Back to the catalog

goldilocks

Select an AI model, reasoning effort, and attempt strategy from validated evidence.

Open source Open in the app JSON README (API)

About

Select an AI model, reasoning effort, and attempt strategy from validated evidence.

Details

Kind
Plugins
Topic
AI, RAG & memory
Publisher
newnonsick
Origin
gemini
Category
ferramentas
Version
0.1.1
Last push
2026-07-25T12:56:52Z
Repository state
ativo
Language
Python
License
MIT
Added
2026-08-30 14:13:39
Updated
2026-08-30 14:13:39
Origin id
newnonsick/goldilocks

README

# Goldilocks

Goldilocks is an Agent Skill that recommends a model, reasoning effort, and attempt strategy from validated benchmark evidence. It applies hard constraints before ranking, exposes provenance, and abstains when the evidence cannot support a safe choice. Routing remains recommendation-only: the host platform decides whether a model or effort can actually be selected.

## How it runs

The Skill first reuses suitable validated evidence already in context. With a platform-native read-only fetch capability, it can retrieve official data without installing a runtime. If native fetch is unavailable and Python 3.10+ is already present, the bundled deterministic core is the fallback. Goldilocks never installs Python.

Acquisition uses the official versioned JSON endpoint first and the official webpage only as partial fallback evidence. Remote content is treated as data, never instructions. A validated last-known-good cache may be used according to freshness policy.

## Quick start

```sh
git clone https://github.com/newnonsick/goldilocks.git
cd goldilocks
python skills/goldilocks/scripts/goldilocks.py doctor --format json
python skills/goldilocks/scripts/goldilocks.py models list --compact
```

Use your platform's installation procedure from [Platform compatibility](docs/platform-compatibility.md). Native manifests, CLI registration, and manual copies are different integration tiers; the adapter pages contain the platform-specific commands and boundaries.

## CLI

Discover evidence progressively:

```sh
python skills/goldilocks/scripts/goldilocks.py models list --compact
python skills/goldilocks/scripts/goldilocks.py models list --all --compact
python skills/goldilocks/scripts/goldilocks.py models show MODEL_ID --detail decision
python skills/goldilocks/scripts/goldilocks.py versions list --all
python skills/goldilocks/scripts/goldilocks.py shortlist --input examples/low-cost.input.json --limit 3
python skills/goldilocks/scripts/goldilocks.py recommend --input examples/quality-first.input.json
```

`models list` returns a compact catalog. `models show` inspects one model; `--all` includes all efforts and discovered versions for inspection, not automatic cross-version ranking. `shortlist` returns decision-rich candidates and `recommend` returns one deterministic recommendation or an abstention. Use `--allow-prerelease` only when prerelease evidence is acceptable. See [Model selection](docs/model-selection.md) for task inputs and version policies.

Cache diagnostics are read-only unless refresh is requested:

```sh
python skills/goldilocks/scripts/goldilocks.py cache status --version v1.1
python skills/goldilocks/scripts/goldilocks.py cache refresh --version v1.1
python skills/goldilocks/scripts/goldilocks.py doctor --format json
```

## Safety boundaries

- One attempt uses observed `pass_at_1`.
- Four compatible attempts use observed `pass_at_4` only.
- Attempts 2, 3, greater than 4, or incompatible strategies have null reliability. No metric is interpolated or estimated.
- Total attempt cost is exact multiplication when its operand exists; it is labeled derived.
- Unknown platform availability, required capability, or comparability causes rejection or abstention.
- Goldilocks does not invoke subagents, enforce routing, or send task text to benchmark sources.

See [Architecture](docs/architecture.md), [Installation](docs/installation.md), [Troubleshooting](docs/troubleshooting.md), and [Security policy](SECURITY.md).

## Development

The deterministic core uses only the Python standard library at runtime. Run the offline suite and repository validator with Python 3.10+:

```sh
python -m unittest discover -s tests -v
python scripts/validate-repository.py
```

The live contract test is opt-in with `GOLDILOCKS_LIVE_TEST=1`; normal verification performs no network calls.

Licensed under the [MIT License](LICENSE).

More