self-bench

self-bench builds private coding-agent benchmarks from work already completed in your repository, so you can compare coding agents and models on tasks drawn from your own codebase.

It finds completed requests from local coding sessions and merged GitHub pull requests, then reconstructs each task from the commit before the change. For every accepted task, self-bench creates hidden tests and a reference solution, proves that the task fails without a solution and passes with the original implementation, and exports a native task for Harbor, a runner for coding-agent evaluations.

SelfBench is sandbox-agnostic: run generation locally with Docker or on Modal, Vercel Sandbox, or E2B. Generation and Harbor validation are configured independently, so Modal is optional.

The result is a private .tar.gz benchmark that you can run against multiple models:

Your repository history
        ↓
completed requests + implementations
        ↓
validated Harbor tasks with hidden tests
        ↓
gpt-5.6-luna vs gpt-5.6-terra vs gpt-5.6-sol

Quickstart

This path runs the self-bench API, worker, Temporal, and all generation and validation sandboxes locally with Docker. It does not require an account with a hosted sandbox provider.

Prerequisites

• Bun 1.3.14 or newer
• uv for installing Harbor
• Docker with Compose
• gh, authenticated with read access to the repository
• An OpenAI API key with access to gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna
• A Git checkout with a GitHub origin and completed work in its history

Install self-bench and authenticate GitHub:

bun add --global self-bench
gh auth login

Set the model and GitHub credentials used by the local worker, plus a random token that protects the local self-bench API:

export OPENAI_API_KEY=...
export GH_TOKEN="$(gh auth token)"
export SELFBENCH_API_TOKEN="$(openssl rand -hex 24)"

1. Start self-bench

self-bench up --backend docker
export SELFBENCH_API_URL=http://127.0.0.1:8080

This starts Postgres, Temporal, the self-bench API, and a worker. The worker creates disposable local Docker sandboxes; SELFBENCH_API_URL tells subsequent CLI commands where to reach the local API. From a git worktree or any checkout not named self-bench, up gives the stack its own Compose project and ports and prints the URL, so several checkouts run side by side (see Operations).

2. Build a benchmark

self-bench run \
  --repo /absolute/path/to/your/repository \
  --easy-count 10 \
  --medium-count 10 \
  --hard-count 10 \
  --output ./self-bench-evals.tar.gz

Easy, medium, and hard candidates require at least 20, 50, and 100 changed implementation lines across 1, 2, and 3 paths respectively; the counts are generation budgets, not guarantees that every candidate will pass validation.

The repository must be a Git checkout with a GitHub origin. self-bench pins its current HEAD, ignores uncommitted changes, and may take hours to author, validate, review, and export the accepted tasks. --output waits for completion and verifies the downloaded archive with SHA-256.

3. Evaluate with Harbor

Install Harbor and extract the generated tasks:

uv tool install --python 3.12 'harbor==0.23.0'

mkdir -p ./self-bench-export ./self-bench-tasks
tar -xzf ./self-bench-evals.tar.gz -C ./self-bench-export
for archive in ./self-bench-export/tasks/*.tar.gz; do
  task_id="$(basename "$archive" .tar.gz)"
  mkdir -p "./self-bench-tasks/$task_id"
  tar -xzf "$archive" --strip-components=1 -C "./self-bench-tasks/$task_id"
done

Run Harbor's Codex agent adapter directly. Harbor evaluates every extracted task at high reasoning and keeps the model results in one job directory. Repeat with additional --model values when comparing models:

harbor run \
  --path ./self-bench-tasks \
  --agent codex \
  --model gpt-5.6-luna \
  --model gpt-5.6-terra \
  --model gpt-5.6-sol \
  --ak version=0.146.1 \
  --ak reasoning_effort=high \
  --env docker \
  --jobs-dir ./harbor-jobs \
  --n-concurrent 1 \
  --yes

The evaluated agent receives the base repository and task instruction, but not the hidden tests or reference solution.

See Running self-bench evaluations for running tasks directly with Harbor.

The signed-in site also supports repository-scoped solver runs with live output and completed results. See Run tasks from the site for worker configuration and supported choices.

Run management

Closing the waiting CLI does not cancel a submitted workflow. If the local worker or Docker stack stops, work pauses until the worker is restarted.

self-bench list                    # find run IDs
self-bench status RUN_ID
self-bench cancel RUN_ID
self-bench download RUN_ID ./self-bench-evals.tar.gz

Stop the local stack with:

self-bench down

Named Docker volumes retain Temporal history and generated artifacts.

Choose a sandbox backend

The quickstart uses local Docker sandboxes so it works without a hosted sandbox account. For more concurrency or unattended runs, generation can use Modal, Vercel Sandbox, or E2B instead:

• Docker: self-bench up --backend docker keeps generation and validation on your machine.
• Modal: authenticate with modal token new, then run self-bench up --backend modal.
• Vercel Sandbox: run self-bench setup vercel; Harbor validation defaults to Vercel too.
• E2B: build the pinned SelfBench runtime with self-bench setup e2b --name NAME[:TAG]; Harbor validation defaults to E2B too.
• Temporal Cloud + Modal/E2B: use a persistent worker for unattended runs and large repositories.

See Operations and deployment for provider setup, credentials, persistence, object storage, and Temporal Cloud deployment.

Generation sandboxes and Harbor validation are independent choices. Every generation backend defaults to the matching Harbor environment, and --harbor-environment picks a different one. Daytona is available for Harbor only:

# Matching defaults
self-bench up --backend docker
self-bench up --backend modal

# Vercel generation, with Vercel Harbor by default or any other environment
self-bench setup vercel
self-bench up --backend vercel                           # Vercel + Vercel
self-bench up --backend vercel --harbor-environment docker
self-bench up --backend vercel --harbor-environment modal
self-bench up --backend vercel --harbor-environment e2b
DAYTONA_API_KEY=... self-bench up --backend vercel --harbor-environment daytona

# E2B generation uses a required prebuilt template; setup never installs at runtime
export E2B_API_KEY=...
self-bench setup e2b --name selfbench-runtime:v1
export SELFBENCH_E2B_TEMPLATE=selfbench-runtime:v1
self-bench up --backend e2b                              # E2B + E2B
# self-bench up --backend e2b --harbor-environment docker
# self-bench up --backend e2b --harbor-environment modal

# Explicit cross-provider combinations are also supported
self-bench up --backend docker --harbor-environment modal
self-bench up --backend modal --harbor-environment docker

Provider selection belongs to the worker, so all runs on one task queue use the same pairing. See Operations and deployment for Vercel and E2B setup for template builds, credentials, plan limits, resources, cleanup, and unattended deployment.

How tasks are validated

An accepted task must:

1. Preserve a real human request from repository history.
2. Start from the repository state before the completed change.
3. Include hidden tests that fail against the base snapshot.
4. Pass after applying the original implementation.
5. Survive deterministic reruns and an independent model review that rejects tests tied to private details of the reference solution.

Exports contain repository snapshots, hidden tests, and reference solutions. They are sensitive and unencrypted; keep them private.

See Task construction and validation for the full acceptance rules and archive format.

Development

git clone https://github.com/mupt-ai/self-bench.git
cd self-bench
bun install --frozen-lockfile
bun run validate

Run the CLI directly from source:

bun run cli -- --help

Useful development commands:

bun run dev:api
bun run dev:worker
bun run dev:review

Documentation

• Task construction and validation
• Running evaluations
• HTTP API reference
• Operations and deployment

License

MIT © 2026 Mupt AI.