DiffSeal

Turn Python verification signals into one reproducible review-readiness evidence packet and explicit decision.

DiffSeal is a local-first, Python-first review-readiness gate. It runs the verification tools you already use against one exact repository state, normalizes their output into a single evidence bundle, and produces one explicit decision: is this change ready for review?

Quick flow

For a published release, install DiffSeal with:

pip install diffseal

Then, inside a Python repository:

diffseal init
diffseal plan
diffseal run

A run produces one explicit decision and two evidence files:

decision: PASS
evidence.json: ...
evidence.md: ...

DiffSeal orchestrates and normalizes the verification signals you already use — pytest, Ruff, coverage, and dependency checks — into one review-readiness decision. It does not replace those tools:

Python repository
    |
    +-- pytest
    +-- Ruff
    +-- coverage
    +-- dependency check
            |
            v
    normalized evidence
            |
            v
    PASS / FAIL / REVIEW_REQUIRED / INSUFFICIENT_EVIDENCE
            |
            +-- evidence.json
            +-- evidence.md

Why DiffSeal

Verification output is scattered. A change produces pytest output, Ruff output, coverage output, and dependency availability signals across different logs and exit codes. A reviewer has to gather and interpret all of it manually before deciding whether a pull request is ready.

DiffSeal collects those signals into one normalized evidence packet and gives the reviewer an explicit, reproducible decision instead of a pile of logs.

What it produces

Running DiffSeal writes two files:

• evidence.json — the canonical, normalized evidence bundle
• evidence.md — a derived, reviewer-friendly summary

The bundle ends with one decision:

• PASS — all required evidence is satisfactory
• FAIL — evidence established that a required verification failed
• REVIEW_REQUIRED — evidence exists but configured policy requires human review
• INSUFFICIENT_EVIDENCE — required evidence could not be established (for example, a required tool was unavailable)

FAIL means a required check ran and its requirement genuinely failed. INSUFFICIENT_EVIDENCE means required evidence could not be established at all — the gate refuses to pass on missing evidence rather than guessing.

Installation

Release installation

Install a published release from PyPI with:

pip install diffseal

Local development installation

From a checkout of this repository:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Quickstart

Inside a Python repository:

diffseal init      # create a minimal .diffseal.toml (refuses to overwrite)
diffseal plan      # show which checks would run, without running them
diffseal run       # run checks, evaluate, and write evidence.json + evidence.md

diffseal run exits with a stable code reflecting the decision:

Example configuration

[checks.pytest]
enabled = true
required = true
args = ["-q"]

[checks.ruff]
enabled = true
required = true
args = ["check", "."]

[checks.coverage]
enabled = true
required = false
threshold = 80.0
args = ["-q"]

[checks.dependency]
enabled = true
required = false

[evaluation]
# Optional checks listed here force REVIEW_REQUIRED when they fail.
review_on = []

[repository]
# Optional explicit base revision for change context, e.g. "main".
# base_revision = "main"

GitHub Actions

Use the release-specific Action tag in a workflow (see examples/workflow/diffseal.yml for a complete example):

- uses: zuli2021/diffseal@v0.1.1
  with:
    output-dir: evidence

The Action runs the same DiffSeal CLI against the checked-out repository. Upload the evidence artifacts in a later step:

- uses: actions/upload-artifact@<sha>
  with:
    name: diffseal-evidence
    path: evidence/

The usage example follows least-privilege defaults:

• triggered on pull_request (never pull_request_target)
• permissions: contents: read
• no repository secrets required
• no pull request comments posted

Evidence example

A compact excerpt from evidence.json:

{
  "schema_version": "0.1",
  "product_version": "0.1.0",
  "invocation_mode": "local-cli",
  "checks": [
    {
      "id": "pytest",
      "required": true,
      "outcome": "PASS",
      "summary": "2 passed"
    },
    {
      "id": "ruff",
      "required": true,
      "outcome": "PASS",
      "summary": "no lint violations"
    }
  ],
  "decision": "PASS",
  "decision_reasons": ["all required evidence satisfactory"]
}

See examples/evidence/ for a full committed sample generated by a real diffseal run against a standalone Git repository containing the contents of examples/python-basic/.

Only the following are normalized for a stable public example:

• evidence_id and run_id (per-run identifiers vary every run);
• started_at and finished_at (timestamps vary every run);
• the absolute, machine-specific temporary filesystem prefix of repository and environment_metadata.cwd, replaced with examples/python-basic.

Everything else is preserved from the real run: the head_revision is the actual commit SHA of the standalone demo repository used to generate the sample, and check outcomes, the decision, decision reasons, tool results/versions, configuration and policy hashes, and the change summary are all genuine run output.

Check semantics

Each check reports one outcome:

• PASS — the check established that its requirement passed
• FAIL — the check established that its requirement failed
• ERROR — required evidence could not be established successfully (tool missing, crashed, or unusable)
• SKIPPED — the check did not run by design, policy, or an applicable collector condition

ERROR and SKIPPED are distinct. Neither is silently treated as a pass when the check is required.

Decision semantics

The gate combines normalized checks deterministically:

1. any required check with ERROR or SKIPPED → INSUFFICIENT_EVIDENCE
2. otherwise, any required check with FAIL → FAIL
3. otherwise, a configured optional review condition → REVIEW_REQUIRED
4. otherwise → PASS

The same normalized evidence plus the same policy always produces the same decision.

Community scope

Community v0.1 collectors:

The dependency check verifies only that dependencies declared in pyproject.toml and/or requirements.txt are installed in the local environment and satisfy their declared version constraints. It is not:

• a vulnerability scanner
• a dependency security audit
• an SBOM generator
• an advisory database integration

No code is claimed "safe" or "secure" on the basis of DiffSeal output. DiffSeal reports verification evidence and a review-readiness decision.

Local-first / privacy

DiffSeal:

• executes locally or in your CI provider
• requires no DiffSeal SaaS or account
• requires no API key
• sends no telemetry
• uploads no repository source to a DiffSeal-operated service

DiffSeal runs the tools present in its own environment. Third-party CI providers and their runners have their own policies and are not guaranteed by DiffSeal.

Community vs Pro

Community is the independently useful base product. Pro is a future, commercial extension and is not yet released. Pro capabilities are not represented as available.

License

Community is licensed under the Apache License 2.0. See LICENSE.

Status

DiffSeal Community is an early/Alpha release line focused on local-first Python review-readiness evidence. The repository is public. Current release availability is represented by the repository's GitHub Releases and PyPI project page.