EvidentLoop

Turn local Git diffs into evidence-backed, traceable audit reports.

English · 简体中文

Use EvidentLoop to inspect a local Git diff, trace each finding to the changed lines, record a human decision, and explicitly verify a later fix.

Your coding agent makes the judgment. EvidentLoop validates the result and saves the evidence in a self-contained HTML report with its JSON record. The result stays traceable without being presented as proof that the code is safe.

Prefer to see the report first? Jump to the real audit evidence.

Quick start

Requirements: Git, Python 3.10 or newer, uv, Node.js/npx for Skill installation, and a coding agent with Skill support.

# Try the bundled offline demo
uvx evidentloop demo

# Install the CLI and Skill
uv tool install evidentloop
npx skills@latest add evidentloop/evidentloop --skill evidentloop -g
evidentloop doctor

pipx install evidentloop is the CLI fallback.

Inside the Git repository to inspect, ask your coding agent:

Use EvidentLoop to audit my staged changes and generate the HTML report.

Real audit evidence

Representative schema 0.5 report · decision and comment · Alpha UI in Simplified Chinese

The animation uses a representative schema 0.5 report generated by the current renderer to show a finding, its complete diff, and a local decision with a comment. The separate self-audit report covers 43/43 files and ends complete / pass_candidate with 0 findings. It proves report generation, validation, and rendering for that range. The public evidence does not yet include a real two-diff fix-verification run.

How it works

• Review a complete diff. Your coding agent returns its judgment; EvidentLoop validates the result and renders one self-contained report.
• Revise the report, not the code. Feedback on the same diff updates only the existing report and JSON record without asking the model to review again.
• Verify a later diff explicitly. A source report, chosen finding, and claim start a new full-diff review. Old reports remain unchanged and EvidentLoop does not guess finding matches.

Output

Pending feedback stays in each viewer's browser until copied or downloaded. The HTML can be shared after redaction, but it is not a multi-user review service.

Alpha scope

Integration and development

Use EvidentLoop through the Skill as a standalone product; no other workflow is required. Integrators can follow the prepare -> external review -> finalize path in AI host integration, use the public API in evidentloop.api, and rely on diff_version plus report_version for artifact identity.

For local development:

python -m pip install -e '.[dev]'
python -m pytest -q
python -m ruff check .
python -m build

References: Pages · V0 scope · Data model · AI host integration

Alpha feedback and bug reports are welcome in GitHub Issues.

License

Licensed under the MIT License.