Blast Radius
aharwelik/blast-radius
Avalia risco e alcance de diffs antes de fazer merge
Ver no GitHubAnalisa diffs de código antes do merge para detectar risco e alcance da mudança. Identifica secrets commitados, mudanças em autenticação, testes deletados, alterações em CI/CD e dependências, funcionando como gate pré-merge offline ou skill Claude Code com receipt de operação. Usa apenas stdlib.
Instalação
npx convez-ai aharwelik/blast-radius1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Bloquear merges com secrets ou mudanças críticas em autenticação
- Validar qualidade de PRs automaticamente em CI/CD
- Detectar deletions de testes que passam falsamente em builds
- Analisar impacto de mudanças em dependências e infraestrutura
Veredito

Ferramenta sólida e prática para DevSecOps que mede risco real antes de merge, não tamanho da mudança. Implementa bem a detecção de secrets e mudanças críticas com zero dependências, ideal como gate automático em pipelines.
README
A tradução para o português deste README ainda está na fila.
blast-radius
Score the reach and risk of a code diff before you merge it. Flags diffs that touch auth, commit secrets, delete tests, change CI/CD, or widen dependencies -- offline, as a pre-merge gate or a Claude Code skill.
Keywords: code review, AI code review, diff risk, pre-merge gate, secret detection, CI security, blast radius, change risk, DevSecOps, Claude Code, Codex, agent guardrails, pull request automation.
The problem
An agent stops when the diff looks done, and a plausible-looking diff is the most common way a real problem slips through (the trust-then-verify gap). The danger of a change is not its size. A two-line edit to authentication, an accidentally committed key, or a quietly deleted test outranks a large mechanical refactor. Reviewers and gates need the reach measured, not assumed.
blast-radius reads a unified diff and reports which high-reach signals fired, the overall scope, and a severity, so review effort matches actual risk and an unattended run can refuse to merge a dangerous change on its own.
What it flags
- Secrets committed in the diff (AWS keys, private keys, GitHub tokens,
api_key/token/passwordassignments) -- critical. - Authentication / session / OAuth paths -- high.
- Deleted or gutted tests (removing tests to make a build pass) -- high.
- CI/CD, Dockerfiles, deploy changes -- high.
- Database migrations and infrastructure as code -- high.
- Dependency manifests and lockfiles -- medium (supply-chain reach).
- Large changes (many files or lines) raise the scope to service or global.
Install
No dependencies. Python 3.10+.
git clone https://github.com/aharwelik/blast-radius
cd blast-radius
Use it
git diff | python -m blastradius score
# or a saved diff:
python -m blastradius score examples/risky.diff
Output on the bundled risky example:
blast-radius: severity=critical scope=service
files=3 dirs=3 lines=9
high-reach signals:
- secret committed in diff (src/auth/login.py)
- touches authentication (src/auth/login.py)
- deletes a test file (tests/test_login.py)
- changes a dependency manifest or lockfile (requirements.txt)
Exit code is 1 when severity meets the --fail-on threshold (default high),
so it gates a merge. Verify a diff as a change receipt:
git diff | python -m blastradius score --receipt | python -m pco verify -
Use it as a Claude Code skill
mkdir -p ~/.claude/skills/blast-radius
cp skill/SKILL.md ~/.claude/skills/blast-radius/SKILL.md
Then ask Claude to "score the blast radius of this diff before merging."
Theory
This implements the Blast-Radius declaration of the proof-carrying-ops model: the reach of a change must be measured before the change is accepted, because the same request can produce a local or a global edit and the cost of accepting an unmeasured global one is not proportional to the agent's average quality.
Honesty boundary
The signals are transparent rules, not a verdict that the change is wrong. A high score means "look closely", not "reject"; a low score is not a correctness guarantee. Pair it with tests. Example diffs use documentation-only placeholder secrets.
Design
Standard library only, tested on Python 3.10 through 3.13, no-emoji guard in CI.
Author
Anthony Harwelik -- aharwelik@gmail.com -- https://github.com/aharwelik
License
MIT. See LICENSE.