eval-banana 0.0.6

Lightweight aspect-based evaluation framework with YAML check definitions.

Eval Banana

Aspect-based evaluation framework - deterministic checks + harness judges. Score anything (agentic outputs, workflows, banana!) with simple YAML check definitions.

What it does

Eval Banana discovers YAML check definitions from eval_checks/ directories, runs them, and produces a report. Every check scores 0 or 1 with equal weight.

Two check types:

Type	Purpose	How it works
deterministic	Objective assertions (file existence, content, structure)	Runs a Python script via subprocess; exit 0 = pass
harness_judge	LLM-as-a-judge (coherence, accuracy, tone)	Invokes the configured AI agent to score target files; expects {"score": 0|1}

The harness judge uses one of the following: codex, gemini, claude, openhands, opencode, pi

Writing checks

Create a directory called eval_checks/ anywhere in your project. Add YAML files -- one per check.

Deterministic check

schema_version: 1
id: output_file_exists
type: deterministic
description: Verify that output.json was generated.
target_paths:
  - output.json
script: |
  import json, sys
  from pathlib import Path
  ctx = json.loads(Path(sys.argv[1]).read_text())
  target = ctx["targets"][0]
  assert target["exists"], f"{target['path']} not found"

Harness judge check

schema_version: 1
id: summary_is_accurate
type: harness_judge
description: The generated summary accurately reflects source data.
target_paths:
  - summary.txt
  - source_data.json
instructions: |
  Compare the summary against the source data.
  Score 1 if accurate, 0 if it contains fabricated claims.

Requires a configured harness agent. Set [harness] agent in config or pass --harness-agent.

Quick start

# Install
uv sync

# Initialize project config and example check
eb init

# Run all discovered checks
eb run

# List discovered checks without running
eb list

# Validate YAML definitions without running
eb validate

Installation

# Using uv (recommended)
uv add eval-banana

# Using pip
pip install eval-banana

# From source (development)
git clone https://github.com/writeitai/eval-banana.git
cd eval-banana
uv sync --extra dev

After installation the CLI is available as eb.

Harness configuration

harness_judge checks require a configured harness agent. Configure it via TOML or CLI flags.

TOML

# .eval-banana/config.toml
[harness]
agent = "codex"
model = "gpt-5.4"
# reasoning_effort = "high"

Custom agent templates

Add [agents.<name>] sections to override built-in templates or define new ones:

[agents.myagent]
command = ["my-cli", "run"]
shared_flags = ["--headless"]
prompt_flag = "--prompt"
model_flag = "--model"

Skills

Install bundled skills into a target project's native agent directories:

eb install
eb install --target-agents codex
eb install --skills gemini_media_use --dry-run

Agent	Destination
claude	.claude/skills/
codex	.codex/skills/
openhands	.agents/skills/
opencode	.agents/skills/
gemini	.gemini/skills/

See docs/configuration.md for details on bundled skills, authentication, and the deprecated distribute-skills command.

Configuration

Eval Banana uses a single project-level TOML config at .eval-banana/config.toml.

Create it with eb init.

Config precedence (highest to lowest)

1. CLI arguments (--output-dir, --harness-model, etc.)
2. Environment variables (EVAL_BANANA_*)
3. Project config (.eval-banana/config.toml)
4. Built-in defaults

Key settings

Setting	Default	Env var
output_dir	.eval-banana/results	EVAL_BANANA_OUTPUT_DIR
pass_threshold	1.0	EVAL_BANANA_PASS_THRESHOLD
llm_max_input_chars	0	EVAL_BANANA_LLM_MAX_INPUT_CHARS
harness.agent	unset	EVAL_BANANA_HARNESS_AGENT
harness.model	unset	EVAL_BANANA_HARNESS_MODEL

CLI reference

eb init [--force]                Create config + example check
eb run [OPTIONS]                  Run all discovered checks
eb list [OPTIONS]                 List discovered checks
eb validate [OPTIONS]             Validate YAML without running
eb install [OPTIONS]              Install bundled skills into agent dirs

Options for run/list/validate:
  --check-dir PATH              Scan only this directory
  --check-id TEXT               Run only this check ID
  --output-dir TEXT             Override output directory
  --pass-threshold FLOAT        Minimum pass ratio (0.0-1.0)
  --verbose                     Enable debug logging
  --cwd TEXT                    Working directory

Harness options (run only):
  --harness-agent TEXT          Agent CLI used by harness_judge checks
  --harness-model TEXT          Model override for the agent
  --harness-reasoning-effort TEXT  Reasoning effort level

Output

Each run creates a timestamped directory under the configured output_dir:

.eval-banana/results/<run_id>/
  report.json       # Machine-readable full report
  report.md         # Human-readable Markdown report
  checks/
    <check_id>.json       # Per-check result
    <check_id>.stdout.txt # Captured stdout (if any)
    <check_id>.stderr.txt # Captured stderr (if any)

Development

uv sync --extra dev
make test         # Run tests
make fix          # Auto-fix lint + format
make pyright      # Type check
make all-check    # Lint + format + types + tests (matches CI)

Inspiration

Eval Banana's binary 0/1 scoring philosophy draws directly on two earlier bodies of work:

• Hamel Husain's Creating LLM-as-a-Judge that drives business results — argues that binary pass/fail judgments produce more reliable, actionable evals than Likert-style 1-5 scales.
• RAGAS's Aspect Critic metric — evaluates outputs against a natural-language aspect definition and returns a binary verdict.

The harness_judge check type is essentially an Aspect Critic: you describe what "good" looks like in plain language, and the judge returns {"score": 0|1}.

Contributing

Issues and pull requests are welcome. Please run make all-check before opening a PR.

Changelog

See CHANGELOG.md for release notes.

License

Apache License 2.0 — see LICENSE for details.

Copyright 2026 WriteIt.ai s.r.o.