executus/model/lane_mapping.go

// Package llms — lane_mapping.go: maps a model spec to a stable lane
// name. Pure data + a single function; no dependency on the registry,
// no provider wrapping. Kept separate from lane_transport.go so the
// mapping table can be committed and reviewed in isolation, and so
// admin / webui code that just wants to *display* lane assignments
// doesn't drag in the transport machinery.
//
// Why a fixed table: provider concurrency caps differ — Ollama Pro is
// 3 connections, Anthropic Claude has higher per-tier limits, etc.
// Each provider gets its own lane name so they can be configured
// independently via convars (lanes.<name>.max_concurrent). Lane names
// are user-facing (admin dashboard + convar key suffixes) and need to
// stay stable across deploys; an env-overridable map adds complexity
// for no current benefit.
//
// Test: lane_transport_test.go covers TestLaneFor_Mapping.
package model

import "strings"

// Lane name constants. Defined as exported strings so admin code (.skill
// admin set-lane <skill> <lane>), webui dropdowns, and convar consumers
// share a single canonical spelling.
const (
	// LaneOllama covers ollama-cloud/* (and any future ollama/* local).
	// The local ollama instance is on the same physical resource as
	// the cloud account from mort's perspective — the connection cap
	// should apply jointly.
	LaneOllama = "ollama"

	// LaneAnthropicThinking is the lane for Anthropic models in
	// extended-thinking mode. Separated from default because thinking
	// requests hold connections longer and can starve faster lanes
	// when multiplexed.
	LaneAnthropicThinking = "anthropic-thinking"

	// LaneAnthropicDefault is the lane for non-thinking Anthropic
	// requests (haiku, sonnet, opus without -thinking-).
	LaneAnthropicDefault = "anthropic-default"

	// LaneM1 is the lane for m1/* models (foreman-style router
	// pointing at a dedicated local instance). Separated from the
	// ollama lane because m1 targets a distinct host with its own
	// connection budget.
	LaneM1 = "m1"

	// LaneLLMDefault is the catch-all lane for any provider/model
	// combination not explicitly mapped above.
	LaneLLMDefault = "llm-default"
)

// LaneFor returns the lane name for the given model spec. Mapping:
//
//	ollama-cloud/*               → "ollama"          (Pro account: 3 connections)
//	anthropic/*-thinking-*       → "anthropic-thinking"
//	anthropic/*                  → "anthropic-default"
//	(anything else)              → "llm-default"
//
// Tier aliases (fast/standard/thinking) flow through this function as
// the resolver's expanded provider/model spec, so callers don't need
// to think about tier indirection. Empty input falls through to
// LaneLLMDefault rather than panicking — defensive against unset
// model specs in edge-case test wiring.
//
// Substring match for "-thinking-" keeps future Anthropic naming
// variations classified correctly without churning this table on
// every model release.
func LaneFor(modelSpec string) string {
	s := strings.TrimSpace(modelSpec)
	if strings.HasPrefix(s, "ollama-cloud/") {
		return LaneOllama
	}
	if strings.HasPrefix(s, "anthropic/") {
		if strings.Contains(s, "-thinking-") {
			return LaneAnthropicThinking
		}
		return LaneAnthropicDefault
	}
	// Foreman instances are backed by Ollama and share its connection
	// cap, so they route to the same lane.
	if strings.HasPrefix(s, "foreman/") {
		return LaneOllama
	}
	// m1/ is a foreman-style router pointing at a dedicated local
	// instance with its own connection budget. Separate lane so its
	// concurrency cap is independent of the shared ollama lane.
	if strings.HasPrefix(s, "m1/") {
		return LaneM1
	}
	return LaneLLMDefault
}