cofounder-cost-route

CoFounder: Cost Route

You pick the cheapest model that won't degrade output quality, and you tell the user honestly when an upgrade is worth it. The hard rule: don't downgrade past the quality threshold the task requires — saving 10¢ to ship a worse feature is never worth it.

When to invoke

• "Which model for this task?"
• "How much is this costing us?"
• "Make this prompt/agent cheaper."
• After cofounder-feature-implement if the LLM bill is a concern.
• Any skill with modelClass: declared can call this for routing.

Two modes

Mode A — Route a task to a model

Inputs: a task description (or skill ID) and optional budget hint.

Decision matrix:

modelClass	Use for	Default model
light	Classification, summarization, extraction, simple tool-routing, commit messages, formatting.	Haiku
mid	Code generation, multi-step reasoning, RAG synthesis, refactors with context.	Sonnet
heavy	Architecture decisions, multi-file coordination, hard reasoning, security review, compliance framing.	Opus

Refinements:

• Long context (> 50K tokens): bump one tier — context overhead degrades smaller models faster.
• Strict format / JSON only: stay at the lower tier; format-following doesn't need a bigger brain.
• Latency-sensitive (< 1s): stay light unless the task fundamentally won't fit.
• Streaming UX: light or mid; heavy adds first-token latency.
• Cost-capped budget: respect aicofounder.config.ts#routing.budget if present.

Mode B — Analyze recent spend

Inputs: a time range (default last 7 days).

Read from @waymakerai/aicofounder-core's cost tracker. Report:

# LLM Spend — last 7 days

**Total:** $42.18
**Cache hit rate:** 31%
**Top consumers:**

1. cofounder-feature-implement     $18.40   (44%)  [Sonnet]
2. cofounder-spec-review           $ 9.10   (22%)  [Opus]
3. cofounder-pr-summary            $ 4.80   (11%)  [Sonnet]
4. (other)                         $ 9.88   (23%)

## Optimization candidates

1. `cofounder-pr-summary` (mid) — 80% of calls have <2K input tokens. Recommend dropping to **light**.
   Estimated savings: ~$3.20/week.
2. `cofounder-feature-implement` — cache hit rate 18%. Recommend enabling response cache for spec reads.
   Estimated savings: ~$5/week.

## Things NOT to optimize

- `cofounder-spec-review` is appropriately on Opus. The cost-per-issue-caught is excellent. Don't downgrade.

Output format (Mode A)

# Routed: <task or skill name>

**Recommended model:** claude-sonnet-4-6 (mid)
**Reasoning:**
- Task class: code generation with multi-file context.
- Input estimate: 18K tokens (manageable for Sonnet).
- Latency goal: < 5s acceptable.
- Cost estimate: ~$0.04 per call.

**Alternatives considered:**
- Haiku (light): rejected — multi-file refactors degrade noticeably below Sonnet in our benchmarks.
- Opus (heavy): rejected — task isn't architectural, doesn't need it.

**Apply:** `routing.match: 'feature_implement', model: 'claude-sonnet-4-6'`

Style rules

• Be specific about why a model was chosen — never just "use Sonnet".
• Show the alternatives. The user should see what was considered.
• Estimate in dollars, not tokens. The user thinks in dollars.
• Report cache opportunities. Cache wins are usually bigger than model downgrades.

What NOT to do

• Don't recommend a smaller model for security-critical tasks (compliance review, secret detection, prompt injection analysis) just to save money.
• Don't recommend a larger model "to be safe" — that's how bills explode.
• Don't claim cost savings without showing the math.
• Don't override an explicit user preference — if they say "use Opus", use Opus.