// 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" \
		"together" "wafer-serverless" "zenmux" {
		family "k2.6" {
			stream-idle-timeout-ms 300000
		}
	}
	// Replaces Moonshot-native K3 and K2.7 Code stream-idle branches.
	on "kimi-code" "moonshot" {
		family "k3" {
			stream-idle-timeout-ms 300000
		}
		family "k2.7-code" {
			stream-idle-timeout-ms 300000
		}
	}
	on "aiand" "baseten" "coreweave" "fireworks" "huggingface" "kilo" "nanogpt" "novita" \
		"opencode-go" "opencode-zen" "together" "vercel-ai-gateway" "zenmux" {
		family "k2.7-code" {
			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
		}
	}
	family "k3" {
		thinking-default-level "max"
		thinking-efforts "low" "high" "max"
		thinking-requires-effort #true
		// K3 drives reasoning through OpenAI-style reasoning_effort with a
		// wire low/high/max scale on every chat-completions host; the remap
		// is a lineage truth, not a deployment contract.
		reasoning-effort-map {
			minimal "low"
			medium "high"
			xhigh "max"
			max "max"
		}
	}
	on "alibaba-coding-plan" "cursor" "kimi-code" "moonshot" "ollama-cloud" "openrouter" {
		family "k2.5" {
			thinking-efforts "minimal" "low" "medium" "high"
		}
	}
	on "alibaba-coding-plan" "github-copilot" "wafer-serverless" {
		supports-developer-role #false
	}
	on "baseten" "cloudflare-ai-gateway" "coreweave" "fireworks" "huggingface" "kilo" \
		"nanogpt" "novita" "nvidia" "opencode-go" "opencode-zen" "together" "vercel-ai-gateway" "zenmux" {
		family "k2.6" {
			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
		}
	}
	on "cloudflare-ai-gateway" "coreweave" "fireworks" "huggingface" "kilo" "nanogpt" \
		"novita" "nvidia" "opencode-go" "opencode-zen" "together" "vercel-ai-gateway" \
		"zenmux" {
		family "k2.5" {
			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
		}
	}
	on "cursor" "github-copilot" "moonshot" "ollama-cloud" "openrouter" "wafer-serverless" {
		family "k2.7-code" {
			thinking-efforts "minimal" "low" "medium" "high"
		}
	}
	on "huggingface" "moonshot" "nanogpt" "novita" "nvidia" "ollama-cloud" "zenmux" {
		family "k2" {
			thinking-requires-effort #true
		}
	}
	on "huggingface" "nanogpt" "novita" "nvidia" "opencode-zen" "venice" "vercel-ai-gateway" "zenmux" {
		family "k2" {
			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
		}
	}
	on "kilo" "moonshot" "openrouter" "vercel-ai-gateway" "zenmux" {
		family "k2-thinking" {
			thinking-requires-effort #true
		}
	}
	on "kilo" "vercel-ai-gateway" "zenmux" {
		family "k2-thinking" {
			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
		}
	}
	on "moonshot" "ollama-cloud" {
		family "k2" {
			thinking-efforts "minimal" "low" "medium" "high"
		}
	}
	on "moonshot" "ollama-cloud" "openrouter" "wafer-serverless" {
		family "k2.6" {
			thinking-efforts "minimal" "low" "medium" "high"
		}
	}
	on "moonshot" "openrouter" {
		family "k2-thinking" {
			thinking-efforts "minimal" "low" "medium" "high"
		}
	}
	// K2 chat templates can leak native section/tool-call tokens on hosted and
	// local OpenAI-compatible transports.
	models "*kimi*k2*" {
		stream-markup-healing-pattern "kimi"
	}
	on "kimi-code" "moonshot" {
		family "k3" {
			native-kimi-k3-reasoning #true
			clamp-output-to-model-max #true
		}
		// Native K3 accepts forced tools on the Anthropic surface too;
		// omitting thinking leaves its mandatory provider default intact.
		on-api "anthropic-messages" {
			family "k3" {
				supports-forced-tool-choice #true
			}
		}
	}
	// K3 reasoning reaches omp as its plaintext trace on every host, and K3 was
	// trained to read all of its earlier reasoning back (preserved thinking).
	// A recall probe showed the reasoning slot of moonshot, factory-droid,
	// openrouter and devin reaches the model, so K3 reasoning moves between
	// those hosts natively instead of as visible `<think>` text.
	family "k3" {
		portable-reasoning #true
	}
}