brevio-core

Catches the mistakes coding agents repeat before they merge: a pull-request gate with no model call and no API key.

Save this as .github/workflows/brevio-core.yml:

name: brevio-core
on:
  pull_request:
  merge_group:
permissions:
  contents: read
  pull-requests: write     # the comment on the pull request
  security-events: write   # the upload to code scanning
jobs:
  brevio-core:
    name: brevio-core
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@fbc6f3992d24b796d5a048ff273f7fcc4a7b6c09 # v5.1.0
        with:
          fetch-depth: 0
      - uses: Brevio-Ventures/brevio-core@v0.9.1
        with:
          sarif: true

It catches what review keeps finding in agent-written code:

• secrets written into source, eval on input, and any casts on data from outside;
• type errors silenced with @ts-ignore, and exceptions caught and dropped;
• fixed sleeps in end-to-end tests, images without alt text, and click handlers a keyboard cannot reach.

No API key, no model. scan, the agent hook and the MCP server use no network and send nothing.

These are line-level pattern checks on the lines a change adds, not AST or dataflow analysis. The catalogue has 123 rules: brevio-core scan checks 13 scanned rules, and the other 110 are documented rules, written down for review and for agents, with no scanner in this package yet. Every finding links to its rule page: what it catches, why, a bad and a good example.

Run it without installing

Version 0.9.1, Python 3.10 or newer, no runtime dependencies. In a repository, scan the lines your change adds:

uvx brevio-core@0.9.1 scan --diff
uvx brevio-core@0.9.1 score

pipx run --spec brevio-core==0.9.1 brevio-core scan --diff

Or install it, then run it the same way:

pip install brevio-core==0.9.1

brevio-core scan --diff

Or download brevio-core.pyz from the GitHub release, check where it was built, and run it with any python3:

gh attestation verify brevio-core.pyz --repo Brevio-Ventures/brevio-core
python3 brevio-core.pyz scan --diff

To scan whole files instead, name them, or nothing for the current folder:

brevio-core scan

Set it up in one step

brevio-core init --dry-run

--dry-run shows, as a diff, what brevio-core init would write, and writes nothing. init writes a pull-request workflow like the one above (pinned to this release), a block of agent rules in AGENTS.md and a baseline, so the first local run is quiet. --with precommit,hook,mcp adds the pre-commit entry, the agent hook and the MCP server config. It asks before it writes (--yes skips the question), never rewrites a file you have, and records what it wrote in .brevio/init.json; brevio-core init --undo removes what it wrote and you have not changed since.

Adopt it on a repository that already has findings

brevio-core baseline                 # record the findings the repository has now in .brevio/baseline.json
brevio-core scan                     # report only the findings that are not in the baseline
brevio-core scan --fail-on none      # report them but never fail, for a trial week
brevio-core scan --fail-on all       # leave the baseline out: every finding fails

A baseline entry names the file, the rule and a hash of the line's text, not its line number, so moving a line does not make it new. Commit the file. Run baseline again after a fix and the fixed entries drop out. The file holds one entry per line, sorted, so branches that add entries merge cleanly.

Silence one finding, with a reason

// brevio-core: ignore R9.2 -- the generated client has no type for this field
// @ts-ignore

The comment goes on the line or the line above, and silences that rule there. Without -- reason it does not count, and the scan warns about it and about a rule ID that does not exist. The summary line counts what was silenced, and SARIF carries each one as a suppression with its reason, which code scanning shows.

Only a person can silence the gate. The agent hook and the MCP tools never honour an ignore comment (--forbid-ignores), and a diff scan reports each ignore comment and baseline entry the change adds as a finding, "new suppression: needs a human review", until a reviewer allows it with --allow-new-suppressions.

A report for the pull request

brevio-core scan --diff origin/main --format markdown prints the findings as a Markdown table of rule, file and line, and why, for a job summary or a pull request comment. --rules-url BASE links each rule ID to its page at BASE followed by the ID. Paths and lines from the repository are escaped, so a file name cannot add a mention or a link.

