luv-model-routing

name	luv-model-routing
description	Profile-based model routing for Luv Marketing agents. Load when making programmatic OpenRouter calls for text generation — reads the active profile (quality/budget/speed), resolves the right model per task type, and provides the curl call pattern.
allowed-tools	Read, Bash

Luv Model Routing

Profile-based model selection for the Luv Marketing Suite. Task type → profile → model → OpenRouter call.

Profile Resolution Order

1. $LUV_PROFILE env var (session override)
2. .luv/active-profile file in project root (project-level default)
3. ~/.cks/profiles/luv/active (global Hermes tier)
4. Fallback: quality

get_luv_profile() {
  local p="${LUV_PROFILE:-}"
  [ -z "$p" ] && p=$(cat .luv/active-profile 2>/dev/null)
  [ -z "$p" ] && p=$(cat "${HOME}/.cks/profiles/luv/active" 2>/dev/null)
  echo "${p:-quality}"
}

get_luv_model() {
  local task_type="$1"
  local profile=$(get_luv_profile)
  local pfile="${CLAUDE_PLUGIN_ROOT}/skills/luv-model-routing/profiles/${profile}.yaml"
  # Project override takes precedence over plugin default
  [ -f ".luv/profiles/${profile}.yaml" ] && pfile=".luv/profiles/${profile}.yaml"
  # Global Hermes-tier override
  [ -f "${HOME}/.cks/profiles/luv/${profile}.yaml" ] && pfile="${HOME}/.cks/profiles/luv/${profile}.yaml"
  grep "^${task_type}:" "$pfile" | awk '{print $2}'
}

Task Type → Profile Key

Luv Agent	Profile key	Meaning
luv:brand-strategist	strategy	Deep positioning and reasoning
luv:ads-copywriter	copywriting	Short-form ad copy
luv:alan-sharpe	copywriting	B2B direct response copy
luv:long-form-copywriter	long_form	Blog posts, whitepapers
Quick variations, A/B tests	fast_copy	High-volume iterations
Competitive analysis, research	analysis	Research and synthesis

OpenRouter Call Pattern (Deterministic)

luv_or_generate() {
  local model="$1"
  local system_prompt="$2"
  local user_prompt="$3"

  local payload
  payload=$(python3 -c "
import json, sys
print(json.dumps({
  'model': sys.argv[1],
  'messages': [
    {'role': 'system', 'content': sys.argv[2]},
    {'role': 'user',   'content': sys.argv[3]}
  ]
}))" "$model" "$system_prompt" "$user_prompt")

  curl -s -X POST https://openrouter.ai/api/v1/chat/completions \
    -H "Authorization: Bearer $OPENROUTER_API_KEY" \
    -H "Content-Type: application/json" \
    -H "HTTP-Referer: https://github.com/cardinalconseils/claude-starter" \
    -H "X-Title: Luv Marketing Suite" \
    -d "$payload" \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(d['choices'][0]['message']['content'])"
}

Usage in an agent:

MODEL=$(get_luv_model "copywriting")
RESULT=$(luv_or_generate "$MODEL" "You are a VoC copywriter..." "Write 5 Google Ads headlines for...")

Profiles at a Glance

Profile	Strategy	Copy	Long-form	Fast	Analysis
quality	claude-opus-4-7	claude-sonnet-4-6	claude-sonnet-4-6	claude-haiku-4-5	claude-opus-4-7
budget	claude-sonnet-4-6	gpt-4o-mini	gpt-4o-mini	gemini-flash-2.0	claude-sonnet-4-6
speed	gemini-2.5-flash	gpt-4o-mini	gpt-4o-mini	gpt-4o-mini	gemini-2.5-flash

Image (openai/gpt-image-1) and video (kling/kling-v1-5) are direct-API, not OpenRouter — all profiles use the same values.

Switching Profiles

# Session override
export LUV_PROFILE=budget

# Project default
mkdir -p .luv && echo "budget" > .luv/active-profile

# Global default (Hermes tier)
mkdir -p ~/.cks/profiles/luv && echo "quality" > ~/.cks/profiles/luv/active

Or use /cks:luv-profile [name] to switch from within a session.

When Luv Agents Call OpenRouter

Luv text agents ARE Claude — they generate copy using their own intelligence by default. Use OpenRouter programmatically only when:

• Batch generation: producing 10+ variants in a single run (efficiency)
• Non-Claude model specified: profile routes to GPT-4o-mini, Gemini, etc.
• User explicitly requests a specific model: "use the budget profile for this one"
• Speed profile + rapid iteration: workshop or live client session

For single-task generation (the common case), the agent uses its own capabilities and the profile just informs which model tier is appropriate.

Common Rationalizations

Rationalization	Reality
"Just use quality for everything"	For 20 ad variations at 2am, Haiku is 20× cheaper and fast enough. Profile routing exists for this.
"Profile reading is complex"	3 grep lines. The resolution order is deterministic.
"Speed profile for final deliverables"	Speed = drafts and workshops only. Final copy → quality.
"Direct OpenAI for images is inconsistent"	Images are not on OpenRouter. Direct is correct. Profile documents this explicitly.