token-miser 0.4.0

Benchmark Claude Code configuration packages

token-miser

Claude Code's behavior changes dramatically depending on what's in CLAUDE.md, settings.json, and any hooks you've wired up. token-miser gives you reproducible numbers instead of vibes.

It runs identical tasks under different configurations (called packages) and measures token usage, cost, and task-completion quality. It supports Claude Code and Codex from the same CLI, so you can compare a vanilla setup against a tuned package — across one agent or both — and know whether the improvement is real.

Benchmark Results

Benchmark data is committed to docs/results-{suite}.json in this repo after each CI run. Open docs/index.html locally to view the dashboard.

All experiment data is collected by the repo owner on a consistent setup. See Submitting a package if you want your package included.

Concepts

• Task — A YAML file describing work for an agent: a prompt, a target repo, success criteria, and a quality rubric.
• Package — A configuration directory with a manifest.yaml that bundles CLAUDE.md/AGENTS.md, optional settings.json (hooks, permissions), and optional services.
• Suite — A named collection of tasks (e.g. quick, standard, axis, domain-python-api).
• Run — A single execution of a task under one package.
• Experiment — One or more runs across baseline/package and agent combinations.

Requirements

• Python 3.11+
• Claude Code CLI installed and authenticated (claude auth status)
• Codex CLI installed and authenticated for Codex runs (codex)

Install

pip install token-miser

Or with uv:

uv tool install token-miser

Quick start

The primary workflow is three commands:

# 1. Run the benchmark suite
token-miser tune --suite quick --package caveman --yes

# 2. View the cross-package comparison
token-miser matrix --suite quick

# 3. Export a CI score artifact
token-miser score --suite quick --package caveman --output score.json

Additional tune examples:

# Test with both Claude and Codex
token-miser tune --suite quick --package caveman --agent both --yes

# Compare multiple packages across the axis suite
token-miser tune --suite axis --package c-structured --agent both --yes

For single-task experiments:

# Run one task against a package
token-miser run \
  --task benchmarks/tasks/bm-axis-explore.yaml \
  --baseline vanilla \
  --package caveman

# Compare results
token-miser compare --task bm-axis-explore

# Inspect a specific run
token-miser history
token-miser show 1

Commands

Command	Purpose
tune	Run a benchmark suite against baseline + package (primary workflow)
matrix	Cross-package comparison grid (text + JSON export)
score	Quality-first score report for CI artifacts
run	Execute a single task under baseline and/or package
compare	Side-by-side comparison of runs for a task
analyze	Statistical summary (mean, stdev, median per package)
history	List all recorded runs
show <id>	Inspect a specific run in detail
digest	Export run data for git tracking
suite	List, validate, or pre-clone suite repos
packages / list	List available packages
publish	Publish a package to a git repo for kanon distribution
tasks	List available task YAML files
migrate	Initialize or migrate the database

tune

token-miser tune \
  --suite axis \
  --package caveman \
  --agent both \
  --model sonnet \
  --timeout 600 \
  --yes

Flag	Default	Purpose
--suite	standard	Benchmark suite name
--package	(optional)	Package to test: name or path
--agent	claude	claude, codex, openai (alias), or both
--model	agent-specific	Claude defaults to sonnet; Codex to gpt-5.4. For Bedrock, use the full cross-region inference ID (e.g. us.anthropic.claude-sonnet-4-5-20250929-v1:0)
--timeout	600	Per-task timeout in seconds
--skip-baseline	off	Reuse the last baseline run
--bare	off	Skip hooks/plugins (cheaper, less realistic)
--yes	off	Skip confirmation prompts

run

token-miser run \
  --task benchmarks/tasks/bm-axis-explore.yaml \
  --baseline vanilla \
  --package caveman \
  --agent codex \
  --timeout 600

Flag	Default	Purpose
--task	(required)	Path to task YAML
--baseline	(required)	Baseline: vanilla, package name, or path
--package	(optional)	Package to test: name or path
--agent	claude	claude, codex, openai (alias), or both
--model	agent-specific	Claude defaults to sonnet; Codex to gpt-5.4
--timeout	600	Per-invocation timeout in seconds

score

token-miser score --suite quick --package caveman --output score.json

Generates a quality-first score for the most recent tune session. The output JSON is the artifact uploaded by the CI grading workflow and consumed by badge URLs.

