Skip to content

OPENCODE.md

Latest commit

 

History

History
51 lines (42 loc) · 2.37 KB

OPENCODE.md

File metadata and controls

51 lines (42 loc) · 2.37 KB
Raw

Supergraph for OpenCode

Supergraph skills are installed as OpenCode skills. Use /skills and pick the skill name (scan, analyze, plan, tdd, execute, fix, verify, review, etc.). Do not use /supergraph:* slash commands in OpenCode.

Mandatory workflow

  1. Start every session with the scan skill.
  2. Run analyze to assess ambiguity/risk and select approach — required before plan (skip only for Micro: <20 lines, ≤2 files).
  3. For architectural, data model, or API contract changes, draft sdd before planning.
  4. Create plan with plan (graph-informed, plan-reviewer approval required).
  5. Implement through tdd or execute a saved plan with execute.
  6. After coding, use fix.
  7. Before claiming done, use verify.
  8. Before merge, use review.

Skill routing

Need OpenCode skill
Start session / load graph scan
Ambiguous scope / risk analysis analyze
Software design & contracts sdd
Create implementation plan plan
Implement one task with RED → GREEN → REFACTOR tdd
Execute saved plan execute
Auto-fix tests/lint/graph issues fix
Integration/e2e checks integration
Evidence gate before done verify
Final independent review review
Unknown bug cause diagnose
Need module map zoom-out
Architecture report architecture
Requirements → PRD prd
Issue triage triage
Throwaway validation prototype
Session compaction handoff
Token compression style caveman — always-on, ~75% reduction

Hard rules

  • Never code without a plan unless the change is trivial (<20 lines, ≤2 files, no hub/bridge).
  • Never implement without a verified failing test first.
  • Never modify hub/bridge nodes without user approval.
  • Never claim done without fresh verification evidence.
  • Use graph MCP tools before assuming file relationships.
  • Use Serena MCP tools when available for diagnostics and symbol impact.
  • Caveman always-on: strip articles/filler, keep code/numbers/paths exact, bullets over prose, auto-suspend for safety warnings only.

OpenCode limitations

OpenCode skills are invoked through /skills, not /supergraph:* slash commands. Bash hooks from Claude Code are not active on OpenCode; trigger workflow skills manually. Caveman compression is still always-on via this instruction file — no hook needed.