RepoRipple

Know what a code change can break before you merge it.

RepoRipple is a local-first change-impact analyzer for engineers and coding agents. It builds a lightweight dependency graph from a repository, reads the current Git diff, traces reverse dependencies, recommends relevant tests, and emits review-ready Markdown, JSON, or a standalone visual impact map.

The deterministic analyzer requires no API key, hosted service, source upload, or framework integration. An optional, explicitly requested OpenRouter explanation can turn its anonymized evidence into a short narrative without changing the underlying analysis.

PyPI · Portfolio case study · MCP setup · v0.6.0 release

Code graphs explain how a repository is connected. RepoRipple applies that graph to the change in front of you:

• Which files depend on the code I changed?
• How far does the blast radius travel?
• Which tests are most relevant?
• Did I touch authentication, payments, deployment, or another sensitive area?
• Should this pull request be blocked for additional review?
• What structured context should I give a coding agent?

pip install reporipple

# Analyze staged, unstaged, and untracked changes.
reporipple .

# Compare a feature branch with main.
reporipple . --base origin/main

# Generate machine-readable context for an agent or automation.
reporipple . --base origin/main --format json --output impact.json

# Generate and open a self-contained visual blast-radius map.
reporipple . --base origin/main --format html --output impact.html --open

# Create the Markdown, JSON, and visual reports together, then open the map.
reporipple . --base origin/main --format all --output reporipple-reports --open

# Fail CI when a change reaches high risk.
reporipple . --base origin/main --fail-on high

# Audit the exact anonymized OpenRouter request without sending it.
reporipple . --base origin/main --explain-preview

Run it once without installing:

uvx reporipple . --base origin/main

--format html turns the same deterministic evidence into a portable review dashboard. The map lays out changed files, direct dependents, transitive dependents, and suggested tests from left to right, then keeps the complete review paths, risk signals, and verification list below it.

reporipple . --base origin/main --format html --output impact.html --open

Use --format all --output reporipple-reports to create impact.md, impact.json, and impact.html from one repository scan. Adding --open opens the HTML map after writing the bundle.

The report has no external scripts, fonts, analytics, or network requests. Repository source code is never embedded; only the file paths and evidence already present in the standard report appear. Large results keep the full review list while limiting the visual canvas to 80 nodes. See the impact-map guide or run the included deterministic demo with ./scripts/demo-impact-map.sh.

Watch the 22-second impact-map demo — a silent, 720p recording generated from the included checkout fixture.

Add RepoRipple to a pull-request workflow to create impact reports and append the Markdown report to the GitHub job summary. This immutable v0.6.0 example produces Markdown, JSON, and visual HTML:

name: RepoRipple

on:
  pull_request:

permissions:
  contents: read

jobs:
  impact:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0
          persist-credentials: false
      - id: reporipple
        uses: adiarora06/RepoRipple@de0ba8ff4059928cdf59481d4fb77989da482735 # v0.6.0 release source
        with:
          base: ${{ github.event.pull_request.base.sha }}
      - uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
        with:
          name: reporipple-impact
          path: |
            ${{ steps.reporipple.outputs.markdown-report }}
            ${{ steps.reporipple.outputs.json-report }}
            ${{ steps.reporipple.outputs.html-report }}

The Action accepts path, base, max-depth, fail-on, and reports-directory. Version 0.6.0 outputs markdown-report, json-report, html-report, and risk-level. The example is pinned to the release source commit rather than a mutable tag. When upgrading, replace it with the commit SHA for the desired release. Use fetch-depth: 0 so the requested base revision is available.

This repository's own pull-request workflow demonstrates a fork-safe commenting design. Analysis runs with read-only contents access. A separate job updates one stable RepoRipple comment only for branches in the same repository; fork pull requests keep their report in the job summary and downloadable artifact without receiving a write-capable token.

RepoRipple can expose the same deterministic analysis to Codex, OpenCode, Cursor, Claude Code, and VS Code over a local, read-only MCP server. From a source checkout:

uv run reporipple mcp --root /path/to/allowed/workspace

Launch the published package without a clone:

uvx reporipple mcp --root /path/to/allowed/workspace

The server provides forecast_change for proposed edits and analyze_worktree for actual Git changes. See the MCP setup guide for client-specific commands and configuration.

