SuperAPI AI Directory

using-meisijiya-skills

General & other

View repository

Updated 10-16 00:23

About this skill

IMPORTANT: When dispatching sub-agents via task(), ALWAYS pass the COMPLETE load_skills set from the Category × Skill Matrix main table (references/category-matrix.md) — never [] for matrix-mapped categories.

This is not optional. Skills encode team-validated discipline; bypassing them because "this is simple" is exactly when they matter.

Overview

Meta dispatcher. Hard-injected every session by the OpenCode plugin (~/.config/opencode/plugins/meisijiya-skills.js) so the model actively invokes skills instead of letting them sit in <available_skills>. Coordinates with omo Sisyphus + IntentGate; this skill does NOT do work — it routes. SOT: each skill's description field; cross-skill hints in references/priority-table.md.

When to Use

Use when:

  • Starting any session in a project where meisijiya-skills are installed (you're in one right now)
  • About to take any action on the user's behalf — every turn, before responding

NOT for:

  • Sub-agents (you were dispatched to execute a specific task; ignore this skill — see <SUBAGENT-STOP>)
  • Executors — receive domain-specific skills in the dispatch prompt, NOT this dispatcher
  • Doing the work itself — this skill only routes (PROCESS); the <EXTREMELY-IMPORTANT> discipline applies to the invoked skill

Process

  1. Check .omo/handoff/ for unconsumed documents (cross-session resumption). If any .md file has consumed: false, this takes precedence over trigger matching — user expects RESUME FROM PHASE <to_phase>, not fresh routing. Wait for consumed or consume --reject <reason> before proceeding. See meisijiya-handoff.
  2. Consult <available_skills> (injected by the harness) for the session's skill roster + each skill's description.
  3. Match the incoming request against each skill's description — the description is the source of truth for routing.
  4. Cross-skill hints: references/priority-table.md (trigger patterns); references/process-chains.md (multi-stage sequences); references/controller-executor.md + references/model-selection.md (sub-agent split / model selection).
  5. Announce "Using [skill] to [purpose]" when invoking, and follow the invoked skill's checklist exactly.
  6. When delegating to sub-agents, follow the Sisyphus Dispatch Protocol — always specify the COMPLETE load_skills set from the Category × Skill Matrix main table (Hard Rule; see references/category-matrix.md).

Category × Skill Matrix

→ See references/category-matrix.md (matrix + Security 5-lane fan-out + specialist agents + Common Dispatch Scenarios + Sisyphus Dispatch Protocol).

Plugin layer (where this bootstrap comes from)

This SKILL.md is injected into firstUser.parts each session — see meisijiya-skills.js:113-132 (Editing rules below). The EXTREMELY_IMPORTANT marker guard makes re-injection idempotent after compaction.

  • ~/.config/opencode/plugins/meisijiya-skills.js reads this SKILL.md on session start, strips frontmatter, wraps in <EXTREMELY-IMPORTANT>, and unshifts into firstUser.parts once per session (idempotent against compaction re-fires).
  • ~/.config/opencode/plugins/meisijiya-review-router.js injects per-Edit reminders for ai-code-blindspots + security-and-hardening + verification-before-completion after write / edit / apply_patch (deduped per turn).

Editing rules — Changes to SKILL.md do not auto-reload; OpenCode plugins load once at process start. Restart OpenCode after editing. If <EXTREMELY-IMPORTANT> is missing, the bootstrap plugin is not installed — fix the install before assuming this skill is active.

User Instructions

User instructions (AGENTS.md, direct requests) take precedence over skills, which override default behavior. Only skip skill workflows when your human partner has explicitly told you to.

Verification

Before responding, confirm:

  • You have either invoked a matching skill OR explicitly stated no skill matches
  • You announced "Using [skill] to [purpose]" when invoking
  • The invoked skill's allowed-tools covers the tools you need (otherwise escalate)
  • For sub-agent dispatches: load_skills=[...] is specified (consult references/category-matrix.md)
  • Rationalizations / Red Flags checked (see references/dispatcher-rationalizations.md) — no skipped skill on "this is simple" grounds