Skip to content

dispatch

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

dispatch

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
agents
agents
 
 
checklists
checklists
 
 
data
data
 
 
lib
lib
 
 
scripts
scripts
 
 
tasks
tasks
 
 
templates
templates
 
 
workflows
workflows
 
 
README.md
README.md
 
 
config.yaml
config.yaml
 
 
requirements.txt
requirements.txt
 
 

README.md

⚡ Dispatch Squad — Parallel Execution Engine

Version: 1.0.0 | Minds: Gene Kim, Reinertsen, Goldratt, Deming, Pedro Valério | Status: Active

Overview

The Dispatch Squad is the execution backbone of AIOS. It receives stories, PRDs, or task lists, decomposes them into atomic sub-tasks, optimizes into DAG-based waves, routes to the correct agents/models, and executes in parallel via subagents — all while keeping the main Opus context clean.

Core philosophy: "Never contaminate the main context. Always subagents. Always."

Cost savings: ~43-58x cheaper than executing in the main Opus context.

Architecture

                    ┌──────────────────────────────────┐
                    │         USER INPUT               │
                    │  Story / PRD / Free Text / Batch  │
                    └───────────────┬──────────────────┘
                                    │
                    ┌───────────────▼──────────────────┐
                    │      DISPATCH-CHIEF              │
                    │      (Gene Kim — Three Ways)      │
                    │  Orchestrates full pipeline        │
                    └──┬──────────┬──────────┬─────────┘
                       │          │          │
           ┌───────────▼──┐  ┌───▼────┐  ┌──▼──────────┐
           │ QUALITY-GATE  │  │ WAVE-  │  │ TASK-       │
           │ (Deming + PV) │  │PLANNER │  │ ROUTER      │
           │ Pre/post      │  │(Reiner-│  │ (Registry   │
           │ validation    │  │ tsen + │  │  based)     │
           │               │  │Goldratt│  │             │
           └───────────────┘  └────────┘  └─────────────┘
                                    │
                    ┌───────────────▼──────────────────┐
                    │         WAVE EXECUTION            │
                    │                                    │
                    │  Wave 1: [T1, T2, T3, T4] ← ALL  │
                    │          parallel via subagents    │
                    │  Wave 2: [T5, T6, T7] ← depends  │
                    │          on Wave 1 outputs         │
                    │  Wave N: [...]                     │
                    └──────────────────────────────────┘

Elite Minds

Mind Framework Role
Gene Kim Three Ways of DevOps (Flow, Feedback, Learning) dispatch-chief — pipeline orchestration
Donald Reinertsen Product Development Flow (175 principles, queue theory) wave-planner — batch optimization, WIP limits
Eliyahu Goldratt Theory of Constraints + Critical Chain wave-planner — bottleneck ID, Drum-Buffer-Rope
W. Edwards Deming PDSA + Statistical Process Control quality-gate — zero-ambiguity, predict-then-verify
Pedro Valério Veto Conditions + Self-Healing quality-gate — block impossible paths

Agents

Agent Tier Mind Purpose
dispatch-chief Orchestrator Gene Kim Receives work, validates sufficiency, orchestrates pipeline
quality-gate Tier 0 Deming + PV Pre/post-execution validation with veto conditions
wave-planner Tier 1 Reinertsen + Goldratt DAG optimization, wave sizing, critical chain
task-router Tier 2 Registry-based Agent/model/enrichment selection per task

Tasks

Task Purpose
plan-execution.md Story/task → execution plan
decompose-to-waves.md Atomic decomposition + DAG optimization
route-tasks.md Agent/model/enrichment selection
execute-wave.md Execute single wave in parallel
verify-wave.md Post-execution verification
report-execution.md Generate execution + cost reports
convert-input.md Normalize any input → dispatch format

Workflows

Workflow Purpose Phases
wf-dispatch-main.yaml Story → Plan → Gate → Execute → Report 7
wf-dispatch-free.yaml Free text → Dispatch (backward compat) 8
wf-dispatch-batch.yaml Multiple stories in sequence 4

Checklists

Checklist Purpose
pre-dispatch-gate.md Sufficiency veto conditions (V0.*)
pre-execution-gate.md Quality veto conditions (V1.*)
post-execution-gate.md Output veto conditions (V2.*)
haiku-prompt-checklist.md ALWAYS/NEVER rules for Haiku prompts
dispatch-health-score.md 12-point run health score
batch-discovery-gate.md Batch Phase 0: story discovery validation (VB0.*)
batch-prioritization-gate.md Batch Phase 1: prioritization + cost confirmation (VB1.*)
batch-dispatch-gate.md Batch Phase 2: circuit breaker + cost control (VB2.*)