The architecture guide explains how the CLI, GitHub Action, MCP server, and optional explanation layer share one deterministic analysis engine.

# RepoRipple impact report: checkout-service

**Risk:** `high` · **Graph:** 84 files / 137 dependency edges

## Changed files
- `src/billing/store.py`

## Impacted files
- `src/billing/service.py` — distance 1 via `store.py` → `service.py`
- `src/api/checkout.py` — distance 2 via `store.py` → `service.py` → `checkout.py`

## Suggested verification
- `tests/test_billing_service.py`
- `tests/test_checkout.py`

1. Discovers Python, JavaScript, and TypeScript source files.
2. Extracts Python imports with the standard-library AST and JS/TS imports with a focused parser.
3. Resolves JS/TS dependencies in a deterministic order: relative paths, the nearest tsconfig.json or jsconfig.json paths mappings, its baseUrl, and then npm/Yarn workspace packages declared by the root package.json.
4. Records which JS/TS files are governed by project configuration and package manifests.
5. Builds forward and reverse dependency relationships across source and metadata.
6. Gets added, modified, deleted, and renamed files from an explicit base revision or the local worktree.
7. Retains pre-change source and resolver metadata for deleted and renamed files so their remaining dependents can still be traced.
8. Traverses dependents to a configurable depth.
9. Adds test recommendations, documentation reminders, and deterministic risk signals.

Configuration-only changes now produce a concrete blast radius. Each JS/TS source is linked to its nearest tsconfig.json or jsconfig.json, that configuration's local extends chain, and its nearest live package.json. When Git supplies a deleted or renamed manifest, RepoRipple retains the nearest historical package boundary below the current live boundary as a second impact target without letting it shadow that live ancestor. A deleted project config likewise remains an impact target but cannot supply aliases or baseUrl for the current tree, including when a live config still names it in extends. For extensionless extends, a live fallback is preferred while a deleted candidate remains direct historical metadata. Deleted config chains are not traversed; this prevents stale history from consuming work needed for the live graph. Successfully matched workspace imports are additionally linked to the root workspace manifest and the imported package's manifest; unrelated external imports are not. Changing any of that metadata reports directly governed or dependent source files first, then continues through their downstream importers. Metadata remains outside the source-file inventory, while its deterministic relationships are included in the reported graph edge count.

Workspace resolution is local and source-aware. RepoRipple matches imports to explicitly declared package.json workspaces by package name, respects their public export boundaries, and follows entry points that resolve to source files in the repository. Workspace globs support recursive patterns, bounded brace alternatives and ranges, and exclusions. Conditional exports distinguish ES module and CommonJS entry points: .mts/.mjs use the import profile, .cts/.cjs use the require profile, and ambiguous TypeScript/JavaScript files conservatively inspect both. RepoRipple also keeps applicable custom conditions declared before a syntax or default fallback so it does not silently miss browser, platform, or source-only dependency edges.

Custom conditions are evaluated without assuming one runtime condition set. Nested or repeated custom conditions can therefore produce conservative extra edges that no single runtime would select; impact analysis favors that visible over-approximation over silently missing a dependent.

RepoRipple does not inspect installed node_modules, contact a registry, or run a package manager. pnpm-workspace.yaml, external packages, and workspace entry points that exist only as ignored or missing build output are not currently resolved. Pathological workspace glob sets that exceed the resolver's fixed safety budgets fail closed instead of consuming unbounded scan time. The current limits are 512 effective patterns, 2,048 aggregate path components across unique patterns, 64 KiB of raw patterns, and 256 KiB after brace expansion. Project-config traversal is limited to 16 levels, 64 loaded config nodes per root resolution, and 64 config metadata dependencies per source. A repository scan stores at most 100,000 unique source-to-metadata relationships. Scan warnings identify every exceeded limit, and completed top-level config resolutions are reused without bypassing cycle or depth checks inside an extends traversal. Loaded live config paths have priority in the per-source metadata cap; skipped live targets are next, and direct historical targets use the remaining capacity. Changes to tsconfig.json, jsconfig.json, and package manifests are both flagged as configuration risk and traced through their governed or dependent JS/TS files. Deleted and renamed metadata uses the same bounded in-memory Git overlay as deleted source. For an in-place workspace-manifest edit that removes an old package name entirely, RepoRipple can still report files owned by that package, but cannot reconstruct importers of the removed name without a second full historical scan.

