aarshsharma2803-stack One command, seven modes. /optimusprime alone = activate + status. /optimusprime <mode> = run that mode.
All data lives in .optimusprime/ — find it by walking up from cwd (10 levels max). If missing anywhere: mkdir -p .optimusprime, continue silently. Never ask the user to create folders.
/optimusprime (no args) — activate
1. Detect session type. Read .optimusprime/session-state.json → prompt_count. Over 5 = existing session; otherwise new.
New session: show banner, ask for goal, write contract.json:
⚡ OPTIMUSPRIME ACTIVE
Caveman Bot on · compression on · guardrails armed
Modes: status | compact | repair | bots | tokens | quality | dashboard
What are you working on?
Existing session: recover context BEFORE anything else. Read contract.json (goal), session-snapshot.md (last state + next action), decisions.md (last 5), attempts.md (failed — never retry these), todos.md, loop-state.json. Show:
⚡ CONTEXT RECOVERED
Goal: <goal> · <n> decisions · <n>k tokens · loop <n>
Avoid: <failed approaches or "nothing">
Next: <next action from snapshot>
2. Both flows: Caveman-style compression active from this response onward — drop articles/filler/pleasantries, fragments OK, all technical substance stays. Persists whole session.
/optimusprime status
One panel: tokens (cost-log.json), decisions count (decisions.md), loop streak (loop-state.json, ⚠️ at 3+), compression avg (compression-log.json), active bots (skills.json). Nothing else.
/optimusprime compact
Compact conversation now. Keep: goal, decisions + reasons, requirements, errors + context, code written this session, failed approaches, task state. Drop: repeated explanations, narration, filler. Output the compact summary, write it to .optimusprime/session-snapshot.md.
/optimusprime repair
Read loop-state.json + attempts.md. Show: current error, attempt count, what failed. Propose ONE different approach — different strategy, not a retry. Then reset loop-state.json to {"consecutive_failures":[]}.
/optimusprime bots
Table of 5 bots from skills.json × registry.json: Caveman (auto, tokens>40k), Superpowers (full-budget builds), UI/UX Pro Max (frontend files), Ponytail (minimal budget), Gstack (deploy/ship goals). Offer mode change → write skills.json.
/optimusprime tokens
Session tokens + cost, last 5 sessions trend, compression savings total. Recommendation: <40k healthy · 40-80k caveman auto-on · 80k+ run compact · 150k+ new session with snapshot.
/optimusprime quality
Scan files from session-snapshot.md CHANGED list (or ask which). Check: functions >30 lines, duplicate logic, hardcoded secrets, SQL concatenation, eval/exec on user input. Report critical/important/minor. Clean = "✅ passed".
/optimusprime dashboard
status + last 5 decisions + open TODOs + next action, one screen.
Auto Bot rules (always on once activated)
- tokens > 40k → compress responses harder (caveman full)
- tokens > 80k → maximum compression + suggest
compact - loop streak ≥ 3 → suggest
repair - frontend files touched → apply UI/UX design rigor
- minimal budget → smallest correct implementation, no abstractions
Off only: "stop optimusprime" / "deactivate op".