weaverx 0.1.1

AI-powered GitHub issue and PR triage for the medical AI community

WeaveRx

Medical AI GitHub issue triage with auditable drafts, local safeguards, and human-in-the-loop defaults.

WeaveRx helps maintainers triage issues faster — reproducibility blockers, dataset access, subgroup performance, privacy/DICOM, and clinical validation requests — with sources (issue excerpts that grounded the decision) and safeguard scores (local heuristics, no extra LLM calls).

Built for medical AI maintainers and research groups who need practical tooling — not a gatekeeper bot.

Contents: Features · Quickstart · See it in action · Installation · Safety · Reference · Docs

---

Features

• Domain-tuned — eight medical AI categories (reproducibility, DICOM/privacy, clinical validation, subgroup performance, and more)
• Safety by default — dry-run unless you explicitly post; --confirm required for GitHub writes; local safeguard heuristics on every draft
• Auditable — sources cite issue excerpts; safeguard scores are computed locally with no LLM on that path
• Your LLM stack — Grok, Anthropic, or OpenAI-compatible endpoints via LiteLLM; mock mode for offline CI and demos

Full docs: docs/index.md · Configuration: docs/configuration.md

---

Quickstart

Requires Python 3.11+. Environment variables: docs/configuration.md.

1. Mock (zero API keys)

weaverx triage --repo Project-MONAI/MONAI --issue 42 --mock

2. Dry-run (real GitHub, offline LLM)

weaverx triage --repo Project-MONAI/MONAI --issue 1234 --mock-llm --dry-run

3. Real LLM (Grok example)

export XAI_API_KEY=xai-...
weaverx triage --repo Project-MONAI/MONAI --issue 1234 --dry-run --json

More providers: LLM providers · examples/llm_provider_examples.md

4. JSON for automation

weaverx triage --repo Project-MONAI/MONAI --issue 42 --mock --json

---

See it in action

Command (no API keys): weaverx triage --repo Project-MONAI/MONAI --issue 42 --mock

Clean triage

Typical output: reproducibility category, ready_for_review status, source excerpts, CLEAN safeguard (0.0/10), and a postable draft in the green panel.

Text capture (accessibility / no images)

See examples/captures/triage-clean.txt.

Safeguard warning

Safeguard checks are advisory — they flag drafts for review; they never auto-block posting.

When heuristics fire (e.g. credential-like patterns, heavy repetition), the table shows Safeguard flags, status escalates to HIGH RISK / REVIEW RECOMMENDED, and the draft panel border turns yellow or red.

Text capture + JSON

• Text: examples/captures/safeguard-warning.txt
• JSON: examples/sample_safeguard_warning.json

Try it: weaverx triage --repo Project-MONAI/MONAI --issue 42 --mock -v

Example JSON output

{
  "repo": "Project-MONAI/MONAI",
  "status": "ready_for_review",
  "issue": { "number": 42, "title": "Unable to reproduce nnU-Net training results on BraTS subset" },
  "analysis": { "category": "reproducibility-environment", "priority": "high" },
  "sources": [{ "type": "issue_body", "snippet": "...", "reason": "..." }],
  "draft_response": "Hi @researcher-dev — thank you for documenting this carefully...",
  "safeguard": { "score": 0.0, "status": "clean", "triggered": [] },
  "llm": { "provider": "mock", "model": "mock" },
  "dry_run": true
}

Full JSON: examples/sample_triage_output.json

---

Installation

pip install weaverx

Or install from GitHub:

pip install git+https://github.com/FratresMedAI/WeaveRx.git@v0.1.1

Contributors / local dev:

git clone https://github.com/FratresMedAI/WeaveRx.git
cd WeaveRx
pip install -e ".[dev]"

Documentation: fratresmedai.github.io/WeaveRx

---

Safety and responsible use

WeaveRx is human-in-the-loop by design. Drafts require maintainer review before posting.

1. Never paste patient data in GitHub issues. Privacy flags are heuristic, not guaranteed.
2. Default is read-only. Writes need --post-comment/--apply-labels and --confirm.
3. Use --dry-run on repos you don't maintain.
4. Use --mock in CI and local demos without tokens.
5. Review safeguard warnings before posting flagged drafts.

Not for clinical use — maintainer support tooling only, not medical advice or a clinical decision system. Does not replace IRB, legal, or compliance review.

Read more: ETHICS.md · SECURITY.md · SUPPORT.md

---

GitHub Action

Dry-run triage when issues are opened:

- uses: FratresMedAI/WeaveRx@v0.1.1
  with:
    repo: ${{ github.repository }}
    issue_number: ${{ github.event.issue.number }}
    dry_run: "true"
    llm_provider: "grok"
  env:
    XAI_API_KEY: ${{ secrets.XAI_API_KEY }}

See action.yml and .github/workflows/triage-on-issue.yml.

---

Reference

LLM providers

Provider	CLI	API key env	Default model
Grok	--llm-provider grok	XAI_API_KEY	xai/grok-2-latest
Anthropic	--llm-provider anthropic	ANTHROPIC_API_KEY	anthropic/claude-3-5-sonnet-20241022
OpenAI-compatible	--llm-provider openai	OPENAI_API_KEY	openai/gpt-4o

Override: WEAVERX_LLM_MODEL, WEAVERX_LLM_PROVIDER. Details: docs/reference/llm-providers.md

Medical AI categories

Category	What we look for
Dataset Access & Licensing	Download links, usage terms, attribution
Model Performance (Pathology/Subgroup)	Accuracy on specific diseases or patient groups
Reproducibility & Environment	MONAI/nnU-Net versions, CUDA/PyTorch, can't reproduce results
Clinical Validation Request	External validation, reader studies, deployment
Privacy/Compliance/DICOM	PHI, de-identification, HIPAA/GDPR, DICOM metadata
Bug	Crashes, incorrect outputs
Feature/Integration Request	New capabilities, framework hooks
Documentation	Missing or unclear tutorials and API docs

Full table: docs/reference/categories.md

Draft safeguards

Local-only checks after every draft — advisory, never auto-block posting. See Safeguard warning above.

Score	Status	Meaning
0.0 – 2.9	clean	No meaningful red flags
3.0 – 6.9	review_recommended	Skim draft before posting
7.0 – 10.0	high_risk	Multiple or severe heuristics fired

Full reference: docs/reference/safeguards.md

CLI reference

weaverx triage --repo owner/repo [--issue N | --recent N] [options]

Key flags: --mock, --dry-run, --json, --llm-provider, --confirm, --post-comment, --safeguards

Full reference: docs/reference/cli.md

---

Related projects

• MONAI — open-source medical AI framework
• nnU-Net — self-configuring segmentation
• LiteLLM — unified LLM API (used by WeaveRx)
• Safire — related audit tooling from the same org

---

Citing WeaveRx

If you use WeaveRx in research or evaluations, cite via CITATION.cff (GitHub can generate a BibTeX entry from that file).

---

Roadmap (near-term)

1. Embedding-based duplicate detection — optional weaverx[embeddings] extra
2. PR triage mode — --pr for pull request review drafts
3. Expanded provider presets — more medical-AI-tuned prompt templates

See CHANGELOG.md for release history.

---

Development and contributing

pip install -e ".[dev]"
pre-commit install
ruff check .
mypy src/weaverx
pytest --cov=weaverx

CONTRIBUTING.md · CODE_OF_CONDUCT.md · docs/releasing.md

---

License

MIT — see LICENSE.