Skip to main content

3.2 Pi-Orchestrator

Pi-Orchestrator

The Pi-Orchestrator (~/system/kernel/pi-orchestrator.js) is the central brain of the ALAI system. It's a daemon that runs 24/7.

What It Does

Every 30 seconds, the orchestrator:

  1. Polls MC for the next eligible task (mc.js next-task)
  2. Skips tasks matching safety patterns (CEO decisions, regulated services)
  3. Checks retry cap — max 3 attempts per task (prevents infinite loops)
  4. Classifies the task (complexity 1-5, domain, type) using Ollama
  5. Selects model based on classification tier
  6. Checks capacity — worker pool limits per host
  7. Claims the task (assigns to pi-orchestrator, sets status to started)
  8. Builds prompt — injects company soul, blueprints, task context
  9. Executes via the selected pipeline (ollama-tool → simple fallback → llama-server)
  10. Quality gate — checks response length, detects placeholders, escalates if needed
  11. Creates proof-of-work — writes GOTCHA file + verify dir
  12. Completes — marks task done in MC, posts to Slack, feeds HiveMind
  13. Advances pipeline — triggers next pipeline stage if applicable

Configuration

~/system/config/pi-orchestrator-config.json:

  • Concurrency: FORGE: 2, ANVIL: 3, Claude: 2 (total max: 6 parallel workers)
  • Timeouts: Tier 1: 60s, Tier 2: 120s, Tier 3: 300s, Tier 4: 600s
  • Quality gate: Min response lengths, placeholder detection, max 1 escalation
  • Safety patterns: CEO, DECISION, BankID, Vipps, Finanstilsynet, TENDER

Execution Pipelines

Pipeline How It Works
ollama-tool Ollama with tool-use (read/write files, run commands) — preferred
ollama-simple Ollama text-only (fallback when tool agent fails)
llama-tool Kimi K2.5 via llama-server with tool-use wrapper
llama-server Kimi K2.5 raw text completion (fallback)
claude-cli Claude Code CLI for client-facing/critical work
human-queue Queued for human — Alem reviews manually
Back to top