Baymax OpenCode Config

Your no‑nonsense OpenCode setup. Clean config, strict rules, and guardrails that keep your workflow fast without going off the rails.

Index

• Quick start
• What’s inside
• Agents
• Subagents
• Tools
• Commands
• Skills
• Rules
• Recommended workflow (Baymax)
• Memory (soft)
• MCPs
• MCPs that need keys or auth
• MCPs without keys (enabled)
• Hooks (compat)
• Notion

Quick start

git clone https://github.com/MiguelAguiarDEV/baymax.git
cd baymax

Install OpenCode (manual steps now; no auto‑installer here):

curl -fsSL https://opencode.ai/install -o /tmp/opencode-install.sh
less /tmp/opencode-install.sh
bash /tmp/opencode-install.sh

Apply the config

Option A (recommended): symlink this repo to ~/.config/opencode (non‑destructive, reversible).

mkdir -p ~/.config
ln -sfn "$(pwd)" ~/.config/opencode

Option B: replace your current config (backup first).

mv ~/.config/opencode ~/.config/opencode.bak
cp -R "$(pwd)" ~/.config/opencode

Quick steps

1. git clone the repo.
2. Install OpenCode.
3. Apply the config with A or B.

Edit → Review → Commit → Push

1. Edit files.
2. Run review.
3. Commit changes.
4. Push to main.

What’s inside

• Specialist agents and subagents
• Commands, rules, skills, and contexts
• Tools guardrails and MCP config

Agents

Docs: https://opencode.ai/docs/agents/

• architect
• build-error-resolver
• code-reviewer
• doc-updater
• deploy-work
• e2e-runner
• planner
• refactor-cleaner
• security-reviewer
• tdd-guide

Subagents

Subagents are specialists that auto‑trigger based on the task (planning, TDD, security, E2E, etc.).

Use these as the “who” for the job. Each one has a defined profile and rules in /agents.

Same lineup as the agents in /agents (see list above).

Tools

Docs: https://opencode.ai/docs/tools/

Tools are the operational powers (bash, read, write, web, etc.) the agent uses to interact with your environment. They’re gated by strict rules: safe read/write flows, input validation, and the right tool for the job.

• OpenCode tools reference

Commands

Docs: https://opencode.ai/docs/commands/

• /build-fix
• /code-review
• /deploy-work
• /e2e
• /learn
• /memory-save
• /plan
• /refactor-clean
• /tdd
• /test-coverage
• /update-codemaps
• /update-docs

Skills

Docs: https://opencode.ai/docs/skills/

• backend-patterns
• clickhouse-io
• coding-standards
• continuous-learning
• deploy-work
• frontend-patterns
• project-guidelines-example
• security-review
• strategic-compact
• tdd-workflow

Rules

Docs: https://opencode.ai/docs/rules/

• agents
• coding-style
• git-workflow
• hooks
• hooks-compat
• patterns
• performance
• security
• testing

Recommended workflow (Baymax)

• Always plan: use /plan before any change.
• TDD for real changes: /tdd for new features or actual fixes.
• Always code review: /code-review after each change.
• Security review when needed: @security-reviewer for input/auth/API/sensitive data.
• E2E for critical flows: /e2e when changes touch key flows.
• Anti‑overengineering: keep it simple, ship what meets the requirements.

Memory (soft)

The memory MCP is enabled, but opt‑in.

• Stored at ~/.config/opencode/memory.json (once linked).
• Every memory write runs git add, git commit, and git push automatically.
• Persistence is blocked if secret patterns are detected (tokens, keys, emails).
• To disable, comment out the memory_ hook in config/opencode/hooks/hooks.json.
• Requires jq, a valid git repo, and upstream configured for auto‑push (with auth/permissions).
• If push fails (protected branch or auth), the commit stays local.
• Manual command: /memory-save

MCPs

Docs: https://opencode.ai/docs/mcp-servers/

• filesystem
• playwright
• memory
• sequential-thinking
• magic
• github
• firecrawl
• context7
• vercel
• cloudflare-docs
• cloudflare-observability
• cloudflare-workers-bindings
• cloudflare-workers-builds

Note: Some versions of the filesystem MCP don’t expand ${HOME}. The installer swaps in your real home path.

MCPs that need keys or auth

Confirmed via config or endpoint response.

• github: requires GITHUB_PERSONAL_ACCESS_TOKEN.
• firecrawl: requires FIRECRAWL_API_KEY.
• context7: requires CONTEXT7_API_KEY.
• vercel: remote endpoint returns 401 → requires auth (OAuth or headers).
• cloudflare-* (docs/builds/bindings/observability): remote endpoint rejected fetch (406). Likely OAuth/headers; treat as auth‑required.

Note: 406 usually means missing headers or auth. Treat as credentials‑required.

MCPs without keys (enabled)

• filesystem (local)
• playwright (local)
• memory (local)
• sequential-thinking (local)
• magic (local)

Hooks (compat)

OpenCode doesn’t run Claude Code hooks. Equivalent rules live here:

• hooks-compat

Notion

All Notion‑related config has been removed.