Command: /cm-start [your objective]

TL;DR

• Use to kick off a CM session — entry point
• Detects: stack (Phase 2), suggests skills, reads continuity + learnings
• Autonomy: selects project level from evidence; clear objectives do not need level confirmation
• Next: cm-brainstorm-idea or cm-planning

Role: Workflow Orchestrator — You assess complexity, select the right workflow depth, and drive execution from objective to production code.

Follow _shared/autonomy-policy.md. Its decision table controls confirmations across this workflow.

When this workflow is called, the AI Assistant should execute the following action sequence in the spirit of the CodyMaster Kit:

0. Load Working Memory: Per _shared/helpers.md#Load-Working-Memory — use Smart Spine order:

   1. Check .cm/context-bus.json → any active pipeline? any prior skill output to reuse?
   2. Load L0 indexes: learnings-index.md (~100 tok) + skeleton-index.md (~500 tok)
   3. Scope-filter learnings via cm_query — only load what matches current objective
   4. Read CONTINUITY.md → set Active Goal to the new objective
   5. Run token budget check: cm continuity budget → confirm no category is over soft limit

   ⚡ Total context load: ~700 tokens. Full load used to be ~3,200. Only escalate to L2 (full files) if L0 index explicitly flags a match. 0.5. Skill Coverage Check (Adaptive Discovery):

   • Scan the objective for technologies, frameworks, or patterns mentioned
   • Cross-reference with cm-skill-index Layer 1 triggers
   • If gap detected → trigger Discovery Loop from cm-skill-index: npx skills find "{keyword}" → review → ask user → install if approved
   • Log any discovered skills to .cm-skills-log.json

0.6. Stack & Tier Detection (Phase 2): - cm stack detect --write → writes .cm/project-skills.md (frameworks + suggested skills) - cm tier classify --write → writes .cm/project-tier.md (LITE/STANDARD/PROFESSIONAL/ENTERPRISE) - The tier sets the default Vibecoding mode and adaptive depth: - LITE/STANDARD → render skill TL;DR only - PROFESSIONAL/ENTERPRISE → render full protocol - Inject the suggested-skills list into the skill chain shortlist - These reports are token-light (~300 tok combined) and skipped if files exist and are <24h old

0.7. Code Intelligence Setup (cm-codeintell): - ALWAYS: Run skeleton indexer → bash scripts/index-codebase.sh → .cm/skeleton.md - Read .cm/skeleton.md (~5K tokens) → instant codebase understanding - Count source files → determine intelligence level (MINIMAL/LITE/STANDARD/FULL) - IF level >= LITE: generate architecture diagram → .cm/architecture.mmd - IF level >= STANDARD: check CodeGraph → codegraph status → index if needed - IF level >= STANDARD: also check qmd (cm-deep-search) for existing semantic vector databases and initialize/update if needed. - Log intelligence level to CONTINUITY.md

1. Understand Requirements (Planning & JTBD):

   • Read the objective provided in the /cm-start command.
   • Analyze requirements. Ask once only when ambiguity would materially change scope; include a recommendation and default.
   • Consider multi-language support (i18n) from the start if the project requires it.

2. Detect Project Level: Per _shared/helpers.md#Project-Level-Detection

   • Select the L0/L1/L2/L3 project level from objective and repository evidence
   • State the detected level and recommended skill chain, then continue without confirmation when the objective is clear
   • Allow the user to override the level at any time; an override applies from the next safe boundary
   • Do not treat level selection as the plan-to-execution approval boundary

3. Execute Based on Level:

   L0 (Micro): Code + Test only

   • Skip planning. A clear, reversible micro task may proceed with zero approval.
   • Apply cm-tdd directly → cm-quality-gate

   L1 (Small): Planning lite → Code → Deploy

   • Apply cm-planning (lightweight implementation plan)
   • For meaningful code changes, request one plan approval that grants scoped execution authorization
   • Apply cm-tdd + cm-execution → cm-quality-gate

   L2 (Medium): Full analysis flow

   • Init OpenSpec (create openspec/changes/[initiative-name]/ folder and artifacts manually)
   • Apply cm-brainstorm-idea if problem is ambiguous
   • Apply cm-planning (full implementation plan with OpenSpec tasks.md)
   • Request one plan approval that grants scoped execution authorization
   • Create cm-tasks.json from tasks.md → launch RARV autonomous execution
   • Apply cm-quality-gate → cm-safe-deploy

   L3 (Large): Full + PRD + Architecture + Sprint

   • Init OpenSpec (create openspec/changes/[initiative-name]/ folder and artifacts manually)
   • Apply cm-brainstorm-idea (mandatory)
   • Apply cm-planning with FR/NFR requirement tracing
   • Request one plan approval that grants scoped execution authorization
   • Sprint planning → openspec/changes/[objective]/tasks.md sync with cm-tasks.json
   • Apply cm-execution (Mode E: TRIZ-Parallel for speed)
   • Apply cm-quality-gate → cm-safe-deploy

4. Track Progress:

   • Create openspec/changes/[objective]/tasks.md (for standardized spec tracking)
   • Create or update cm-tasks.json (for autonomous agent execution)
   • Suggest /cm-dashboard for visual tracking
   • Suggest /cm-status for quick terminal summary

5. Complete: Per _shared/helpers.md#Update-Continuity

   • Record any new learnings or decisions made during this workflow
   • If inside a skill chain: cm continuity bus → verify context bus reflects completed step
   • Refresh L0 indexes: cm continuity index (auto-runs on addLearning, manual refresh here)

Note for AI: If this is a brand new project, suggest running cm-project-bootstrap first. If the working environment has a risk of accidentally switching accounts/projects, remind about cm-identity-guard (Per _shared/helpers.md#Identity-Check).