Token optimization plugin for Claude Code

agent-trace-ops (ato) analyzes your conversation’s tool-call chains and auto-suggests refactors to reduce tokens and latency.

1. Extracts and enriches tool call metadata with timing, token counts, and file ranges
2. Compresses session data with RLE for efficient pattern detection
3. Generates actionable optimization suggestions:
   • Quick Commands: One-liner chains for package.json/Makefile
   • Parameterized Scripts: Reusable workflows that handle different inputs
   • File Refactorings: Merge/split files accessed together to reduce token overhead

https://github.com/user-attachments/assets/cac017c2-0fe7-4a8e-841c-e431a9324769

Usage

Claude Code Plugin

/plugin marketplace add peerbot-ai/claude-code-optimizer
/plugin install agent-trace-ops

After installing the plugin, use the /plan command for on-demand analysis:

/agent-trace-ops:plan

CLI

npx agent-trace-ops

Install globally for repeated use:

npm install -g agent-trace-ops
ato --project-path=<path> --agent=claude

This will:

1. Check for existing analysis reports
2. Ask if you want to reuse or regenerate the report
3. Let you select which optimization categories to analyze:
   • Quick Commands (one-liner chains for package.json/Makefile)
   • Parameterized Scripts (reusable workflows)
   • File Refactorings (merge/split frequently accessed files)
4. Analyze patterns and generate optimization suggestions
5. Show helpers with potential token savings calculations

How it works

Every conversation in Claude Code is saved as JSONL (JSON Lines) files in ~/.claude/projects/<hash>/. The tool analyzes these sessions and generates a report:

## Sessions

### Session agent-567356cb [2025-11-04 03:15:59] (29s)
1. Read: [3106b] .zshrc[L1-L64]
💭 [+3s in=89t out=8t]
2. Edit: [+9s 1048b] settings.json[L1-L4], statusline-command.sh[L1-L31]
3. Read: [+1s 9069b] count_tokens.js[L1-L245]
💭 [+2s in=156t out=12t]
4. Edit: [+6s 740b] settings.json[L1-L4]
⏹️ MessageEnd [+4s in=12t out=45t]

### Session 396d2518 [2025-11-04 01:42:53] (2m 10s)
1. Bash: [cmd=96b out=177b] node index.js --project-path=$(pwd) --list
💭 [+2s in=45t out=6t]
2. Bash: [+4s cmd=97b out=1396b] node index.js --project-path=$(pwd) --print
⏹️ MessageEnd [+3s in=18t out=234t]

Claude inspects the behavioral data from it's own agent and suggests different optimizations.

Example Optimizations

Quick Commands - Combine repeated bash sequences (package.json/Makefile):

# Detected pattern (found 23 times in sessions):
4. Bash: [+1s cmd=8b] make build-packages
5. Bash: [+8s cmd=16b] docker compose restart gateway
6. Bash: [+1s cmd=17b] sleep 5
7. Bash: [+6s cmd=23b] docker compose logs gateway --tail=50

# Suggested: Add to Makefile or package.json
make deploy-gateway  # Reduces 4 calls → 1 call

Parameterized Scripts - Reusable workflows with different parameters:

# Detected pattern (found 15 times with different worker IDs):
8. Bash: [+0s cmd=30b] docker ps --filter "name=worker-1762140643"
9. Bash: [+2s cmd=32b] docker logs worker-1762140643-abc --tail=50
10. Bash: [+1s cmd=33b] docker logs worker-1762140643-abc | grep "error"

# Suggested script: ./scripts/worker-logs.sh
./scripts/worker-logs.sh 1762140643 50 "error"  # Reduces 3 calls → 1 call

File Refactorings - Merge co-accessed files:

# Detected pattern (found 23 times):
1. Read: [+0s 6156b] src/interactions.ts[L1-L812]
2. Read: [+2s 4720b] src/types.ts[L1-L203]
3. Read: [+1s 2696b] src/custom-tools.ts[L1-L156]

# Suggested: Merge into src/core.ts (1171 lines)
1. Read: [+0s 13572b] src/core.ts[L1-L1171]  # Reduces 3 reads → 1 read

See instructions.md for detailed pattern examples and optimization strategies.