Bulba

Native OpenCode V2 setup: agents, commands, skills, always-on rules. Markdown only — zero plugin code (V1 plugins do not run in V2, so the old plugin was removed entirely, nothing ported).

Layout

Dir	What
agents/	14 agents: bulba (primary router, default) + 13 subagents — planner, implementer, reviewer, verifier, debugger, benchmarker, paper-explainer, optimizer, webdev, security, skillfinder, researcher, critic
commands/	29 commands: dev loop (/develop /plan /go /verify /publish /status /goal), research/docs/quality, tutor /teach, ponytail set /ponytail*
skills/	14 skills (see Provenance)
AGENTS.md	Always-on rules: ponytail (full level) + Bulba core

Install

The repo is the source of truth; the global config gets symlinks:

ln -s "$PWD/agents"    ~/.config/opencode/agents
ln -s "$PWD/commands"  ~/.config/opencode/commands
ln -s "$PWD/skills"    ~/.config/opencode/skills
ln -s "$PWD/AGENTS.md" ~/.config/opencode/AGENTS.md

Plus "default_agent": "bulba" in ~/.config/opencode/opencode.jsonc. Providers stay unset — pick the model in the TUI.

Gotcha: a running opencode service rebuilds its agent registry on config value changes (or restart), not on plain edits to agents/*.md — edited agents show up after the next launch or a config change.

Key flows

• /develop — Manage-Execute-Audit loop: plan → questions → tasks → subagents → commits → adversarial review → verify gate → report. No push without /publish.
• /teach <topic> — tutor mode: loads the teach skill (mission-grounded, evidence-based: retrieval, spacing, interleaving), HTML lessons + reference docs, stateful workspace the agent picks.
• /away — work autonomously until the user returns (plain prompt, no permission hacks — V2 allows all by default).
• /research — deep research on the built-in websearch/webfetch, cited report in .bulba/research/.
• /overhaul, /critique, /simplify, /audit, /security-review, /graph, /test-ui, /design …
• /ponytail [lite|full|ultra|off] — level switch; always-on at full via AGENTS.md.

Philosophy

Trust, not enforcement: zero deny-rules. Read-only roles (planner, reviewer, security) are stated in prompts, not permissions. Outward actions (push, PR, delete, creating CI) ask the user. Nothing auto-nags; state lives in .bulba/ (plan.md, goal.md, memory.md, lessons.md — per-project scratch, gitignored there).

Provenance (skills keep their original IDs)

• grilling, grill-me, teach, to-spec, diagnosing-bugs, writing-for-agents — mattpocock/skills
• unslop, blast-radius — cursor/plugins (pstack)
• ponytail, ponytail-review, ponytail-audit, ponytail-gain, ponytail-debt, ponytail-help — DietrichGebert/ponytail (ruleset also inlined in AGENTS.md)

User-invoked-only skills (grill-me, to-spec, blast-radius, teach) carry metadata.opencode/autoinvoke: false; unslop stays model-visible. Slash entries for teach and ponytail* belong to the command files (slash: false on those skills).