First run: the interview¶

The first time you run /gru in a project (and the first time with this version of Gru), it interviews you before it does anything else. Knowing the project and you is what makes the roles work the way you expect. It pre-fills what the repo already answers and asks only the rest.

Does Gru leave files in this repo? team (default): a small gru.yml policy and the project notes in AGENTS.md/CLAUDE.md are versioned so teammates share them. none: nothing is versioned, for projects where people do not want Gru around.
Models. Gru works with three strengths: strong (thinks, plans, reviews), mid and weak. Defaults: Opus, Sonnet, Haiku. Stay or change.
Harness for the minions. "The default is Claude Code, but I suggest tether (recommended)." You can also choose opencode or another. If tether breaks on a card, Gru falls back to Claude Code for that card and offers to report the problem to the tether repo.
Your system. Gru detects the OS. Linux and WSL2 use bwrap; macOS and native Windows need Docker. You also say whether QA should run in Docker, what installs the dependencies, and which services tests need.
Limits. Any local model (a Qwen on your machine) or rate-limited provider (NVIDIA free tier)? Those run sequentially or throttled.
What you expect from QA and review. What you want to see in each report, and checks specific to this project (for example "the linked Asana task exists and matches the definition of done"). These stay with the project, are never generalized, and are written next to the project facts. Tool-backed checks are gathered by Gru with its own session's tools.
The project. Stack, check commands, e2e, branch and PR style, who merges, delivery mode, what "done" means in your team, risky areas, hard stops, design docs. Written to AGENTS.md/CLAUDE.md.
You. How you want to be asked, how much Gru may decide alone, verbosity, notifications. Written once per machine to ~/.config/gru/profile.yml.

Then Gru offers a selftest. Redo it any time with /gru setup.

Files¶

Two files in the store, neither with secrets.

<store>/local.yml (git-ignored, yours): versioning, slots, harness, sandbox, limits, sizes, prices, verified harnesses and calibration.

version: 2
versioning: team          # team | none
sandbox: {default: bwrap, qa: docker}
slots:
  strong: {model: opus,   harness: claude}
  mid:    {model: sonnet, harness: claude}
  weak:   {model: haiku,  harness: claude}
minion_harness: claude
size_slots: {XS: weak, S: weak, M: mid, L: strong, XL: strong}
limits:
  - {match: "tether:ollama/*", max_parallel: 1, kind: local}
  - {match: "tether:nvidia/*", max_parallel: 2, rpm: 40}

<store>/gru.yml (versioned, team policy):

version: 2
tether_repo: igormcsouza/tether
budget: {warn_tokens: 300000, stop_tokens: 400000}
timeout_min: 10
max_parallel: 3
sensitive: [{path: "infra/**", min_slot: mid}]
permissions: {file_tether_issues: ask, post_reviews: ask, file_gru_issues: ask}

~/.config/gru/profile.yml: your working style (autonomy, questions, verbosity, notify) and machine specifics such as claude_config_dir.

An old .claude/gru.json is migrated: the tier ladder is dropped, the rest is kept.