact="google/gemini-3.8-flash" exact="MiniMaxAI/MiniMax-M3" exact="stepfun/Step-5-Preview" \
+			exact="xai/grok-4.5" exact="xai/grok-4.6" exact="xai/grok-4.7" exact="meta/muse-spark-1.1" \
+			exact="meta/muse-spark-1.2" exact="meta/muse-spark-1.2-contributor" exact="meta/muse-spark-1.3" \
+			exact="meta/muse-spark-1.3-contributor" exact="stealth/pixel-canary" exact="sakana/fugu-ultra"
 	}
 	// ZenMux fronts Anthropic ids on a dedicated Messages endpoint; every
 	// other namespace stays on chat completions.
@@ -232,6 +269,30 @@ behavior {
 		route "openai-completions" exact="minimax-m2.7" exact="minimax-m3" exact="minimax-m3-free" \
 			exact="qwen3.5-plus" exact="qwen3.6-plus"
 	}
+	// Factory Droid's proxy multiplexes four wires behind one client identity
+	// (packages/ai/src/providers/factory-droid.ts); `google-generate` is its
+	// native Gemini client. The registry names each model's wire, so every
+	// roster id routes exactly; unknown ids fall back to chat completions.
+	api-routes provider="factory-droid" default="openai-completions" {
+		route "anthropic-messages" exact="claude-fable-5.1" exact="claude-fable-5" exact="claude-opus-5-5" \
+			exact="claude-opus-5-5-fast" exact="claude-opus-5" exact="claude-opus-5-fast" \
+			exact="claude-opus-4-8" exact="claude-opus-4-8-fast" exact="claude-opus-4-7" \
+			exact="claude-opus-4-6" exact="claude-opus-4-5-20251101" exact="claude-sonnet-5-5" \
+			exact="claude-sonnet-5" exact="claude-sonnet-4-6" exact="claude-sonnet-4-5-20250929" \
+			exact="claude-haiku-4-5-20251001" exact="atlas-07-21" exact="aster-07-15" exact="minimax-m2.7"
+		route "openai-responses" exact="gpt-6-astra" exact="gpt-6.1-sol" exact="gpt-6-sol" exact="gpt-6-luna" \
+			exact="gpt-5.6-sol" exact="gpt-5.6-sol-fast" exact="gpt-5.6-terra" exact="gpt-5.6-luna" \
+			exact="gpt-5.5" exact="gpt-5.5-fast" exact="gpt-5.5-pro" exact="gpt-5.4" exact="gpt-5.4-fast" \
+			exact="gpt-5.4-mini" exact="gpt-5.4-mini-fast" exact="gpt-5.3-codex" exact="gpt-5.3-codex-fast" \
+			exact="gpt-5.2" exact="grok-4.7" exact="grok-4.6" exact="grok-4.5"
+		route "google-generate" exact="garnet-07-15" exact="gemini-3.1-pro-preview" exact="gemini-3.8-flash" \
+			exact="gemini-3.7-flash" exact="gemini-3.6-flash" exact="gemini-3.5-flash" \
+			exact="gemini-3-flash-preview"
+		route "openai-completions" exact="inkling" exact="mistral-medium-3.5" exact="glm-5.3-flash" \
+			exact="glm-5.3" exact="glm-5.2" exact="glm-5.2-fast" exact="kimi-k3" exact="qwen3.8-max" \
+			exact="nemotron-3-ultra" exact="deepseek-v4.1-flash" exact="deepseek-v4-flash-0731" \
+			exact="deepseek-v4-pro" exact="minimax-m3"
+	}
 
 	// Copilot serves these models with tighter effective limits than its
 	// /models metadata reports.

