clawtfup AGENTS.md

Agent rules: policy evaluation is mandatory

You are working in a repo that uses clawtfup (OPA + .clawtfup/). Treat this as a hard gate, not optional hygiene.

Required behavior

1. Before you treat a coding task as finished, a batch of edits as safe to ship, or a change as “ready for review”, run the evaluator from the repository root (or pass --workspace to that root):

   clawtfup evaluate --pretty
   

   If clawtfup is not on PATH, use the project venv and module form:

   python -m policy_eval evaluate --pretty
   

   The CLI is strict by default: exit code 2 if allow is false or there are error-severity findings (the JSON report is still printed on stdout). You do not pass a flag to “turn strict on”—that is already the behavior.

2. You do not skip this because the diff is “small”, “only docs”, or “obviously fine”. If the command is missing or fails to run, fix the environment (e.g. pip install -e ., OPA on PATH or tools/opa) and re-run—do not silently continue.

3. Only use --no-strict when a human explicitly asked for a non-failing exit while inspecting a denied report (e.g. scripting that parses JSON regardless). Agents must not use --no-strict to avoid fixing policy violations.

4. If evaluation fails (non-zero exit), read findings[] and each item’s feedback.remediation (when present). Implement fixes that satisfy policy, then run the command again until exit 0—or stop and say clearly that policy blocks the change and why.

5. Default semantics: the evaluator indexes the working tree on disk (no git diff HEAD for the default full scan), applies .cfupignore patterns when that file exists (gitignore syntax), then runs policy on the entire indexed tree by default (inputs.scan_mode: full_tree). Pass --diff-file or stdin to layer a unified diff on top of the disk index. Use --diff-only without --diff-file only when you intentionally scope to git diff HEAD (legacy). Agents must not use --diff-only to evade fixes unless the user asked.