The core intentionally avoids LLM scoring. Results stay reproducible, private, fast, and usable in CI. Optional AI explanations can be layered over the JSON output without making correctness depend on a model.

AI explanation is off by default and is activated only by --explain. The deterministic risk level, evidence, report, and --fail-on exit status remain authoritative even if OpenRouter is unavailable or returns an invalid response.

First inspect the exact key-free body that would leave your machine:

reporipple /path/to/repository --base main --explain-preview

The preview does not read an API key and does not make a network request. When you are satisfied with the payload, provide the key through the environment and explicitly request an explanation:

export OPENROUTER_API_KEY="your-key"
reporipple /path/to/repository --base main --explain

There is no command-line key option and RepoRipple does not read a key from a file. The integration uses only the Python standard library and applies all of these controls to every request:

• Repository names, paths, source, diffs, scanner-warning text, and user identifiers are not transmitted.
• Files and risk signals become opaque evidence IDs such as E001; returned IDs are schema-validated and mapped back to labels locally.
• The response must satisfy a strict JSON Schema, and unknown evidence IDs or extra fields are rejected.
• OpenRouter routing requires zero data retention (zdr: true), denies data collection, and selects only providers that support every requested parameter.
• Provider prices are capped at $0.25 per million input tokens and $1 per million output tokens, with a 450-token response limit.
• Requests have a 10-second timeout and retry at most once after an explicit retryable HTTP response; ambiguous network timeouts are not retried.
• Token usage and reported request cost are included in Markdown and JSON output.

The pinned default model is openai/gpt-6-luna. Override it with --explain-model or REPORIPPLE_OPENROUTER_MODEL; privacy, schema, and price controls remain mandatory. Privacy filters may reduce provider availability, and the request fails closed when no eligible provider satisfies them. OpenRouter's ZDR policy prevents eligible providers from retaining prompts and responses after processing; it does not mean the request stays on your machine. See OpenRouter's structured outputs, ZDR controls, and provider routing controls.

reporipple [PATH]
  --version             show the installed RepoRipple version and exit
  --base REVISION       compare REVISION...HEAD
  --changed PATH        analyze an explicit path; repeat as needed
  --max-depth N         reverse dependency traversal depth (default: 4)
  --format markdown|json
  --output FILE
  --fail-on medium|high
  --explain             request an optional privacy-filtered OpenRouter explanation
  --explain-preview     show the exact sanitized request without sending it
  --explain-model NAME  use a pinned provider/model slug for either explanation mode

reporipple mcp
  --root PATH            restrict tool access to PATH and its descendants

When no base or explicit path is supplied, RepoRipple analyzes staged, unstaged, and untracked files. On a clean checkout it falls back to the latest commit.

--explain and --explain-preview are mutually exclusive. With JSON output, either flag returns a wrapper containing deterministic_report plus the explanation or preview; ordinary JSON output is unchanged.

uv venv --python 3.12
uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/ruff check .
.venv/bin/pytest
uv build

See CONTRIBUTING.md for RepoRipple's design contracts, test map, and pull-request expectations.

RepoRipple's deterministic analysis runs locally and does not transmit repository contents. It invokes Git using argument arrays rather than a shell and reads only supported, non-symlinked source files outside common generated and dependency directories. The MCP server constrains every repository request to its configured --root and exposes read-only tools only. Network access occurs only when --explain is present; --explain-preview remains fully local so the outbound body can be audited first.

Report suspected vulnerabilities through GitHub private vulnerability reporting; see SECURITY.md for scope and reporting guidance.

MIT