Agent: loop token consumption too high — 9.6M effective input for simple task

[federiconeri]

Summary

A simple scaffolding task (issue-1-scaffolding on l2l project) consumed 9.6M effective input tokens across 3 sessions and 180 API turns. The primary driver is the high number of API turns, each re-reading ~50K tokens of cached system prompt.

Problem / Context

Token breakdown from the second agent test:

• cache_read: 9,160,663 tokens (95.2%)
• cache_create: 461,016 tokens (4.8%)
• direct_input: 204 tokens (0.0%)
• Total: 9,621,883 effective input tokens
• Output: 30,371 tokens

180 API turns across 3 sessions:

• Session 1 (planning): 43 turns — ENTIRELY WASTED (planned already-completed work)
• Session 2 (verification): 87 turns — ~50% waste from sequential single-tool operations
• Session 3 (pr-review): 50 turns — moderate waste

Average context per turn: ~53K tokens (mostly Claude Code system prompt + tool definitions).

Analysis

Session 1 waste (43 turns saved entirely):

• Planning phase produced a full implementation plan for work already done on feat/config-module
• Pre-check git diff main..HEAD -- src/ tests/ would have revealed existing code
• Savings: 43 turns × 53K = ~2.3M tokens

Session 2 waste (~57 turns saveable):

• Turns 6-9: 4 sequential Glob calls → could be 1 parallel turn
• Turns 10-13: 4 sequential test Bash calls → could be 1 batched turn
• Turns 16-21: 6 sequential git commands → could be 2 turns
• Turns 29-37: 8 sequential git show calls → could be 1-2 turns
• Turns 48-65: 9 sequential Edit calls for spec checkboxes → could batch
• Savings: ~57 turns × 53K = ~3M tokens

Session 3 waste (~20 turns saveable):

• Re-verifies work already verified in session 2
• Failed reads of non-existent files (AGENTS.md)

Proposed Solution

P0: Pre-implementation detection in feature-loop.sh

# Before planning, check if implementation exists
CODE_DIFF=$(git diff main..HEAD -- src/ tests/ | wc -l)
if [ "$CODE_DIFF" -gt 0 ]; then
    echo "Implementation exists, skipping planning phase"
    # Skip directly to verification
fi

P1: Batch operations in prompts

Update verification prompt (PROMPT_verify.md) to instruct:

• "Run all verification commands in a SINGLE Bash call"
• "Use parallel tool calls (multiple Glob/Read in one turn)"
• "Batch spec checkbox updates into a single Edit"

P2: Merge verification + PR into single session

Eliminate session 3 entirely by having the verification session also create the PR.

P3: Prompt compression

• Remove AGENTS.md references (saves failed reads)
• Trim verbose prompt sections
• Remove unnecessary skill invocations

Projected Impact

Optimization	Turns saved	Tokens saved
Skip planning when code exists	43	~2.3M
Batch verification operations	57	~3.0M
Merge verify+PR sessions	20	~1.1M
Total	120	~6.4M

Current: 180 turns, 9.6M tokens
Optimized: ~60 turns, ~3.2M tokens (67% reduction)

Acceptance Criteria

• feature-loop.sh checks for existing code before planning phase
• Verification prompt instructs batching of operations
• Token consumption for equivalent task drops below 5M effective input
• Turn count drops below 100 for a single-iteration loop