supergraph
Meta-orchestrator — dispatches to the right skill for the task. Use before any non-trivial work.
Files27
View on GitHub- .claude-plugin/2 files
- .codex-plugin/2 files
- .dsh-plugin/3 files
- .githooks/1 files
- .github/
- .opencode-plugin/
- agents/
- bin/
- docs/
- extensions/
- hooks/
- references/
- scripts/
- skills/
- src/
- team/
- tests/
- themes/
- .gitignore.supergraph41 B
- .mcp.json646 B
- AGENTS.md14 KB
- CHANGELOG.md23 KB
- CLAUDE.md14 KB
- client.js10 KB
- cordis.patch.yml11 KB
- hooks.json1.2 KB
- index.js6.2 KB
- install.sh10 KB
- LICENSE1.0 KB
- mcp_config.json641 B
- OPENCODE.md2.5 KB
- package.json1.7 KB
- plugin.json1.2 KB
- SECURITY.md1.4 KB
- settings.json3.6 KB
- SKILL.md2.8 KB
- SOUL.md3.4 KB
SKILL.mdplugins/supergraph/SKILL.md
GitHubsupergraph
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 |
| "video demo", "demo video", "walkthrough", "quay video demo" | /supergraph:video-demo |
| "done?", "verify", before commit/PR | /supergraph:verify |
| "review", "merge", "PR", before merge | /supergraph:review |
| Need project context | /supergraph:scan |
Adaptive Gate
- < 20 lines, ≤2 files, no hub/bridge, complexity <10 → use
/supergraph:tdddirectly - ≤5 files, clear requirement, no cross-boundary →
/supergraph:planwith lightweight tasks - >5 files, ambiguous, hub/bridge, cross-boundary →
/supergraph:analyzefirst, 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, qa-tester, security-auditor, diagnostician):
- 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)