AgentLint copilot-instructions.md

Copilot instructions for AgentLint

AgentLint is a shell-based linter for AI-friendly repos. It scans for CLAUDE.md/AGENTS.md quality, CI hygiene, and agent-ready patterns, then emits SARIF + scored reports. Implementation is bash + JavaScript: src/scanner.sh runs 51 deterministic checks across 6 core dimensions (findability, instructions, workability, safety, continuity, harness), and JS modules add 7 opt-in extended checks across 2 dimensions (Deep + Session) for 58 total. The CLI wrapper is shell + JS, published to npm.

Any Copilot agent (Coding Agent, Code Review, ci-doctor) working on this repo must follow the rules below.

Hard rules — never violate

1. No test assertion mutation. If a test is failing, fix the code or file a bug — never change the expected value in an assertion to make it pass.
2. No lint / type / test suppressions. No @ts-ignore, # noqa, // eslint-disable, it.skip, @pytest.mark.skip, or equivalent. The tool raised it for a reason.
3. No weakening of CI, workflows, or build config. Do not edit .github/workflows/**, Makefile, tsconfig.json, package.json scripts, or test config to make things pass. Fix the code instead.
4. No deleting or skipping tests to resolve a failure.
5. No unrelated refactors. Only change lines directly required by the task. A bug fix does not need surrounding cleanup.
6. No features, abstractions, or error handling the task did not ask for. Three similar lines beats a premature abstraction. Don't validate scenarios that cannot happen.
7. No absolute hardcoded paths (/Users/..., /home/...). Use ${CLAUDE_PLUGIN_ROOT}, env vars, or repo-relative paths.
8. No emoji in code, comments, commit messages, or markdown unless explicitly requested.
9. No new top-level files (README alternates, design docs, notes) unless the task explicitly asks.
10. No commits to main. All changes go through a PR.

Scope discipline for this repo

• Adding a check requires evidence. Cite an Anthropic source, academic paper, or documented real-world failure in standards/evidence.json. Opinions are not acceptable.
• Scanner output is a contract. Changing the JSONL schema in src/scanner.sh requires updating the scorer, plan-generator, and reporter in the same PR.
• Scoring changes require rationale in standards/weights.json. Do not adjust weights silently.
• Thresholds in shell code must match standards/evidence.json. Do not hardcode a new number without updating the evidence file.
• Skills in commands/*.md must reference bundled assets via ${CLAUDE_PLUGIN_ROOT}. Persistent data goes in ${CLAUDE_PLUGIN_DATA} or ~/.al/, never the plugin root.