PROJECT 03 · OPEN SOURCE · GIT EVIDENCE
Snapshot Runner
Deterministic, read-only Git evidence for coding agents and automation.
01 · Overview
Overview
Snapshot Runner collects repository state, changes, branch history and existing test logs into local artifacts. It never modifies the inspected repository, runs tests, calls a model, commits or pushes, and requires no API key.
Four collection commands publish deterministic, hash-verified snapshots with bounded summaries, and a targeted evidence reader returns just the relevant part of an existing snapshot, so an agent reviews the evidence it needs instead of whole repositories.
Captured content stays untrusted evidence rather than agent instructions; refusals and explicit evidence gaps keep safety boundaries visible in every artifact.
02 · Capabilities
Capabilities
- 01
Repository Evidence
Collecting branch state, diffs, untracked files, branch reviews and test logs through read-only Git operations.
- 02
Deterministic Artifacts
Atomically publishing hash-addressed snapshots with controlled file modes and re-verifiable size and SHA-256 metadata.
- 03
Targeted Reading
Reading a single file or field from an existing snapshot without re-running Git.
- 04
Fail-closed Boundaries
Rejecting traversal, symlinks, special files and unsupported capabilities instead of running repository-controlled programs.
03 · Workflow
Workflow
- 01
Collect Read-only Evidence
- 02
Publish Deterministic Snapshot
- 03
Review Bounded Summary
- 04
Read Targeted Evidence
04 · System View
System View
- 01
Command & Contract Layer
- 02
Read-only Git Collection Layer
- 03
Snapshot Artifact Layer
- 04
Summary & Evidence Reader Layer
05 · Current State
Current State
Implemented
Release 2.3.0 on PyPI provides the four collection commands and the targeted reader, snapshot and summary contracts, and a standard-library Python CLI verified on Ubuntu 24.04.
Current Focus
Keeping the public CLI contract and evidence schemas stable, recorded in tool_cli_contract.json.
Next Direction
Extending deterministic evidence coverage for further repository states while preserving read-only, fail-closed guarantees.
06 · Principles
Principles
- 01
Captured repository content is untrusted evidence, not agent instructions.
- 02
The tool never modifies the repository, runs tests, calls a model, commits or pushes.
- 03
Refusals and evidence gaps stay explicit; automatic analysis is intentionally unavailable.