Session miser

Session-Miser — Intelligent Session Cost Optimization

Core Principle

Use the cheapest model that can do the job well. Opus for thinking, Sonnet for doing, Haiku for fetching.

Model Selection Matrix

Task Type	Model	Effort	Why
Architecture, design, complex reasoning	Opus	high	Needs deep reasoning
Security audits, code review with judgment	Opus	medium	Quality-sensitive
Code generation (new features, complex logic)	Opus or Sonnet	medium	Sonnet handles most code well
Code generation (boilerplate, simple functions)	Sonnet	low	Templated, predictable
Text editing, formatting, refactoring	Sonnet	low	Mechanical transformation
Git operations (commit, push, branch, merge)	Sonnet	low	Deterministic commands
File search, grep, glob, reading files	Haiku (subagent)	—	Read-only, no reasoning
Config file edits (.yaml, .json, .toml)	Sonnet	low	Structural, not creative
Documentation writing	Sonnet	medium	Needs coherence, not deep reasoning
Debugging (root cause analysis)	Opus	high	Multi-step reasoning
Debugging (applying a known fix)	Sonnet	low	Just follow the plan
Research (web search, GitHub crawl)	Sonnet	medium	Synthesis, not deep reasoning
Planning (PRD, architecture)	Opus	high	Open-ended judgment
Running tests, checking status	Haiku (subagent)	—	Mechanical execution

Session Start Recommendation

At the start of every session, assess the likely work and recommend:

You're on Opus. Based on the project and recent activity:
- If primarily editing/committing/deploying → suggest /model sonnet
- If primarily debugging/designing/architecting → stay on Opus
- If mixing both → stay on Opus but DELEGATE mechanical tasks to subagents

Delegation Rules

ALWAYS delegate these to subagents (automatic, don't ask):

• File exploration (Explore agent → haiku)
• Running bash commands for status checks → haiku
• Git status, diff, log → haiku
• Searching codebase → haiku

SUGGEST delegating these (mention to user):

• Bulk file edits across multiple repos → sonnet subagent
• Running test suites → haiku subagent
• Documentation generation → sonnet subagent
• Config file updates → sonnet subagent

NEVER delegate (must stay in main thread):

• Anything the user is actively discussing
• Decisions that need user confirmation
• Security-sensitive operations
• Complex multi-step reasoning chains

Effort Level Guide

Claude Code supports effort levels: opus:low, opus:medium, opus:high (and sonnet:low/medium/high)

Effort	Use For	Cost
high	First attempt at complex problems, architecture, security	1x
medium	Most coding, debugging, research	~0.7x
low	Simple edits, formatting, git ops, status checks	~0.4x

Recommend /model opus:low for simple tasks instead of switching to Sonnet entirely. This preserves Opus quality for unexpected complexity while saving cost.

Integration with Multi-Instance Coordination

When coordinating work across multiple nodes:

• Route GPU-intensive tasks to GPU hosts (create coordination task with requires: [gpu, ollama])
• Route mechanical fleet tasks to primary/CPU hosts (lower cost hosts)
• When creating coordination tasks, specify the model tier in the task description: "This is a sonnet-tier task: bulk edit config files across 5 repos"
• This helps the claiming instance pick the right model

Cost Tracking

After significant work blocks, check session-snitch:

session-snitch stats       # session cost so far
session-snitch cost        # cost breakdown by model

If session cost is climbing fast, suggest:

1. Delegate more to subagents
2. Switch to /model sonnet for the current phase
3. Use /model opus:low instead of opus:high

Anti-Patterns

DON'T:

• Use Opus to read a file and report its contents (that's a haiku job)
• Use Opus to run git status and parse the output (haiku)
• Use Opus to format JSON or YAML (sonnet:low)
• Stay on Opus when doing 30 minutes of file edits (switch to sonnet)

DO:

• Start on Opus for planning, switch to Sonnet for execution
• Delegate parallel searches to haiku subagents
• Come back to Opus for review/verification of Sonnet's work
• Use effort levels to fine-tune cost within a model tier