Flag	Default	Purpose
--suite	(required)	Suite name matching the tune run
--package	(required)	Package name (e.g. lean) or full label (e.g. claude:lean)
--output	token-miser-score.json	Path to write JSON

The score report includes quality rate, cost and token totals for qualifying runs, per-task normalized figures (cost_per_qual_task, tokens_per_qual_task), and deltas against the baseline package. Output defaults to token-miser-score.json in the current directory.

Global flags

Flag	Default	Purpose
--packages-dir	$TOKEN_MISER_PACKAGES_DIR or ./packages	Directory containing packages

Package names (no /) resolve to {packages-dir}/{name}/. Paths with / are used as-is.

Suites

Six benchmark suites ship in benchmarks/suites/:

Suite	Tasks	Focus
quick	8	Fast screening
standard	15	General-purpose benchmark
axis	8	Interaction patterns (explore, multiturn, diff, testrun, bashheavy, smallio, reasoning, bigoutput)
domain-python-api	8	Python API development
domain-iac	8	Infrastructure as code
domain-frontend	8	Frontend development

token-miser suite list         # show available suites
token-miser suite validate     # check all task YAMLs are valid
token-miser suite prep         # pre-clone repos to speed up runs

Task format

id: my-task
name: "Human-readable name"
repo: "https://github.com/owner/repo"
starting_commit: "abc1234"
prompt: |
  What you want the agent to do...
success_criteria:
  - type: file_exists
    paths: ["some/file.py"]
  - type: command_exits_zero
    command: "uv run pytest"
quality_rubric:
  - dimension: "correctness"
    prompt: "Score 0-1 based on..."

Task repo fields support ${VAR} expansion from environment variables.

How it works

For each run in an experiment:

1. Create an isolated temp directory as HOME
2. Clone the task's target repo and checkout the starting commit
3. Copy the selected agent's auth/config into the isolated HOME
4. Apply the selected package (copy targets, deep-merge settings.json)
5. Invoke the selected backend:
   • Claude: claude --print --dangerously-skip-permissions --output-format json
   • Codex: codex exec --json --sandbox workspace-write --full-auto
6. Check success criteria against the workspace
7. Optionally score quality via Claude-as-judge (requires ANTHROPIC_API_KEY)
8. Store results in SQLite (~/.token_miser/results.db)

Packages

24 packages ship in packages/, spanning different optimization strategies:

Package	Strategy
caveman	Minimal tool use, think before acting
c-structured	Structured output, systematic approach
tdd-strict	Strict TDD — failing test first, always
thorough	Maximize correctness — read everything, explain reasoning
token-miser	Minimize tokens — terse output, lazy reads, no extras
lean	Minimal overhead, skip unnecessary steps
personal	Rubin's full personal config — everything, everywhere, all at once
rtk	Rust CLI hook that compresses verbose command outputs
...	See token-miser list for all 24

Each package is a directory with a manifest.yaml:

name: my-package
version: 0.1.0
author: your-name
description: What this package does
targets:
  - path: AGENTS.md
    dest: AGENTS.md
  - path: settings.json    # optional — deep-merged into experiment config
    dest: settings.json

Packages can include CLAUDE.md/AGENTS.md instructions, settings.json for hooks and permissions, and hook scripts. The settings.json is deep-merged with any existing experiment configuration rather than overwriting it.