Configure it for a team

Put a brevio-core.toml at the root of the repository, and another in any folder that needs its own settings. For a file, the files from the root down to its folder apply in that order, so the nearest one wins.

extends = "policy/brevio-core.toml"   # an org policy, read first (see below)
select = ["R9", "R13.1"]              # the rules that run: an ID, a group or a glob (default: every rule)
ignore = ["R6.3"]                     # rules that do not run
exclude = ["generated/**", "*.pb.ts"] # paths not scanned, relative to this file's folder
level = "warn"                        # the level of every rule (default: error)

[rules."R13.1"]
level = "error"                       # the level of one rule

[[overrides]]
paths = ["legacy/**"]                 # relative to this file's folder
ignore = ["R9.1"]                     # rules that do not run on these paths
rules = ["R9"]                        # the rules the level applies to (default: every rule)
level = "warn"

A nearer file replaces select and ignore; exclude and [[overrides]] add up; for a level, the last setting that names the rule wins. An unknown key, a wrong value or a broken file stops the scan as a usage error, and a rule pattern that matches no rule is a warning. Write a rule ID in quotes as a table name, as in [rules."R13.1"].

Every scan already skips code a project keeps but did not write: dependencies and build output (node_modules/, dist/, build/, .next/, virtual environments), vendored folders (vendor/, third_party/, third-party/), minified files (*.min.js, *.min.mjs, *.min.cjs, *.min.css), bundle chunks (*.chunk.js, *.bundle.js, chunk-*.js, and names that end in a hex hash, such as index-<hash>.js) and any file whose first two lines are longer than five hundred characters. A file named on the command line is scanned as asked. No rule reads a TypeScript declaration file (*.d.ts, *.d.mts, *.d.cts), named or not: it holds types, no code that runs. Exclude more, such as generated code or a third-party folder with another name, with exclude.

brevio-core config src/app.ts shows what applies to a file or folder: the files it comes from, whether it is excluded, and each rule that runs there with its level and status (--json for a tool).

brevio-core scan --fail-on error (in the Action, fail-on: error) fails only on error-level findings and reports the warn-level ones; the default, --fail-on new, fails on every new finding whatever its level. A preview rule never fails a scan (see RULES-STABILITY.md). A config can turn rules off, so a diff scan reports a change that adds or edits a brevio-core.toml as a new suppression, as it does an ignore comment.

An org policy with extends

extends reads a policy file before the file that names it; the policy's paths are relative to that file's folder.

• extends = "policy/brevio-core.toml": a file vendored into the repository, pinned by your own history.
• extends = "policy@<commit SHA>": a folder that is a git checkout, such as a submodule, at that commit.
• extends = "<git URL>@<commit SHA>": the submodule whose URL in .gitmodules is that URL, at that commit.

extends reads only files already in your checkout: brevio-core does not fetch a policy (scan, the agent hook and the MCP server use no network and send nothing). When the checkout is missing or at another commit, the scan stops and says so, and git submodule update --init fixes it. To move to a new policy, update the submodule and the SHA in one pull request.

Roll it out

1. Observe. Run brevio-core scan --fail-on none in CI for a week: everything is reported, nothing fails.
2. Warn on new. Record what is there with brevio-core baseline, set level = "warn" in brevio-core.toml and run --fail-on error: new findings show on the pull request and nothing blocks.
3. Block on new. Raise the rules you trust to level = "error", one by one or with [[overrides]] for the paths that are ready. Their new findings now fail the check.
4. Tighten. Fix baseline entries and run brevio-core baseline again so they drop out, raise the last rules to error, then go back to the default --fail-on new.

GitHub Enterprise Server

It works with the .pyz from the GitHub release, vendored into your repository, and with a PyPI mirror for pip. This is untested: no one has run it on GitHub Enterprise Server yet.

Use it from your agent

Three ways, all local: no model call and no API key.

Wire the scan into your coding agents. In the project folder:

brevio-core hook install

It finds the coding agents set up in the folder or on the machine and attaches the scan to each, so every edit is checked against the rules. It also installs the agent skill (skills/brevio-core/SKILL.md), which tells the agent when to run the commands and how to read a finding. It replaces any older skill text and a second run changes nothing.

