Changelog

0.7.0 (unreleased)

brevio-core scan now checks 13 scanned rules of the 123 in the catalogue (110 documented): R5.5 flags subprocess calls with shell=True and os.system calls with interpolated strings in Python; R10.1 flags single-line requests calls without timeout= in Python; R16.2 flags "*" and "latest" dependency ranges in package.json; and R18.2 flags single-line target="_blank" anchor tags without rel="noopener noreferrer".
R2.8 also flags raw SQL string interpolation in .execute(f"SELECT ...") and .$queryRawUnsafe template literals, as well as hidden directional and Unicode tag control characters.
R13.1 also flags secret-named environment variables exposed to client bundles (NEXT_PUBLIC_, VITE_, REACT_APP_).
brevio-core scan --fix and the MCP fix tool now rewrite target="_blank" anchor tags that have no rel= attribute to add rel="noopener noreferrer" (R18.2).
Every rule page in docs/rules now links to the interactive rule playground on brevio.pro.

brevio-core baseline records the findings a repository has today in .brevio/baseline.json, so a scan reports only new ones. An entry names the file, the rule and a hash of the line, so moving a line does not make it new.
scan --fail-on chooses what fails: new (the default), all, none for a trial, or error for error-level findings only.
brevio-core init sets a repository up in one step: the pull-request workflow, a block of agent rules in AGENTS.md and a baseline, with --with precommit,hook,mcp for more. It shows a diff and asks first, never rewrites a file you have, and init --undo takes back what it wrote.

An ignore comment with a reason (brevio-core: ignore R9.2 -- reason) silences one finding in a scan a person runs. Without a reason it does not count.
The agent hook and the MCP server never honour an ignore comment, and a diff scan reports every ignore comment, baseline entry and config change it adds as a new suppression until a reviewer allows it.
SARIF keeps each silenced finding as a suppression with its reason.

brevio-core.toml at the root and in any folder chooses the rules, leaves paths out, and sets warn or error per rule and per path. extends reads an organisation policy from a file in the repository or a checkout pinned to a commit; it never fetches one.
brevio-core config shows what applies to a file or folder.
The README has a rollout guide: observe, warn on new findings, block them, then tighten.

Every rule has a status: stable, preview (reports, never fails) or deprecated (reports nothing, keeps its ID).
A rule ID is never removed, reused or renumbered; the release refuses one that is.
The release notes list the rules added, changed and retired.

R13.1 (a secret in code) flags a known credential format, a JWT, or a random-looking value bound to a secret, token, password or API key name. It no longer flags placeholders or keys that are public by design, such as Algolia search keys and Stripe pk_ keys.
Every scan skips vendored folders, minified files and bundles. Exclude more with exclude in brevio-core.toml.
No rule reads a TypeScript declaration file (.d.ts, .d.mts, .d.cts): it holds types, no code that runs.
The eval, new Function and dangerouslySetInnerHTML findings have their own rule, R2.8. Earlier builds reported them as R2.5, the SSRF and timeout guard.
R6.3 (a sleep in a test) is preview: it reports and never fails a scan.

brevio-core hook run is the hook a coding client calls after an edit and before it stops. It reads the client’s event, scans the changed lines and hands the findings back in that client’s own format, so the agent fixes them without a human turn. It gives up quietly on an error or past its time budget.
hook install writes the hooks for Claude Code and Cursor, and the agent skill for the clients that load skills.
scan --format agent prints one short line per finding with a fix hint, as the hook and the MCP server do.
learn --staged, learn --worktree and learn --dry-run compile a guard from a fix that is not committed yet, and show what it would match before anything is written.

scan --format markdown prints the findings as an escaped table, with links to the rule pages.
The Action writes the job summary and annotations, keeps one pull request comment up to date and marks it resolved once the findings are gone, and can upload SARIF to code scanning. It runs on pull requests, merge queues and pushes only, and pins every action it uses to a commit.

brevio-core aeo scores a folder of built HTML pages, with its robots.txt and llms.txt, against the AEO rubric, as text, Markdown or JSON, and fails below --min. It reads local files only and refuses a URL with a hint to save the page first.

A command reference generated from the code, a guide to the Action, a guide to writing a rule, the guard and SARIF formats, a FAQ and this changelog.
One page per coding agent: Claude Code, Cursor, Codex, Hermes and OpenClaw.
Each rule page leads with a short answer, has a search-friendly title, links related rules and carries structured data; the docs site has a sitemap.
An agent-safe starter template: the workflow, the AGENTS.md block and a README, ready to become a template repository.

Each release has a MANIFEST.json that maps every file to its source, and a .pyz single-file build with its SBOM and checksum in the release notes.
Releases are published by a workflow with approvals, from a signed tag.
The README is tested: its marked examples run in CI.
The security review page states what the package reads and runs, and tests check each claim.
Governance, support, a rule-acceptance rubric, a rule-test template and issue forms for bugs, false positives and new rules.
A trademark brand-use section and a privacy notice.