Session-Miser — Session-Level Cost Optimization
Core Principle
Use the cheapest model that can do the job well. Opus for thinking, Sonnet for doing, Haiku for fetching.
For per-call routing, pricing data, and API-level optimization, see the token-miser skill.
This skill covers session-level decisions only: which model to start on, when to switch, and what to delegate.
Planning Phase Exemption
During the planning phase (before ExitPlanMode), use Opus for all subagents:
- project-management planning — Explore/Plan agents building the work queue
- /plan mode — implementation design
Once ExitPlanMode is called, normal routing resumes.
Session Start Assessment
At session start, assess the work and recommend:
| Work Profile | Recommendation |
|---|---|
| Architecture, design, complex planning | Stay on Opus |
| Debugging root cause analysis | Stay on Opus |
| Mixed planning + execution | Stay on Opus, DELEGATE mechanical tasks to subagents |
| Primarily code edits, file changes, deploys | Suggest /model sonnet |
| Bulk operations, config changes, formatting | Suggest /model sonnet |
When to Switch Models Mid-Session
| Signal | Action |
|---|---|
| Planning phase complete, entering execution | Suggest /model sonnet |
| Hit unexpected architectural decision | Switch back to Opus for that decision |
| 30+ minutes of file edits on Opus | Suggest /model sonnet |
| Debugging shifts from "apply known fix" to "investigate root cause" | Switch to Opus |
| Cost climbing fast with no complex reasoning | Suggest delegation or model switch |
Delegation Rules
ALWAYS delegate (automatic, don't ask):
- File exploration → Explore agent (haiku)
- Git status, diff, log → haiku subagent
- Searching codebase → haiku subagent
- Running bash for status checks → haiku subagent
SUGGEST delegating (mention to user):
- Bulk file edits across repos → sonnet subagent
- Running test suites → haiku subagent
- Documentation generation → sonnet subagent
- Config file updates → sonnet subagent
NEVER delegate:
- Active user discussion topics
- Security-sensitive operations
- Complex multi-step reasoning chains
Effort Level Guide
| Effort | Use For | Relative Cost |
|---|---|---|
| high | Complex problems, architecture, security | 1x |
| medium | Most coding, debugging, research | ~0.7x |
| low | Simple edits, formatting, git ops | ~0.4x |
Prefer opus:low for simple tasks over switching to Sonnet entirely — preserves Opus quality for unexpected complexity.
Cost Tracking
Check session cost via hydra-pulse / session-snitch:
# If hydra-pulse CLI available
hydra-pulse stats
If cost climbing fast: delegate more, switch model, or reduce effort level.
Anti-Patterns
- Using Opus to read a file and report contents (haiku job)
- Using Opus for
git status(haiku job) - Using Opus for JSON/YAML formatting (sonnet:low)
- Staying on Opus for 30+ minutes of mechanical edits
- Switching to Haiku for anything requiring judgment