Bounded autonomy / model-routing evaluation framework for real engineering work.
Use the least expensive adequate model for a tightly scoped slice, keep shared architecture legible, and accept results only on independent evidence — not worker self-report.
Copyright © 2026 Stephen Phillips and HappyMonkeyAI. Released under the MIT License.
Luna Framework is a documentation-first evaluation and execution discipline for coding agents:
- Decompose larger objectives into owned, verifiable task packets
- Gather repository context before editing
- Route cheaper models (e.g. Luna) to narrow slices; escalate (e.g. Terra / break-glass) only on evidence
- Verify with real commands; separate worker claims from owner acceptance
- Retain durable lessons without copying secrets or noisy logs
It is not a claim that one model beats another. Early matched runs are exploratory protocol evidence unless a registered protocol says otherwise.
| Doc | Role |
|---|---|
BRIEF.md |
Origin, design thesis, objectives |
SPINE.md |
End-to-end workflow and component map |
GLOSSARY.md |
Shared vocabulary and routing hypothesis |
TASK-TEMPLATE.md |
Bounded task packet template |
EVALUATION-PROTOCOL.md |
Rules for comparable evaluations |
REPOSITORY-INTELLIGENCE.md |
Provider-neutral context-layer contract |
RESULTS.md |
Evaluation registry (exploratory / matched / incomplete) |
LINKS.md |
Related projects and protocols |
PROGRESS.md |
Current state and limitations |
| Path | Role |
|---|---|
scripts/repo_context.py |
Native repository-context reference implementation |
scripts/test_repo_context.py |
Adversarial contract tests for that implementation |
scripts/run_matched_evaluation.sh |
Matched-evaluation launcher harness |
rubrics/ |
Scoring rubrics used in controlled confirmation tracks |
b4-repeat-track/ |
Frozen B4 confirmation evidence + SHA-256 manifest |
Canonical TypeScript fixtures used for B1–B4 live under fixtures/ as local nested git repositories (with their own tags/baselines). They are not published in this GitHub snapshot — fixtures/ is gitignored because it includes nested .git history, node_modules, and large disposable worktrees under fixtures/fixture-runs/.
What is tracked for differential oracles:
fixtures/diff-tests/— parent-side A1–A4 differential / trace harnesses
To reproduce matched runs you need the fixture repos locally (same baseline tags recorded in RESULTS.md). Publishing those fixtures as standalone public repos is optional follow-up work.
labs/glyph/ — agent-facing code encoding (typed short identifiers + legend), not model routing. Codec v0 + W1 harness are ready; worker matched runs not started. Start at labs/glyph/BRIEF.md.
python labs/glyph/test_codec.py
python labs/glyph/w1_harness.py encode --root <dir> --path src/domain/email.ts
- AgentsProtocol — agent operating protocol used alongside this discipline
- Echoes of Aion — mature-project operational pilot target (separate repo)
Do not read this repo as:
- Luna vs Terra superiority
- Statistically valid model ranking from small samples
- A production product or turnkey agent harness
- Permission to commit, push, deploy, or change ownership outside an explicit task packet
MIT — Copyright (c) 2026 Stephen Phillips and HappyMonkeyAI.