Local PR review tool that orchestrates Claude Code CLI agents over git diffs.
Project description
PR Sentinel
Local pull-request review tool. Reads a git diff, runs four specialized review agents over it via the Claude Code CLI, and emits a structured JSON + Markdown report so you can fix issues before raising the PR.
No API keys. No hosted services. PR Sentinel shells out to claude -p and uses your existing Claude Code authentication.
How it works
git diff main...HEAD
│
▼
[ diff_parser ] ── filters built-in noise + user skip patterns
│ (--skip-files, .prsentinelignore)
▼
[ chunker ] ── packs files into ≤100k-char chunks (hybrid batching)
│
▼
[ router ] ── per chunk, decides which agents are relevant for those file types
│
▼
[ orchestrator ] ── runs (agent × chunk) tasks in a bounded thread pool
│
├── Security Agent (claude -p prompts/security.md)
├── Code Quality Agent (claude -p prompts/quality.md)
├── Performance Agent (claude -p prompts/performance.md)
└── Testing Agent (claude -p prompts/testing.md)
│
▼
[ cache ] ── sha256(model + prompt) → response, on disk
│
▼
[ Summary Agent ] ── single claude call that dedupes/merges findings across agents
│ (prompts/summary.md; falls back to raw findings on failure)
▼
[ report_generator ] ── merges findings, computes risk level
│
▼
reports/report.json + reports/review-report.md
File-type routing
Not every agent has something useful to say about every file. PR Sentinel skips agents that have nothing meaningful to contribute to a chunk's file types, reducing token usage with no loss in review quality.
| File type | Security | Quality | Performance | Testing |
|---|---|---|---|---|
*.cs (source) |
✅ | ✅ | ✅ | ✅ |
*Test*.cs, *Spec.cs, *Fixture.cs |
✅ | ✅ | — | ✅ |
*.cshtml, *.razor, *.html |
✅ | ✅ | — | — |
*.js, *.ts |
✅ | ✅ | ✅ | ✅ |
*.css, *.scss, *.less |
— | ✅ | — | — |
*.svg |
✅ | — | — | — |
*.csproj, *.props, *.targets |
✅ | ✅ | — | — |
*.sln |
— | — | — | — |
appsettings*.json, *.yml, *.yaml, *.xml, *.json |
✅ | ✅ | — | — |
Dockerfile, *.bicep, *.tf |
✅ | ✅ | — | — |
*.sql |
✅ | ✅ | ✅ | — |
*.csv, *.tsv, *.resx |
✅ | — | — | — |
*.md, *.txt, *.http |
✅ | — | — | — |
| Unknown extension | ✅ | ✅ | ✅ | ✅ |
Three rules behind the table:
- Security runs on almost everything — secrets and PII appear in docs, config files, and data files.
- Performance and Testing only run on executable code.
- Unknown extensions always get all four agents — nothing is silently skipped.
If a chunk contains mixed file types (e.g. a .cs file and a .css file together), the agents for the union of both types are run. Binary files (*.png, *.dll, *.zip, etc.) are dropped entirely before routing — their diffs are unreadable.
Requirements
- Python 3.11+
- Claude Code CLI installed and authenticated.
claude --versionmust work from your shell. - Git, if you want to review live branches (not required for
--diffmode).
Install
From PyPI:
pip install pr-sentinel
From source (for development):
git clone https://github.com/kishor2004reddy/Pull-Request-Sentinel-CLI.git
cd Pull-Request-Sentinel-CLI
pip install -e .[dev]
Verify:
pr-sentinel --help
pr-sentinel agents
Quickstart
Review your current branch against main:
pr-sentinel review --base main
Or review a saved diff file (works without a git repo):
git diff main...HEAD > my.diff
pr-sentinel review --diff my.diff --out reports
Open reports/review-report.md.
