yaj-a

name	yajña
aliases	["yajna","autonomous","loop","agentic-loop","coordinate","ritual"]
description	Autonomous development ritual with role-based coordination. Loops until complete using specialized agents (hotṛ→research, adhvaryu→implement, udgātṛ→test).
execution	direct

Yajña (यज्ञ) - Autonomous Development Ritual

Truly autonomous development loop. Runs continuously until done. NO manual coordination.

Critical Rules

1. NEVER ask user for routine decisions - just do it
2. NEVER pause between iterations - loop automatically
3. STOP ONLY for: Blockers | Completion | User interrupt
4. Blockers require user: Daemon down, permissions, unresolvable errors

Pre-Flight Check (REQUIRED)

Before EVERY iteration, verify daemon is alive:

health = mcp__chitta-mcp__health_check();
if (health.error || health.status !== "OK") {
  // BLOCKER - stop and report
  output("[BLOCKER] Daemon unreachable. Run: pkill -9 chittad && chittad daemon");
  long_task_event({ event_type: "blocker", description: "Daemon down" });
  STOP;  // Do not continue
}

The Loop (Execute Without Pausing)

WHILE tasks_remain AND no_blocker:

  1. PRE-FLIGHT
     health_check → if fail → STOP with blocker

  2. LOAD CONTEXT (silent, no output)
     long_task_snapshot
     read fix_plan.md
     count remaining tasks

  3. HOTṚ - Research (if needed)
     Task(Explore) for complex tasks
     VALIDATE: Did agent find correct files?
     If wrong files → retry with explicit paths

  4. ADHVARYU - Implement
     BEFORE ANY CODE CHANGES:
       a. Call EnterWorktree to create isolated workspace
       b. All Read/Edit/Write/Bash file mutations must happen in the worktree path
       c. On completion or failure, call ExitWorktree
       d. Proposed patches are stored as artifacts, not applied directly to main
     Task(general-purpose) with EXPLICIT file paths
     Prompt MUST include: "Edit file X at path Y"
     VALIDATE: Check file was actually modified
     If not modified → retry or mark blocker

  5. UDGĀTṚ - Test
     Task(general-purpose) for validation
     Syntax checks, tests, verification
     If FAIL → log, continue (not a blocker)

  6. CHECKPOINT (silent)
     long_task_update
     Update fix_plan.md

  7. EVALUATE
     All done? → COMPLETION
     Same error 3x? → STAGNATION (blocker)
     Else → continue loop (NO pause, NO output)

Agent Prompts Must Be Explicit

BAD (vague, leads to wrong files):

"Add pattern detection to the hooks"

GOOD (explicit paths, clear instructions):

"Edit /home/user/.claude/hooks/prompt-hook.sh
Add these lines after line 25:
[exact code]
Verify the file exists first with: ls -la /home/user/.claude/hooks/"

Validation After Each Agent

// After Adhvaryu returns
result = Task({ ... });

// Validate the work
if (result.includes("Error") || result.includes("not found")) {
  // Retry with more explicit instructions
  retry_count++;
  if (retry_count >= 3) {
    // BLOCKER
    long_task_event({ event_type: "blocker", description: result });
    STOP;
  }
  continue;  // Retry this iteration
}

// Check file was actually modified
file_check = Bash(`ls -la ${target_file}`);
if (file_check.error) {
  // File doesn't exist - agent edited wrong path
  retry_count++;
  continue;
}

Output Rules

During loop: Minimal output. Just status lines:

━━━ ITERATION 3 ━━━
[HOTṚ] ✓ Found 2 patterns
[ADHVARYU] ✓ Modified 3 files
[UDGĀTṚ] ✓ Tests passing
[4/7 tasks done]

On blocker: Full details + stop:

━━━ BLOCKER ━━━
[ERROR] Daemon not responding
[ACTION] Run: pkill -9 chittad && chittad daemon
[STATUS] Yajna paused at iteration 3

On completion: Summary only:

━━━ PŪRṆĀHUTI ━━━
Complete in 5 iterations.
Files: 8 | Tests: passing | Learnings: 3

Blocker Conditions

These STOP the loop and require user:

Condition	Action
Daemon health_check fails	Stop, show restart command
File not found after 3 retries	Stop, ask user for correct path
Same error 3 consecutive times	Stop, show stagnation
Build/compile fails critically	Stop, show error
Permission denied	Stop, ask user

NOT Blockers (Continue Automatically)

Condition	Action
Test fails	Log, continue to next task
Agent returns partial result	Use what's there, retry rest
Minor validation warning	Log, continue
Task takes long	Wait, don't timeout

Spawn Agents Correctly

Parallel (independent tasks, ONE tool call):

// Send ALL parallel agents in SINGLE message
Task({ description: "Task 1", prompt: "..." });
Task({ description: "Task 2", prompt: "..." });
Task({ description: "Task 3", prompt: "..." });
// Wait for all, then continue

Sequential (dependent):

result1 = Task({ description: "Task 1", prompt: "..." });
// Validate result1
result2 = Task({ description: "Task 2", prompt: `Using ${result1}...` });
// Validate result2

Quick Reference

/yajna              # Start/continue (runs until done)
/yajna init GOAL    # New ritual with explicit goal
/yajna status       # Check progress without running
/yajna pause        # Stop after current iteration

Example: Correct Autonomous Flow

User: /yajna init "Add dark mode"

[PRE-FLIGHT] Daemon OK
[INIT] yajna-myapp-1234
[PLAN] 4 tasks generated

━━━ ITERATION 1 ━━━
[HOTṚ] ✓ Found ThemeProvider, CSS variables
[ADHVARYU] ✓ src/theme/dark.css, src/context/Theme.tsx
[UDGĀTṚ] ✓ Builds, renders
[2/4 done]

━━━ ITERATION 2 ━━━
[ADHVARYU] ✓ Toggle component, localStorage
[UDGĀTṚ] ✓ Tests passing
[4/4 done]

━━━ PŪRṆĀHUTI ━━━
Complete in 2 iterations.
Files: 4 | Tests: 6 passing

[NO USER INTERACTION NEEDED - FULLY AUTONOMOUS]

Anti-Patterns (NEVER DO)

❌ "Let me check with you before proceeding..." ❌ "Should I continue to the next task?" ❌ "I'll wait for your confirmation..." ❌ Outputting between every agent call ❌ Asking which file to edit ❌ Pausing after each iteration

MCP Tools

Tool	Purpose
health_check	Pre-flight daemon check
long_task_start	Initialize ritual
long_task_active	Check for existing
long_task_snapshot	Load context
long_task_update	Record progress
long_task_event	Log blockers/checkpoints
long_task_complete	Mark done