// 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"
		}
	}
}