UCM — Unified Context Model

Probabilistic impact analysis for code changes.

UCM builds a Bayesian dependency graph of your codebase and answers: "I changed this function — what else might break, and how confident are you?"

⚡ TL;DR: What is UCM?

UCM is not a log collector. It is a Code Impact Analysis Engine.

UCM acts like a highly intelligent, real-time map of your codebase:

1. It ingests code and tickets: It parses your source code (functions, imports, API endpoints) and your Jira/Linear tickets to find entities and relationships.
2. It builds a Graph: It connects all of this together in memory. It knows that "Ticket JIRA-42" requires changes to the validateToken function, and that the processPayment endpoint imports validateToken.
3. It answers "What if?": If you ask, "I am about to change validateToken, what else is going to break?", UCM runs a Bayesian probability algorithm to highlight every connected piece of code and calculate a "confidence score" that it will be impacted.

What is the ucm-api server for?

The API server exists so that other tools can talk to that graph cleanly:

• Your React Dashboard uses the API to visualize the graph and show the impact reports to managers.
• A GitHub Action (CI/CD) could ping the API with a git diff to auto-comment on PRs with impact warnings.

Its core job is to sit in the background, hold the codebase graph in memory, listen for new code changes, and answer impact queries.

Install

cargo install ucm

Or build from source:

git clone https://github.com/paritoshk/ucm-core
cd ucm-core
cargo build --release

Quick start

# Scan a TypeScript project
ucm scan ./src --language typescript

# What breaks if I change validateToken?
ucm impact src/auth/service.ts validateToken

# Get test recommendations for that change
ucm intent src/auth/service.ts validateToken

Example output:

UCM Impact Analysis
====================
  Changed: src/auth/service.ts#validateToken

  DIRECT IMPACTS:
    authMiddleware — 95% confidence
      1. authMiddleware imports validateToken directly (StaticAnalysis)

  INDIRECT IMPACTS:
    processPayment — 76% confidence (2 hops)
      1. processPayment depends on authMiddleware
      2. authMiddleware imports validateToken

  NOT IMPACTED:
    generateReport — 90% safe (No graph path to changed entities)

How it works

UCM scans your source files, builds a typed dependency graph, and runs a reverse BFS from the changed entity. Each hop applies confidence decay:

confidence(path) = Π edge_weight_i

When multiple independent sources confirm the same relationship (static analysis + API traffic logs), UCM fuses them with Noisy-OR:

P(edge) = 1 − Π(1 − P(source_i))

This means two 80% signals produce 96% confidence — not 64% (naive multiply). Confidence also decays over time at rates tuned per relationship type (import statements decay slowly; API traffic patterns decay fast).

CLI reference

ucm scan <path> [--language rust|typescript|python]
    Scan source files and print graph statistics.

ucm graph <path> [--export json]
    Show entity list or export full graph as JSON.

ucm impact <file> <symbol> [--min-confidence 0.1] [--max-depth 10] [--json]
    Run reverse BFS from a changed symbol. Print impacted entities with
    confidence scores and explanation chains.

ucm intent <file> <symbol> [--json]
    Same as impact, but formats output as prioritised test scenarios:
    MUST TEST / SHOULD TEST / RISKS / COVERAGE GAPS.

REST API

The ucm-api binary exposes the same analysis over HTTP (default: localhost:3001).

cargo run --bin ucm-api

curl -s http://localhost:3001/health
# {"status":"ok"}

curl -s -X POST http://localhost:3001/impact \
  -H 'Content-Type: application/json' \
  -d '{"changed_entities":[{"file_path":"src/auth/service.ts","symbol":"validateToken"}]}'

Dashboard

Interactive UI for exploring the graph and running impact analysis:

cd dashboard
npm install && npm run dev
# http://localhost:5173

Set VITE_API_URL to point at a remote ucm-api instance.

Architecture

ucm-core      — graph types, Bayesian math, SCIP identity  [open-source]
ucm-ingest    — source adapters: code, git, Jira, Linear   [this repo]
ucm-events    — event store + graph projection             [this repo]
ucm-reason    — BFS impact engine, test intent             [proprietary]
ucm-observe   — event replay, audit trail                  [this repo]
ucm-api       — Axum REST server                           [this repo]
ucm-cli       — terminal interface                         [this repo]

Event sourcing: every parser and adapter emits immutable UcmEvents. The projection replays them to build the graph — any point-in-time state is reproducible by replaying the event log up to that timestamp.

SCIP identity: entities use Sourcegraph SCIP-style strings (scip:local/project/0.0.0/src/auth/service.ts#validateToken), so files can be re-indexed independently without central ID coordination.

Current limitations

Development

cargo test --workspace   # run all tests
cargo clippy --workspace # lint
cargo fmt --all          # format

License

MIT — see LICENSE.