legion-steer

Lint catches a violation after the agent has already written it. By then the oversized component exists and the fix is a refactor nobody asked for. legion-steer moves the rules earlier: into the agent's context, before it writes.

npx legion-steer init            # write or refresh the block
npx legion-steer check           # exit 1 on drift, for CI

The block is generated, never hand-written

It comes from the repo's resolved lint config — the rules that are on, each at its level, what a bypass looks like, which rules are locked, and the lint command.

That is the whole point. A rule the config does not enforce cannot appear in the block, so what the agents are told cannot drift from what the linter does. The same argument LEGION-STANDARDS §8.1 makes about documented rules applies to agent instructions, which until now were hand-maintained prose in every repo.

Steering for a React Native repo names no-rn-button and controlled-text-input and omits button-has-type, because that is what the reactNative preset actually enforces. Nobody has to remember to keep the two in step.

Targets

Detected automatically, or chosen with --targets:

| id | file | | --------- | --------------------------------- | | agents | AGENTS.md | | junie | .junie/guidelines.md | | cursor | .cursor/rules/legion.mdc | | copilot | .github/copilot-instructions.md |

With none present it writes AGENTS.md.

It only owns what is between the markers

<!-- legion-steer:start -->

...generated...
<!-- legion-steer:end -->

Content outside them is never touched. A file that already exists gets the block appended; a file that already has the markers gets only that region replaced; a file already current is left alone entirely.

Relationship to LEGION-STANDARDS

The block states what this repo enforces and points at LEGION-STANDARDS.md for why. It deliberately does not reproduce or condense the standards, which §14 rules out — one file, one location, read by every repo underneath.

The hook: catching it before it lands

Lint tells you after the file exists. legion-steer hook runs the same rules on the content Claude Code is about to write, and blocks the write when it would introduce an error.

init registers it in .claude/settings.json, merging with whatever is already there:

| event | matcher | behaviour | | ------------- | ------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | | PreToolUse | Write | lints tool_input.content in a scratch copy beside the target, so path-based rules see the real directory. Exit 2 blocks the write. | | PostToolUse | Edit\|Write | lints the file that was just written and reports. The edit already happened, so this is feedback one edit later rather than one task later. |

Only errors block; warnings do not. Non-TypeScript writes pass straight through, and if oxlint is not installed the hook stays out of the way.

What the agent actually sees:

legion-steer: 1 violation(s) would land in this file
  src/Panel.tsx:2  `useState` inside `Panel` is state in a component
                   (LEGION-STANDARDS section 1). Own it in `usePanel` and
                   return the value and a `handleX` setter.
Move the logic into a use<Component> hook or lib/, then write the file again.

The path is the file it tried to write, not the scratch copy, and the message says what to do rather than only what is wrong.

Skip registration with --no-hooks.

check in CI

legion-steer check || exit 1

Regenerates the block in memory and compares. Non-zero when a repo's steering has fallen behind its own lint config.

Options

--root <dir>       repo root (default: cwd)
--preset <name>    recommended | strict | reactNative
--targets <ids>    comma separated (default: whichever are present)
--lint <command>   the lint command to name in the block