Commands
pr-sentinel review
| Flag | Default | Description |
|---|---|---|
--base |
main |
Branch to diff against. Runs git diff <base>...<head>. |
--head |
HEAD |
Source branch/ref to review. Use with --base to diff arbitrary refs without checking them out. |
--diff PATH |
— | Review a saved diff file instead of running git. Mutually exclusive with --staged. |
--staged |
off | Review staged changes (git diff --cached). |
--repo PATH |
cwd | Path to the git repository to review. Ignored when --diff is used. |
--agents |
security,quality,performance,testing |
Comma-separated agents to run. |
--out |
./reports |
Output directory. |
--format |
both |
json, markdown, or both. |
--max-file-size |
20000 |
Per-file diff size cap (chars). Larger files get truncated with a marker. |
--chunk-budget |
100000 |
Max combined diff size per Claude call before chunking kicks in. |
--model |
haiku |
Claude model to use. Shortcuts: sonnet, opus, haiku. Or pass a full model ID such as claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001. Forwarded to claude --model. |
--max-parallel |
8 |
Max concurrent claude calls across all (agent, chunk) pairs. |
--timeout |
600 |
Per-call timeout in seconds for each claude subprocess. |
--no-cache |
off | Bypass the response cache for this run. Successful responses are still written to the cache. |
--skip-files |
— | Comma-separated glob patterns to skip on top of built-in noise filters (e.g. "*.lock,vendor/**,fixtures/*.json"). Combines with .prsentinelignore if present. |
pr-sentinel agents
Lists available agents and their implementation status.
pr-sentinel cache
Inspect and manage the on-disk response cache.
| Subcommand | Description |
|---|---|
cache size |
Show cache location, entry count, and disk usage. |
cache clear |
Wipe the entire cache (prompts for confirmation). |
cache prune --older-than 30d |
Delete entries older than the given age. Supports s/m/h/d suffixes. Add --dry-run to preview. |
The cache lives at ~/.pr-sentinel/cache/ by default. Override with the PR_SENTINEL_CACHE_DIR environment variable. Keys are sha256(model + prompt), so changing the model or any prompt content invalidates the entry automatically.
Auto-pruning. Every pr-sentinel review run silently drops cache entries older than 90 days before doing any work. No flag, no output — it just keeps the cache from growing unbounded over time. The threshold is set via AUTO_PRUNE_AGE_DAYS in config.py. Use cache prune --older-than ... for manual prunes at a different age.
Skipping files
PR Sentinel always filters built-in noise (lock files, build/dist dirs, minified, generated). To skip additional files — large fixtures, vendored code, anything that wastes tokens without adding review value — use either of these:
Per-run flag, good for one-offs:
pr-sentinel review --skip-files "vendor/**,fixtures/*.json,*.snap"
Repo-wide ignore file, good for project defaults — drop a .prsentinelignore at the repo root:
# generated docs
docs/api/**
# huge test fixtures
tests/fixtures/large/**
# vendored deps
vendor/**
Syntax: one glob per line, # for comments, blank lines ignored. Patterns use fnmatch (same as the built-in noise filter), matched against both the full path and the basename. The two sources combine — flag patterns are appended to the ignore file's patterns. Skipped files are listed in the "Skipped noise file(s)" panel at the top of the run output.
Examples
Review current branch vs main:
pr-sentinel review --base main
Run only the security agent:
pr-sentinel review --base main --agents security
Review what's currently staged:
pr-sentinel review --staged
Review a feature branch without checking it out:
pr-sentinel review --base main --head feature/new-auth
Use a stronger model for higher-stakes reviews:
pr-sentinel review --base main --model opus
Force a fresh run, ignoring cached responses:
pr-sentinel review --base main --no-cache
Skip extra files for this run only:
pr-sentinel review --base main --skip-files "vendor/**,*.snap"
Prune cache entries older than a week:
pr-sentinel cache prune --older-than 7d
Report structure
The markdown report always emits these five sections in this fixed order, regardless of findings:
- Summary — risk level, source, branch, timestamp, agents, finding counts
- Merge Verdict — deterministic verdict driven by risk level (not a separate Claude call)
- Key Findings — top blocking issues (High + Medium)
- Key Recommendations — deduplicated fixes
- All Findings — per-agent summary table + full per-issue detail
The JSON report contains the same data in a single structured object suitable for piping into other tools. Key fields:
