name: cost-dashboard
description: Analyze model routing efficiency, task distribution, and performance from the SuperPAI cost logger. Shows how effectively the CCR routes tasks to the right model tier.
triggers:
- "cost"
- "spending"
- "token usage"
- "how much"
- "cost dashboard"
- "cost report"
- "cost summary"
- "routing efficiency"
- "model distribution"

Routing & Efficiency Dashboard

Analyze how effectively SuperPAI routes tasks to the optimal model tier.

Why This Matters

For Pro Max subscribers (unlimited usage), the value isn't dollars saved — it's:
- Speed: Haiku responds ~10x faster than opus. Routing simple tasks to haiku means faster iteration.
- Routing accuracy: The right model for the right task. Opus on a typo fix is slow. Haiku on architecture design is under-capable.
- Throughput: More tasks completed per hour when simple work isn't waiting for opus-level processing.

For API users (pay-per-token), the value is direct cost savings:
- Haiku: ~$0.25/MTok input vs Opus: ~$15/MTok input (60x difference)
- Proper routing can reduce monthly API costs by 70-80%

Data Source

Cost data is stored in logs/cost.jsonl at the plugin root. Each line is a JSON object with:
- ts: ISO timestamp
- session: Session designation (WIN1, WSL1, etc.)
- tool: Tool name used
- model: Model tier used (haiku, sonnet, opus) — written by the CCR hook
- type: Entry type (tool_use, session_start, session_end)

For richer data, use the superpai_cost_summary MCP tool which queries the database.

How to Analyze

1. Read the log file: Read logs/cost.jsonl from the plugin root
2. Parse JSONL: Each line is one JSON entry
3. Key metrics to calculate:
4. Model distribution: % of tasks routed to haiku / sonnet / opus
5. Routing efficiency score: Tasks where model matched expected complexity (aim for >80%)
6. Speed index: Estimated time saved by routing simple tasks to haiku
7. Session comparison: Which sessions use more opus (complex work) vs haiku (maintenance)
8. Tool frequency: Which tools are called most (Read, Edit, Bash, Agent)

Output Format

Present results as a clear summary:

SuperPAI Routing Report
=======================
Period: {start_date} to {end_date}
Total agent dispatches: {count}

Model Distribution:
  haiku:  {count} ({pct}%) — simple tasks, lookups, fixes
  sonnet: {count} ({pct}%) — features, bugs, refactoring
  opus:   {count} ({pct}%) — architecture, research, security

Routing Efficiency:
  Estimated correct routes: {pct}%
  Speed benefit: ~{X}x faster avg response for {haiku_count} haiku-routed tasks

By Session:
  WIN1: {count} dispatches (haiku:{h}% sonnet:{s}% opus:{o}%)
  WSL1: {count} dispatches (haiku:{h}% sonnet:{s}% opus:{o}%)

Top Tools:
  Agent: {count} (model routing active)
  Read:  {count}
  Edit:  {count}
  Bash:  {count}

Daily Trend:
  {date}: {count} dispatches — h:{haiku} s:{sonnet} o:{opus}
  {date}: {count} dispatches — h:{haiku} s:{sonnet} o:{opus}

If the user is on the API (not Pro Max), also show estimated cost:

Estimated API Cost (if applicable):
  With CCR routing:    ~${routed_cost}
  Without (all opus):  ~${all_opus_cost}
  Estimated savings:   ~${savings} ({pct}%)

CCR Weight Configuration

Users can adjust the CCR signal weights via superpai_settings:
- SUPERPAI_CCR_WEIGHT_KEYWORDS (default: 40) — keyword pattern matching
- SUPERPAI_CCR_WEIGHT_LENGTH (default: 15) — prompt word count
- SUPERPAI_CCR_WEIGHT_COMPLEXITY (default: 20) — multi-file/single-file scope
- SUPERPAI_CCR_WEIGHT_DEPTH (default: 25) — adaptive depth mode integration

Notes

• Token counts are not available from Claude Code hooks — model distribution is the primary metric
• For actual dollar costs on API plans, cross-reference with Anthropic's usage dashboard
• The CCR logs its routing decision (model + scores) to additionalContext on every Agent dispatch
• Use "cheap mode" / "quality mode" to override automatic routing