Repairs you can
reason about.

Dependency Rescue is a LangGraph application for investigating and repairing dependency upgrades. Its output is a proposed patch with execution evidence for human review.

The problem

A dependency upgrade can break application code even when installation succeeds. The workflow establishes a passing baseline, reproduces the failure against the target version, and uses that concrete failure to guide a bounded repair loop.

The graph

Intake → baseline → upgrade → reproduce → investigate → diagnose → normalize → apply → verify → evidence → approval. Failed verification supplies feedback for a limited retry. Successful verification reaches a review gate.

Repository integration

The application supports GitHub OAuth and optional GitHub MCP inspection. Supported project profiles include Node.js and Python. Dependencies and configuration are inspected before a workflow can be prepared. A connected repository is not a guarantee that its tests can run unattended.

Execution boundaries

• Tests execute in isolated Docker containers with network access disabled and constrained resources.
• Tests and installed packages are protected from repair edits, with integrity checked after execution.
• Python dependency preparation uses a separate wheel-only installation workflow. Test execution receives no model credentials.
• Source changes are validated and unsafe capabilities restricted. Repository content is untrusted input.
• Human approval is separate from agent verification. The agent does not automatically merge repairs.

The recorded demonstration

A small local Python development fixture upgrades HTTPX from 0.27.2 to 0.28.1. Its baseline passes 2 tests, the upgrade passes 1, and the repaired source passes 2. The change replaces Client(app=…) with Client(transport=httpx.WSGITransport(app=…)). The model used was Gemini 3.1 Pro Preview.

The recorded run remains awaiting approval. This fixture is not the upstream HTTPX repository and is not an independent benchmark. Its result demonstrates one reproducible repair, not a general success rate.

What is deployed here

This Azure deployment hosts the interactive showcase and recorded walkthrough. The workspace connects to a separately hosted authenticated API and isolated repair worker. Readiness depends on the deployed worker, model, and prepared repository snapshots. Automatic preparation of new repositories is currently limited to the local Docker workflow.

Current limits

Supported runtime profiles and dependency formats are constrained. Projects requiring external services or credentials may fail baseline validation before repair is attempted. Passing tests are evidence within the tested scope, not proof that every behavior is correct.