Run it as an MCP server. Add this to the mcpServers block of any client that supports MCP servers over stdio (the settings file differs by client, the shape does not):

{
  "mcpServers": {
    "brevio-core": {
      "command": "brevio-core",
      "args": ["mcp"]
    }
  }
}

Without a global install, use "command": "uvx" and "args": ["brevio-core@0.9.1", "mcp"]. Start the client in the project folder: the server reads that folder and nothing else. The tools are scan (paths, or a diff base), explain (a rule id, returns the rule page), list_rules and learn (compiles a fixed defect into a guard).

Learn from a fix. brevio-core learn turns the last fix into a guard. See docs/agent-tools.md for learn, hook install, gain, add and mcp.

What ESLint, Biome and typescript-eslint don't cover

Linters check how code is written: syntax, types and style. Several mistakes agents make are valid code to them, and are what these rules look for. From the rule pages:

Compared with ESLint, Semgrep and CodeQL

brevio-core complements them; keep them.

• Where they win. They parse the code, so they see types, the syntax tree and, in Semgrep and CodeQL, data flowing across lines and files. They have far more rules and editor integration. brevio-core matches patterns line by line on the lines you add, so it misses a mistake spread over several lines and can flag a line that is fine in context.
• What this adds. Rules picked from the mistakes coding agents repeat, an agent hook that scans after each edit, an MCP server an agent can call, learn to turn one fix into a local guard, and a page per rule written for people and agents. Nothing to configure and no dependency.

Use

brevio-core scan src/                 # files or folders
brevio-core scan --diff origin/main   # only the lines added against a base
brevio-core scan --diff --sarif out.sarif

The exit code is 0 when nothing is found, 1 when there are findings and 2 for a usage error.

Additional commands for local workflows:

• brevio-core score: compute the repository's AI-Code Readiness Score and letter grade (--badge, --markdown, --json).
• brevio-core check: pre-PR flight check (rules, test custody, test imports, conflict markers and complexity).
• brevio-core brief: file-scoped rules & trap memory briefing before agent edits.
• brevio-core init --personal: personal stealth mode via .git/info/exclude without touching tracked files.
• brevio-core aeo --generate-llms: score built HTML pages and generate missing llms.txt.

Fix what has one correct fix

brevio-core scan --fix --dry-run    # print the fixes as a unified diff, write nothing
brevio-core scan --fix              # rewrite the files in place, then scan again
brevio-core scan --fix --diff main  # fix only the lines your change adds

A rule gets a fixer only when it has one correct rewrite, applied by code, never by a model: a @ts-ignore comment becomes @ts-expect-error, and a bare Python except: becomes except BaseException:, which catches the same things. Every other finding, a hardcoded secret above all, is reported and left for you. The command prints what it fixed and what remains, and fails while anything remains. Running it twice changes nothing the second time. The MCP server has a fix tool that returns the same diff and writes only when asked.

GitHub Action

The workflow at the top runs the scan on the lines a pull request adds. It writes the report to the job summary and marks each finding on the changed file. With pull-requests: write it keeps one comment on the pull request up to date while there are findings, and marks it resolved once they are gone; a pull request from a fork gets the summary only. With sarif: true and security-events: write it uploads the findings to code scanning. It needs no secret and no API key. The tag v0.9.1 runs exactly brevio-core 0.9.1. To follow every release instead, write uses: Brevio-Ventures/brevio-core@v0; to pin a commit, see SECURITY-REVIEW.md. It runs on pull_request, merge_group and push, and refuses any other event.

pre-commit

repos:
  - repo: https://github.com/Brevio-Ventures/brevio-core
    rev: v0.9.1
    hooks:
      - id: brevio-core

Rules

Each rule has a page under docs/rules: what it catches, why, and what to write instead.

What it does not claim

This release publishes no benchmark. When a reproducible one exists, its method and data will be published with the number.

Questions and support

Ask questions and propose rules in GitHub Discussions; report bugs and false positives as issues, with the forms. Support is best effort, with a first response within a week (SUPPORT.md).

Project

• Security review: what it runs, reads and writes, for a security team
• Contributing (sign-off required), with how this repository is exported and how contributions come back
• Governance: what stays open, and how maintainers are added; the roadmap for the next release
• Security
• Privacy
• Code of conduct
• Trademark policy

Made by Brevio.