Cost Efficiency Analysis — SOLAR LINE Project

Status: RESOLVED

Key recommendations incorporated into CLAUDE.md and project practices:

• TodoWrite discipline (state transitions only)
• Background long commands (run_in_background)
• Sonnet default for subagents (replacing Haiku)
• max_turns on bounded Task calls
• Dedicated tools over Bash in subagents
• Skills defined for repeated workflows (5 skills: episode-analysis, report-review, subtitle-extraction, cost-analysis, nice-friend)

Overview

Analysis of Claude Code token usage patterns across the SOLAR LINE project, based on

ccusage session data from the autonomous agent loop VM.

Key Metrics (Sample: 1 day, 58 sessions, 120 subagents)

Token Distribution

• 97.3% cache reads — excellent cache utilization
• 2.7% cache creation — context setup overhead
• <0.1% actual I/O — the "useful work" is a tiny fraction of total tokens

Cost Drivers

1. Subagent spawning (100% of measured cost)

All $6.57 comes from Haiku subagent calls. Each subagent:

• Gets a fresh context with the project's CLAUDE.md, MEMORY.md, and system prompt
• Averages ~836K tokens (mostly cache reads)
• Makes ~16 tool calls
• Most (88%) are read-only exploration

2. TodoWrite frequency (9.8% of tool calls)

459 TodoWrite calls across 58 sessions (7.9 per session average). Each call includes

the full todo list in both the request and response, consuming context on every turn.

3. Long-running Bash streams

The largest session was 16x the average size due to bash_progress streaming from

Whisper/yt-dlp. These progress lines are stored in the JSONL and inflate context.

4. Subagent model selection

• Haiku subagents: 79% of invocations, 72% of tokens, ~$0.05/invocation avg
• Sonnet subagents: 11% of invocations, 21% of tokens, ~$0.02/invocation but more tokens
• Opus subagents: 10% of invocations, 7% of tokens, $0/invocation (subscription)

Efficiency Recommendations

High Impact

1. Reduce TodoWrite frequency: Update only on state transitions (start/complete), not

between every tool call. Could save ~200 context turns per day.

1. Background long Bash commands: Use run_in_background for yt-dlp, Whisper, and

other long-running processes. Check output later with TaskOutput.

Medium Impact

1. Prefer Haiku for exploration subagents: Most subagent tasks are simple file reading.

Haiku is sufficient and cheaper.

1. Prefer Read/Grep over Bash in subagents: 43% of subagent tool calls are Bash

(often doing cat/grep/find). The dedicated tools are more structured and efficient.

1. Limit subagent scope: Some subagents run 95+ tool calls (9+ minutes). Set

max_turns on Task calls to cap exploration.

Low Impact, High Value

1. Skill-ize repeated workflows: Create Claude Code Skills for:

- Episode analysis pipeline

- Report review checklist

- Subtitle processing (download + transcribe + extract)

1. Trim MEMORY.md: Keep under 200 lines. Move detailed content to topic files.

Currently truncated at 200 lines, wasting some cache creation tokens.

Cost Estimation Framework

For the full project (~65 tasks, ~94 commits), assuming similar patterns:

• ~58 sessions per VM boot × N boots
• ~120 subagents per boot
• ~$6.50 Haiku cost per boot
• Opus cost: $0 (Max subscription)

The project's cost efficiency is generally excellent (97.3% cache hit rate).

The main optimization opportunities are in reducing unnecessary context churn

(TodoWrite) and avoiding inflated sessions from streaming output.

Running the Analysis

# In ts/ directory:
# Option 1: Pipe from ccusage
bunx ccusage@17 session --offline --json | npm run analyze-costs -- --mode session

# Option 2: Save and analyze
bunx ccusage@17 session --offline --json > costs.json
npm run analyze-costs -- --mode session --input costs.json

# Option 3: Daily summary
bunx ccusage@17 daily --offline --json | npm run analyze-costs -- --mode daily