Optimize feature loop: reduce token usage and improve speed

[federiconeri]

Overview

The feature loop runs 13+ claude -p sessions per feature. Each session starts fresh, re-sends full prompt templates, and causes Claude to re-read the same @file references. This results in high cached token volume, unnecessary wall-clock time, and elevated API costs.

Current Baseline (per feature loop)

Metric	Value
Total duration	~30-40 min
Claude sessions	~13
Total cost	~$12.40
Cached token volume	~3.1M (73% of total)

Expected After Optimization

Metric	Value	Improvement
Total duration	~24-32 min	~6-8 min faster (20%)
Claude sessions	~8-9	4-5 fewer sessions
Total cost	~$8-9	~$3.50-4.50 saved (28-36%)
Cached token volume	~2.0-2.3M	~30-35% reduction

Sub-issues (ordered by impact)

P1 — High Impact

• Use --resume for implementation loop iterations 2+ #137
• Merge verification phase into review phase #139

P2 — Medium Impact

• Use --resume for E2E and review loop iterations #138
• Create targeted E2E fix prompt instead of using full PROMPT.md #140

P3 — Lower Impact (prompt trimming & small fixes)

• Trim PROMPT_e2e.md.tmpl — remove reference material Claude already knows #141
• Trim PROMPT_feature.md.tmpl — remove 80-line example format #142
• Trim PROMPT.md.tmpl and review prompts — condense boilerplate #143
• Fix loop inefficiencies: double test run, sleep, config loading #144

Savings Breakdown

Optimization	Cost Saved	Time Saved
--resume for implementation loop	~$2.00	~1.5 min
Merge verification into review	~$1.80	~1.5 min
--resume for E2E/review loops	~$0.80	~1 min
Targeted E2E fix prompt	~$0.50	~45s
Trim PROMPT_e2e.md	~$0.10	—
Trim PROMPT_feature.md	~$0.10	—
Trim PROMPT.md + review prompts	~$0.20	—
Loop inefficiencies (sleep, config, tests)	—	~30s
Total	~$3.50-4.50	~6-8 min

Over 10 features: ~$35-45 saved and ~1 hour less wall-clock time.