| name | supergraph |
| description | Meta-orchestrator — dispatches to the right skill for the task. Use before any non-trivial work. |
| mcp | codebase-memory-mcp |
supergraph
Graph evidence uses codebase-memory-mcp scoped by CBM_PROJECT. Orientation
loads get_architecture aspects overview, layers, boundaries, clusters,
and hotspots, plus validated hubs, bridges, and test-gaps recipes. If
optional graph evidence is unavailable, label it unavailable and use Serena or
filesystem evidence.
Dispatch to the right skill. 1% rule: if it might apply, invoke it.
Priority
- User CLAUDE.md / AGENTS.md (highest)
- This skill set
- Default system behavior
Dispatch Table
| User intent | Skill to invoke |
|---|
| "start", "begin", unclear scope | /supergraph:analyze |
| "plan", "lên kế hoạch", explicit plan request | /supergraph:plan |
| "implement", "build", "execute" (plan saved) | /supergraph:execute |
| "fix bug", "debug", "why failing" | /supergraph:fix |
| "test", "TDD", "add tests" | /supergraph:tdd |
| "refactor", "clean up", "reorganize" | /supergraph:plan (then execute) |
| integration/e2e tests, after unit green | /supergraph:integration |
| "done?", "verify", before commit/PR | /supergraph:verify |
| "review", "merge", "PR", before merge | /supergraph:review |
| Need project context | /supergraph:scan |
Adaptive Gate
- < 10 lines, 1 file, no hub nodes → use
/supergraph:tdd directly
- 1-3 files, clear requirement →
/supergraph:plan with lightweight tasks
- Multi-file, ambiguous, hub/bridge nodes →
/supergraph:analyze first, then plan
User Instructions
"Add X" or "Fix Y" is the WHAT. Skills determine HOW. Never skip a skill because the task "feels simple."
Red Flags (STOP if you catch yourself):
- "I'll just do it quickly" → pick a skill, follow it
- "I already know what to do" → skills evolve, load current one
- "This is too simple for a skill" → simple becomes complex fast
Integration
All skills depend on /supergraph:scan being loaded first in the session.
Subagent Guard
When dispatching subagents (executor, code-reviewer, plan-writer, plan-reviewer):
- Subagents get self-contained prompts — never session history
- Subagents do NOT trigger skills independently — this is the orchestrator's job
- Subagents report results back; orchestrator decides next step (fix, verify, review)