To use packages from another directory:

export TOKEN_MISER_PACKAGES_DIR=~/.claude/packages
token-miser list
token-miser tune --package lean

If you install packages via kanon, they land in ~/.claude-marketplaces/. Point token-miser there:

export TOKEN_MISER_PACKAGES_DIR=~/.claude-marketplaces
token-miser list

Shared instructions

Codex uses AGENTS.md. Claude Code uses CLAUDE.md. To share one instruction file across agents, use the @AGENTS.md import pattern in CLAUDE.md:

# CLAUDE.md
@AGENTS.md

Matrix runs

For package screening across a benchmark suite:

SUITE=quick REPEATS=1 AGENTS=claude,codex \
  ./scripts/run-suite-shared-baseline.sh

Runs one shared vanilla baseline per agent x repeat x suite, then reuses it across packages.

For stricter order-balanced comparisons:

SUITE=quick REPEATS=2 AGENTS=claude,codex \
  ./scripts/run-suite-crossover.sh

Alternates baseline/package order by repeat to control for run-to-run drift.

Data

All results are stored locally in ~/.token_miser/results.db (SQLite). Runs record the agent backend, so Claude and Codex experiments coexist in the same history and reports.

token-miser history            # compact list of all runs
token-miser show 12            # full token breakdown for one run
token-miser matrix --suite axis --json results.json   # export comparison grid

CI grading

Add automated benchmarking and a quality badge to any package repo.

Prerequisites: AWS Bedrock access (Claude Sonnet 4.5 via cross-region inference, us-east-2). Add AWS_ROLE_TO_ASSUME as a repository variable in GitHub → Settings → Variables → Actions pointing to an IAM role with bedrock:InvokeModel permissions for us.anthropic.claude-sonnet-4-5-20250929-v1:0.

Setup:

1. Copy docs/grade-template.yml to .github/workflows/grade.yml in your package repo
2. Configure your AWS OIDC role (AWS OIDC guide)
3. Push — the workflow runs on every PR and posts a quality score comment

Workflow status badge (green = all tasks passed):

[![grade](https://github.com/<owner>/<repo>/actions/workflows/grade.yml/badge.svg)](https://github.com/<owner>/<repo>/actions/workflows/grade.yml)

Score JSON badge (shows quality rate %, updates each run):

[![quality](https://img.shields.io/badge/dynamic/json?url=<raw-score-json-url>&query=%24.quality_rate&label=quality&suffix=%25&color=blue)](https://github.com/<owner>/<repo>/actions/workflows/grade.yml)

Host score.json somewhere public (GitHub Pages, a gist) and replace <raw-score-json-url> with its raw URL.

PR comment format (posted automatically on each pull request):

## token-miser grade

|  | this package | baseline |
|---|---|---|
| **quality rate** | 87% (7/8) | 50% (4/8) |
| **cost / task** | $0.1234 | $0.1456 |
| **tokens / task** | 12,345 | 14,567 |

Development

The source repo is private (caylent/token_miser). If you have access:

git clone https://github.com/caylent/token_miser.git
cd token_miser
uv sync
uv run pytest -q
uv run token-miser --help

Submitting a package for benchmarking

All benchmark data is collected by me (Rubin Johnson) on a consistent hardware and model setup. I don't accept self-reported numbers — I run all experiments myself so results are comparable.

To submit a package:

1. Open a GitHub issue titled benchmark request: <package-name>
2. Include your package files (CLAUDE.md / AGENTS.md, optional settings.json, optional hooks)
3. Describe the strategy in 2-3 sentences: what behavior it changes and why
4. Include reproduction details: model, suite, any env requirements
5. If you have your own numbers, include them — I'll run independently and compare

Packages that test a genuinely new strategy dimension are most likely to get picked up.

If your package is already in a kanon registry, link to it in the issue.

Ecosystem

kanon distributes versioned configuration packages. token-miser benchmarks them by running identical tasks under baseline and package configurations, then comparing token usage, cost, and quality. Results can be published back to kanon so teams converge on the best-performing package.

kanon distributes  ->  token-miser measures  ->  publish to kanon
(versioned packages)   (A/B comparison)         (best package wins)