@sffmc/cognition

Parallel reasoning + structured-workflow skills + plugin health diagnostics for OpenCode. Three formerly-separate sub-features (@sffmc/max-mode, @sffmc/compose, @sffmc/health) live here as internal sub-folders; the package entry composes them into one plugin via mergeHooks().

What's inside

• max-mode/ — maxMode() tool. Run a workflow under several candidate generators in parallel; an LLM judge scores each output on configurable criteria and returns the winner. Useful when one generator is flaky and you want a second opinion rather than a retry loop.
• compose/skills/ — 18 markdown skill files that load into OpenCode's prompt: plan.md, tdd.md, verify.md, review.md, debug.md, brainstorm.md, etc. The LLM knows when to load them based on the user's request.
• health/ — sffmc_health tool. Returns a JSON diagnostic of plugin load order, hook collisions, missing-config warnings, and per-package health. Useful for sffmc doctor-style reports in a chat session.

Architecture

• packages/cognition/src/index.ts — composes the three sub-features with mergeHooks()
• packages/cognition/src/max-mode/src/index.ts — maxMode() implementation
• packages/cognition/src/compose/src/index.ts — skill registry
• packages/cognition/src/health/src/index.ts — health tool implementation

Install

// ~/.config/opencode/opencode.json
{
  "plugins": [
    "npm:@sffmc/cognition@^0.15.4"
  ]
}

Skills

See packages/cognition/src/compose/skills/ for the full list. Each .md file is a self-contained prompt fragment the LLM loads on demand:

• ask.md — clarify ambiguous user input before acting
• benchmark.md — measure baseline cost / latency / accuracy before optimizing
• brainstorm.md — generate multiple candidate plans before picking one
• code-review.md — second-pass review against the spec
• debug.md — diagnose a failing test or build
• execute.md — drive a known-good plan to completion, reporting only deviations
• feedback.md — turn user feedback into concrete action items
• merge.md — orchestrate merges when multiple agents touched the same files
• new-skill.md — meta: write a new skill file
• parallel.md — split a large task across parallel agents
• plan.md — produce a numbered implementation plan from a vague goal
• report.md — summarize progress and remaining work
• review.md — adversarial review of an agent's work
• subagent.md — launch a sub-agent with a clear scope and return format
• tdd.md — write a failing test first, then the implementation
• verify.md — confirm an implementation satisfies the original requirement
• worktree.md — manage isolated git worktrees for parallel work

License

MIT