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

@@ -659,27 +659,49 @@ collapse {
 		requires-effort #true
 		default-member "cursor-grok-4.5-medium-fast"
 	}
-	variant-family "cursor" "cursor-grok-4.6" name="Grok 4.6" {
-		members "cursor-grok-4.6-low" "cursor-grok-4.6-medium" "cursor-grok-4.6-high" "cursor-grok-4.6-xhigh"
-		route "low" "cursor-grok-4.6-low"
-		route "medium" "cursor-grok-4.6-medium"
-		route "high" "cursor-grok-4.6-high"
-		route "xhigh" "cursor-grok-4.6-xhigh"
+	variant-family "cursor" "cursor-grok-{rev}" name="Grok {rev}" revision=">=4.6" {
+		members "cursor-grok-{rev}-low" "cursor-grok-{rev}-medium" "cursor-grok-{rev}-high" "cursor-grok-{rev}-xhigh"
+		route "low" "cursor-grok-{rev}-low"
+		route "medium" "cursor-grok-{rev}-medium"
+		route "high" "cursor-grok-{rev}-high"
+		route "xhigh" "cursor-grok-{rev}-xhigh"
 		mode "effort"
 		efforts "low" "medium" "high" "xhigh"
 		requires-effort #true
-		default-member "cursor-grok-4.6-medium"
+		default-member "cursor-grok-{rev}-medium"
 	}
-	variant-family "cursor" "cursor-grok-4.6-fast" name="Grok 4.6 Fast" {
-		members "cursor-grok-4.6-low-fast" "cursor-grok-4.6-medium-fast" "cursor-grok-4.6-high-fast" "cursor-grok-4.6-xhigh-fast"
-		route "low" "cursor-grok-4.6-low-fast"
-		route "medium" "cursor-grok-4.6-medium-fast"
-		route "high" "cursor-grok-4.6-high-fast"
-		route "xhigh" "cursor-grok-4.6-xhigh-fast"
+	variant-family "cursor" "cursor-grok-{rev}-fast" name="Grok {rev} Fast" revision=">=4.6" {
+		members "cursor-grok-{rev}-low-fast" "cursor-grok-{rev}-medium-fast" "cursor-grok-{rev}-high-fast" "cursor-grok-{rev}-xhigh-fast"
+		route "low" "cursor-grok-{rev}-low-fast"
+		route "medium" "cursor-grok-{rev}-medium-fast"
+		route "high" "cursor-grok-{rev}-high-fast"
+		route "xhigh" "cursor-grok-{rev}-xhigh-fast"
 		mode "effort"
 		efforts "low" "medium" "high" "xhigh"
 		requires-effort #true
-		default-member "cursor-grok-4.6-medium-fast"
+		default-member "cursor-grok-{rev}-medium-fast"
+	}
+	variant-family "cursor" "grok-{rev}" name="Grok {rev}" revision=">=4.7" {
+		members "grok-{rev}-low" "grok-{rev}-medium" "grok-{rev}-high" "grok-{rev}-xhigh"
+		route "low" "grok-{rev}-low"
+		route "medium" "grok-{rev}-medium"
+		route "high" "grok-{rev}-high"
+		route "xhigh" "grok-{rev}-xhigh"
+		mode "effort"
+		efforts "low" "medium" "high" "xhigh"
+		requires-effort #true
+		default-member "grok-{rev}-medium"
+	}
+	variant-family "cursor" "grok-{rev}-fast" name="Grok {rev} Fast" revision=">=4.7" {
+		members "grok-{rev}-low-fast" "grok-{rev}-medium-fast" "grok-{rev}-high-fast" "grok-{rev}-xhigh-fast"
+		route "low" "grok-{rev}-low-fast"
+		route "medium" "grok-{rev}-medium-fast"
+		route "high" "grok-{rev}-high-fast"
+		route "xhigh" "grok-{rev}-xhigh-fast"
+		mode "effort"
+		efforts "low" "medium" "high" "xhigh"
+		requires-effort #true
+		default-member "grok-{rev}-medium-fast"
 	}
 	variant-family "cursor" "gpt-5.6-luna" name="GPT-5.6 Luna" {
 		members "gpt-5.6-luna-none" "gpt-5.6-luna-low" "gpt-5.6-luna-medium" "gpt-5.6-luna-high" "gpt-5.6-luna-xhigh" "gpt-5.6-luna-max"