riskLevel—High/Medium/Low/None/Unknown(see table below)coverageComplete—falseif any agent failed during the runagentsExecuted,failedAgents— who ran and who didn'tfindings— flat list, sorted by severity then filerawFindingCount— only present when the Summary Agent ran; how many findings existed before dedup
Risk levels
| Level | Trigger |
|---|---|
| High | Any High-severity finding, or 5+ Medium-severity findings |
| Medium | One or more Medium-severity findings (and fewer than 5) |
| Low | Only Low-severity findings |
| None | No findings across all executed agents (full coverage) |
| Unknown | All executed agents failed, or some agents failed and the rest found nothing (can't call it clean with incomplete coverage) |
Architecture
src/pr_sentinel/
├── cli.py # Click entrypoint + Rich UI
├── config.py # tunable defaults + shared constants (single source of truth)
├── git_diff.py # git rev-parse, git diff, --staged
├── diff_parser.py # per-file splitting + noise filter (+ --skip-files / .prsentinelignore) + truncation
├── chunker.py # greedy packer to keep prompts under chunk-budget
├── claude_runner.py # subprocess(claude -p) + JSON extraction + 1 retry
├── orchestrator.py # parallel (agent, chunk) execution via ThreadPoolExecutor
├── router.py # file-type routing table — decides which agents run per chunk
├── cache.py # sha256-keyed disk cache + 90-day auto-prune
├── report_generator.py # build_report + JSON/Markdown writers
├── agents/
│ ├── base.py # BaseAgent: load prompt, chunk, call, validate
│ ├── security_agent.py
│ ├── quality_agent.py
│ ├── performance_agent.py
│ ├── testing_agent.py
│ └── summary_agent.py # post-processes findings: dedupe/merge across agents
└── prompts/
├── security.md
├── quality.md
├── performance.md
├── testing.md
└── summary.md
tests/
├── test_diff_parser.py
├── test_chunker.py
└── test_report_generator.py
Running tests
pip install -e .[dev]
pytest -q
Troubleshooting
claude CLI not found on PATH — install Claude Code and confirm claude --version works in the same shell where you run pr-sentinel.
claude returned non-JSON output after retry — Claude occasionally returns prose instead of JSON. The runner retries once; if it still fails, that chunk's findings are dropped and the agent is marked failed for the run. Re-running usually succeeds.
Slow runs — large diffs trigger chunking. Each chunk is one Claude call per agent. Reduce scope with --agents security if you only want one perspective, or with --max-file-size to truncate huge files. The cache amortizes repeat runs against the same diff.
Lock files / minified files showing up — they shouldn't. If they do, add the pattern to NOISE_PATTERNS in config.py, or skip on a per-project basis with .prsentinelignore (see Skipping files).
Want to skip a non-noise file (huge fixture, vendored dep, generated snapshot)? Use --skip-files "pat1,pat2" for a one-off, or commit a .prsentinelignore for project-wide defaults.
Notes & limitations
- The Performance Agent only sees the diff. It can flag pattern-level issues (N+1 queries, sync-over-async) but cannot reason about runtime behavior or system load.
lineHintis approximate — unified diffs have hunk headers, not absolute line numbers. The prompt asks Claude for a description of the location rather than a hallucinated number.- Agents cannot read other files in the repo. Review depth is limited to what's visible in the diff itself.
- If any one chunk fails for an agent (timeout, exit code, unparseable JSON after retry), that agent is marked failed for the run and its partial findings are discarded. Other agents continue.
- The Summary Agent makes one additional
claudecall per run to deduplicate and merge findings across the four review agents. If it fails (timeout, bad JSON), the report falls back to the raw findings — nothing is lost.
License
MIT
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file pr_sentinel-0.1.5.tar.gz.
File metadata
- Download URL: pr_sentinel-0.1.5.tar.gz
- Upload date:
- Size: 42.1 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
a4350b3a5aadd1d3175de536865df1209fea127c60a4df3c6f3b153171dbf0ab
|
|
| MD5 |
31ddd76ef39fd9c7fc31d59aafaf9bbe
|
|
| BLAKE2b-256 |
b417633938b71abeabf262fba5614fe59833d7c918eaad4229cdf58904da8c9a
|
File details
Details for the file pr_sentinel-0.1.5-py3-none-any.whl.
File metadata
- Download URL: pr_sentinel-0.1.5-py3-none-any.whl
- Upload date:
- Size: 42.4 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.11.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
29c2f5d59ed1e3390d78606c02a8cdda41dbafb6c1cda8a5cc023e7d1bdc96cf
|
|
| MD5 |
d1756f96c8fc050e3e12afd7883127cf
|
|
| BLAKE2b-256 |
c9d1c73036fb38cdb0131c1133c1928b6ebe8d30bc614415b09a110194f0fa1d
|
