Parsewall

Make every file inert before it reaches the model.

Quick start · Local install · Offline install · Codex plugin

Parsewall is a deterministic, fail-closed pre-LLM scanner for untrusted uploaded documents. It never asks an LLM to decide whether a document is safe. It parses bounded structures, renders visual formats, compares OCR with embedded text, and returns only this public result contract:

{
  "risk_level": "low | review | quarantine",
  "evidence": [],
  "location": [],
  "structural_anomalies": []
}

Even a low result does not make document content trusted. A low-only plain UTF-8 derivative is prefixed with an untrusted-data marker and must remain data, never an instruction source.

Install the Python CLI

python -m pip install parsewall
parsewall scan --input ./report.txt

The package also installs parsewall-api, parsewall-worker, and parsewall-mcp. The original document-firewall* commands remain available as compatibility aliases.

Use as a GitHub Action

Scan a checked-out document in a workflow. The action fails when the result reaches the configured fail-on level (quarantine by default).

steps:
  - uses: actions/checkout@v4
  - uses: gogooma125732/parsewall@v1
    id: parsewall
    with:
      input: path/to/document.pdf
      fail-on: review
  - run: echo "risk=${{ steps.parsewall.outputs.risk-level }}"

The action installs Parsewall in the runner environment and emits the fixed JSON result contract as an output.

Run the isolated HTTP product

docker compose up --build
curl -F file=@report.pdf http://127.0.0.1:8000/v1/scans

Open http://127.0.0.1:8000/ for the local browser upload interface. It uploads one supported file, follows the isolated worker status, displays only the fixed public result fields, and exposes a derivative download only for low results. The interactive OpenAPI explorer remains at /docs.

For the versioned one-command installation and operational checks, see deploy/LOCAL_INSTALL.md.

The API only accepts uploads and serves status/results. A separate worker scans jobs with no network, a read-only root filesystem, no Linux capabilities, and bounded CPU, memory, processes, and temporary storage.

Endpoints:

• POST /v1/scans
• GET /v1/scans/{job_id}
• GET /v1/scans/{job_id}/result
• GET /v1/scans/{job_id}/derivative (low results only)

Supported inputs are UTF-8/UTF-16 text, Markdown, HTML, DOCX, PPTX, XLSX, PDF, PNG, and JPEG. Unsupported, corrupt, encrypted, incomplete, or dependency- blocked scans fail closed.

Build distribution artifacts

Build the four distribution channels in order after the versioned Docker image exists locally:

python3 scripts/build_release.py --channel all --clean

Artifacts are written under dist/releases/<version>/ with SHA-256 hashes and a machine-readable release manifest:

1. source-backed Docker Compose local installer (*-compose.tar.gz);
2. architecture-specific offline Docker bundle (*-offline-<platform>.tar);
3. Python wheel and sdist (python/);
4. standalone Codex plugin and local marketplace ZIP files.

Installation details are in deploy/LOCAL_INSTALL.md, deploy/OFFLINE_INSTALL.md, deploy/PYTHON_CLI.md, and deploy/CODEX_PLUGIN.md.

Codex plugin and MCP

The distributable plugin is under plugins/document-injection-firewall. It contains the inspect-untrusted-files skill, a PreToolUse hook that blocks local raw-document reads, and an optional root-confined stdio MCP service.

Install the PyPI command and the Git-backed marketplace, then start a new Codex task:

uv tool install parsewall==1.0.0
codex plugin marketplace add gogooma125732/parsewall --ref main
codex plugin add document-injection-firewall@parsewall

Codex's current UserPromptSubmit hook schema exposes prompt text but not an attachment list, so the hook cannot claim to intercept native attachment ingestion. It is a local-tool guardrail; the skill and MCP workflow remain the mandatory pre-read gate.