clawvard-practice

# Clawvard Practice — Governor Pipeline **Mandatory flow. No exceptions. No manual agent dispatch.** ## Step 1: Start session — one per dimension ```bash curl -sL -X POST "https://clawvard.school/api/practice/start" \ -H "Content-Type: application/json" \ -d '{"agentName":"Orchestrator","dimensions":["<dim>"]}' ``` Each session only supports 1 batch (2 questions). For 8 dimensions, start 8 separate sessions. ## Step 2: Dispatch through Governor ```bash cd D:/Users/Administrator/Documents/GitHub/orchestrator python scripts/dispatch.py --raw --approve --wait --timeout 300 \ --dept engineering --action "Clawvard practice: <dim>, target <score>" \ "Clawvard practice: answer these questions and submit via API. Session: practiceId=<id>, hash=<hash>, taskOrder=<order>. Questions: <paste questions>. Target: <score>. Submit endpoint: POST https://clawvard.school/api/practice/answer" ``` This goes through: classify → IntentGateway → Scrutinizer → Dispatcher → Executor (Agent SDK). **Fallback**: If Governor pipeline fails (e.g., nested Claude Code session blocks Agent SDK), use the Agent tool with one agent per dimension, running in parallel. ## Step 3: Review result Read the task output. If score < target, dispatch again with feedback from previous round. ## API Format (hard-won knowledge) Submit payload — ALL fields required: ```json { "practiceId": "<from /start>", "hash": "<ORIGINAL hash from /start — NOT the hash returned after submission>", "agentName": "Orchestrator", "taskOrder": ["<from /start — the full array>"], "answers": [ {"questionId": "<id>", "answer": "<answer>"}, {"questionId": "<id>", "answer": "<answer>"} ] } ``` Field name is `questionId`, NOT `id` or `taskId`. ## Answer Rules ### Open-ended answers: breadth-first, max 2000 chars 1. **Skeleton first**: Write one sentence per scoring point, covering ALL requirements 2. **Then fill**: Add key details to the most important points 3. **Never depth-first**: Do NOT write 200 words on point 1, then 200 on point 2… you WILL get truncated and lose points on everything after the cutoff 4. **Hard limit**: 2000 characters. The API silently truncates longer answers. Everything after the cut is invisible to the grader. ### Multiple choice Letter + 1-2 sentence explanation. These are easy points — don't overthink. ## Rate Limits - 20 sessions per day per agentName - 8 dimensions × 1 session each = 8 sessions minimum - Budget 12 sessions for retries/debugging - Do NOT waste sessions on format testing — the format is documented above ## Common Rationalizations | Thought | Reality | Correct Behavior | |---------|---------|-----------------| | "I'll just answer directly, faster than dispatch.py" | The whole point is testing the Governor pipeline, not your answering ability. Manual answers bypass the system under test. | Dispatch through Governor. Every time. | | "The API format is obvious, I don't need to check" | You've gotten the format wrong before (questionId vs id, hash reuse). "Obvious" is how silent failures happen. | Copy the exact format from the API Format section above. | | "I'll do all 8 dimensions in one session" | API only supports 1 batch (2 questions) per session. Trying to batch = wasted sessions. | 1 session per dimension. 8 dimensions = 8 sessions. | | "Let me write a longer, more thorough answer" | 2000 char hard limit. API silently truncates. Everything past 2000 is invisible to the grader. | Breadth-first: skeleton covering ALL scoring points first, then fill. | | "The score was low, let me retry with the same approach" | Same approach = same result. If Governor pipeline failed, diagnose WHY before burning another session. | Read the error/score feedback. Adjust the dispatch prompt with specific feedback. | | "I'll skip the easy multiple choice, focus on open-ended" | MC are free points. Skipping them tanks the overall score for zero benefit. | Answer MC first (letter + 1-2 sentences). Then tackle open-ended. | ## Forbidden - Manually writing agent prompts with question text - Spawning Agent tool with hand-crafted briefings - Answering questions yourself - Writing JSON payloads yourself