bstack is bercail's opinionated opencode. opencode signs in with a Claude plan (Claude Agent), a ChatGPT plan, and an X subscription, so one harness runs all three, on subscriptions you already pay for. bstack sets opencode up the way bercail works: the job is a beads issue, agents stage and leave changes uncommitted for tuicr, and each kind of work goes to the model that is best at it.
It borrows from pstack (rigor, playbooks, roles by model strength, how/why/unslop), cursor-team-kit (babysit-pr and CI loops, which became autopilot), tstack (a council of providers), and gstack, on top of Shep's handoff workflow.
New installs run bstack. Claude Code, Codex CLI, and cursor-agent stay equal harnesses beside it. Bercail does not prefer one for a handoff, pick one, or fall back from one to another.
bercail bstack setup # installs opencode, connects your plans; models are picked for you
bercail bstack status # what is connected, which model each role uses
dev # the desk starts opencode as the bstack agentNew configs set harness = "pi" under [bstack]. Set it in an existing config.toml and run bercail bstack setup to install Pi (>=1.2.0) and npm:pi-claude-bridge@0.11.0. Claude uses your Claude Code login (claude auth login), not Pi's /login anthropic or per-token Extra Usage. GPT and Grok use /login openai and /login xai inside Pi. Existing configs without harness keep using opencode; harness = "opencode" selects it explicitly. With bstack on Pi, a desk runs Pi as the lead: bercail update writes its command line to pi/lead.json, and the desk shows its status and comes back after a reboot like an opencode desk.
bercail renders the profile to ~/.local/share/bercail/pi/. Installing the bridge adds its pinned package to Pi's settings, leaving unrelated settings and auth intact. Every role explicitly loads the bridge and the managed extensions/fallback.ts with -e, and unsets ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL, and ANTHROPIC_AUTH_TOKEN so fallback to Claude uses your plan. Role boundaries and git rules remain instructions, not enforced tool restrictions. The builder, judge, council, and autopilot run as interactive hc sibling agents from templates in ~/.config/herdr-core/templates/. Pi handoffs use desks or interactive siblings, not one-shot runs.
After Pi exhausts its own rate-limit retries, the extension tries the next model in [bstack] fallback (default ["claude", "gpt", "grok"]). It scans from the start, skipping the current and already attempted models, and leaves the initial role picks unchanged. Set fallback = ["gpt", "grok"] to limit candidates, or [] to disable them. Families must be unique. The profile stores resolved family=model lines in pi/models and one fallback family per line in pi/fallback. bercail bstack status reports the order, bridge availability, and Claude Code subscription auth (authMethod claude.ai or oauth_token, not an API key). bercail doctor checks the installed bridge and flags missing or stale profile data.
The bstack guide is the full story, with prompts to copy. Each fact lives on one page:
| Topic | Page |
|---|---|
setup, sign-ins, how models are picked, status, /bstack-setup, turning it on |
Set up bstack |
| The playbooks, the roles (lead, build, judge, council, autopilot), what a prompt holds | Work with the bstack agent |
/how, /why, taking over a branch |
Understand the code |
The build playbooks, /unslop |
Build and clean the change |
/council, /ready |
Review it |
/autopilot and what bstack-autopilot may do |
Ship with autopilot |
/handoff, stages bstack and autopilot |
Run jobs in parallel |
Pinning models, effort, role fallback, the files bstack writes, bercail bstack off |
Make it yours |
| Prompts and pitfalls | Recipes and pitfalls |
The skills (bstack, how, why, unslop, council, autopilot) are listed in commands, the [bstack] defaults in config, and the headless stages in handoff.