DraftOps

Privacy-aware, draft-only support triage automation with explicit human review.

DraftOps reads support tickets, removes personal and policy-defined identifiers, classifies intent and priority, prepares reply drafts, and writes review artifacts. It contains no delivery connector and every output stays pending until a named human records a decision.

The included rules classifier makes demos and tests deterministic. A separate model runner can supply recorded AI predictions through a JSON file; DraftOps validates those predictions and redacts generated drafts again before storing them. This keeps model credentials and network calls outside the pipeline.

Quick start

Requirements: Python 3.10+.

python -m pip install -e .

draftops process examples/tickets.jsonl \
  --policy examples/policy.json \
  --predictions examples/recorded-predictions.json \
  --out .tmp/demo-run

Generated artifacts:

.tmp/demo-run/
├── RUN.md
├── privacy-report.json
├── queue.json
└── reviewgate-proposals.jsonl

The privacy report stores redaction labels and counts, never original values. The queue contains sanitized ticket context and drafts. The JSONL proposals can be submitted one by one to a compatible ReviewGate policy.

Run without --predictions to use deterministic classification:

draftops process examples/tickets.jsonl \
  --policy examples/policy.json \
  --out .tmp/rules-run

Local review flow

Record a final human decision:

draftops decide .tmp/demo-run/queue.json DEMO-1042 \
  --approve \
  --actor reviewer@example.invalid \
  --note "Fictional demo reviewed" \
  --decisions .tmp/demo-run/decisions.jsonl

Export only approved drafts:

draftops export .tmp/demo-run/queue.json \
  --decisions .tmp/demo-run/decisions.jsonl \
  --out .tmp/demo-run/approved

Approved exports deliberately omit delivery addresses. A separate, authenticated connector must map the internal ticket ID to a destination and should re-check the approval before sending.

Pipeline

flowchart LR
  I[Ticket JSONL] --> R[PII and policy redaction]
  R --> C[Rules or recorded AI classification]
  C --> D[Draft redaction]
  D --> Q[Pending review queue]
  Q --> H[Named human decision]
  H --> E[Address-free approved export]
  Q --> G[ReviewGate proposal JSONL]

Loading

Input and recorded predictions

Ticket JSONL fields:

{"id":"DEMO-1","subject":"Question","body":"Fictional text","customer_email":"person@example.invalid"}

Recorded predictions are keyed by ticket ID:

{
  "DEMO-1": {
    "category": "product",
    "priority": "normal",
    "confidence": 0.86,
    "draft": "A specialist will review the fictional setup."
  }
}

The category must exist in the selected policy. Priority and confidence are strictly validated. Drafts are always reprocessed by the redaction layer.

Safety boundary

Regex redaction is best-effort. It cannot detect every personal identifier, trade secret, or contextual disclosure. Minimize input, keep real runs outside public repositories, review artifacts manually, and apply organization-specific data-retention and access controls.

This project demonstrates engineering controls; it is not a compliance claim or legal advice.

Development

python scripts/public_safety.py
python -m unittest discover -s tests -v
python -m compileall -q src tests

The public-safety check scans every tracked file and the full Git history for high-confidence credential patterns and forbidden secret-bearing filenames. It reports rule names and locations without printing matched values.

Roadmap

• Signed prediction import format
• IMAP/Zendesk adapters that retain draft-only behavior
• Policy-specific retention windows and deletion command
• Reviewer UI with side-by-side redaction evidence

License

MIT. See LICENSE.

If DraftOps is useful for building safer support automation, a star helps other builders find it.