twix

Twix — a reviewer pair on the diff

Roster — твікс = Agent agatha + Agent arthur (default); твікс лайт = Agent arthur + Agent dash; two names in the trigger = exactly those two. прекоги = the precogs workflow (workflows/precogs.js in this repo, installed per its README; args in its header). Roster and synthesis come from the script; the rules below the script cannot enforce itself the calling session supplies through its args — the scope with untracked names, the spec/ADR path, the context. Its debate mode is a deliberate second round on the same findings, not a re-run after patches; rule 5 applies to re-runs.

Panel rules (shared by twix, precogs, /implement step 5)

1. Scope = uncommitted working-tree changes: git diff HEAD plus untracked files. git diff does not show untracked files and git status --short collapses a new directory into one line, so list them with git status --short --untracked-files=all and name every file in the prompt with "read these directly" — a reviewer that only diffs sees nothing and reports a clean change. The user may name another scope (PR, commit range, files). Put the exact git command and the untracked names in every prompt so they review the same thing.
2. Launch all reviewers in ONE message — parallel. Each prompt carries:
   • the scope command; what changed and why (they never see the conversation); the spec/ADR path if one exists;
   • where the project's packages live for API-signature checks (the service venv named in the project's CLAUDE.md, not system python);
   • severity buckets BLOCKING / IMPORTANT / NIT;
   • the explicit out: "If a category is empty, say 'BLOCKING: none.' etc.; if everything is clean, say 'No issues found.' A clean review is a valid result.";
   • "An assumption about the environment stated in a comment (timezone of a column or process, encoding, locale, a default) is verified against the running environment or reported as UNVERIFIED — never confirmed by reading the comment."
3. Frozen tree while they run — no edits, no stash, no checkout. Scratch goes to the scratchpad.
4. Merged verdict: dedupe, mark findings several reviewers raised, verify every BLOCKING against the code before accepting it — a surprising one empirically (run it, query it). Report in plain language, four buckets: found / fixing myself / needs your yes / rejected and why. Implementation-level fixes: the session applies them itself; ADR/product-level ones go to the user.
5. Re-run only on the user's word, with prompts byte-identical to the first pass — no "we fixed A, B, C, verify" priming.