Templates

Template Purpose
execution-plan-tmpl.yaml Plan output structure
wave-manifest-tmpl.yaml Wave definition
task-prompt-tmpl.md Enriched prompt for subagent
execution-report-tmpl.md Post-execution report
cost-report-tmpl.md Cost breakdown
dispatch-state-tmpl.json Initial state structure

Scripts (CODE > LLM)

Script Purpose
wave-optimizer.py DAG topological sort (Kahn's algorithm)
build-command-registry.py Scan all squads → command-registry.yaml
cost-tracker.py Track actual costs per run
validate-dispatch-gate.sh Veto condition checker
validate-wave-results.py Post-execution verification
enrich-task.py KB injection (template fill)
dispatch-health-score.py Calculate 12-point health score

Optimization Scripts (v1.1 — CODE > LLM)

Script Purpose Replaces
route-tasks.py Full routing pipeline orchestrator 60% of task-router LLM
select-model.py Q1-Q4 model decision tree LLM model selection
assign-enrichment.py MINIMAL/STANDARD/FULL assignment LLM enrichment logic
assign-timeout.py Timeout + max_turns assignment LLM timeout logic
batch-size-optimizer.py Reinertsen U-curve batch sizing LLM batch decisions
critical-chain-analyzer.py Goldratt critical chain + slack LLM critical path
estimate-batch-cost.py Per-story + total cost estimation LLM cost calculation
validate-haiku-prompt.py 14-rule ALWAYS/NEVER validation Manual checklist

Hooks Scripts

Script Hook Event Purpose
log-task-start.py SubagentStart Track wave progress
check-wave-complete.py SubagentStop Trigger next wave
save-dispatch-state.py PreCompact Auto-save before compaction
reload-dispatch-context.py SessionStart(compact) Re-inject context
log-task-failure.py PostToolUseFailure Log + retry logic

Data Files

File Purpose
domain-registry.yaml Domain → Agent → KB mapping (extensible)
model-selection-rules.yaml When Haiku vs Sonnet vs Opus
veto-conditions.yaml All blocking conditions (V0., V1., V2.*)
enrichment-rules.yaml MINIMAL/STANDARD/FULL definitions
timeout-rules.yaml Per-task, per-wave, per-run timeouts
haiku-patterns.yaml 6 validated prompt patterns
dispatch-heuristics.yaml Gate classification, self-healing, health score
dispatch-kb.md Consolidated knowledge base

Quick Start

# Dispatch from story (PREFERRED)
*new plan/stories/story-feature.md

# Dry run — see plan without executing
*analyze plan/stories/story-feature.md

# Quick cost estimate
*estimate plan/stories/story-feature.md

# Resume interrupted dispatch
*resume

# Check dispatch health
*health

Pipeline Phases

Phase 0: SUFFICIENCY GATE ──────── Is input well-defined?
Phase 1: DECOMPOSITION ─────────── Story → atomic tasks
Phase 2: ROUTING ────────────────── Task → agent + model
Phase 3: WAVE OPTIMIZATION ──────── DAG → parallel waves
Phase 4: ENRICHMENT ─────────────── Inject KB context
Phase 4.5: PRE-EXECUTION GATE ──── Veto conditions (V1.*)
Phase 5: EXECUTION ──────────────── Parallel subagents
Phase 5.5: POST-EXECUTION GATE ─── Verify outputs (V2.*)
Phase 6: REPORTING ──────────────── Report + cost + state

Pack Structure

squads/dispatch/
├── config.yaml
├── README.md
├── agents/
│   ├── dispatch-chief.md          # Orchestrator (Gene Kim)
│   ├── quality-gate.md            # Tier 0 (Deming + PV)
│   ├── wave-planner.md            # Tier 1 (Reinertsen + Goldratt)
│   └── task-router.md             # Tier 2 (Registry-based)
├── tasks/
│   ├── plan-execution.md
│   ├── decompose-to-waves.md
│   ├── route-tasks.md
│   ├── execute-wave.md
│   ├── verify-wave.md
│   ├── report-execution.md
│   └── convert-input.md
├── workflows/
│   ├── wf-dispatch-main.yaml
│   ├── wf-dispatch-free.yaml
│   └── wf-dispatch-batch.yaml
├── checklists/
│   ├── pre-dispatch-gate.md
│   ├── pre-execution-gate.md
│   ├── post-execution-gate.md
│   ├── haiku-prompt-checklist.md
│   └── dispatch-health-score.md
├── data/
│   ├── domain-registry.yaml
│   ├── model-selection-rules.yaml
│   ├── veto-conditions.yaml
│   ├── dispatch-heuristics.yaml
│   ├── haiku-patterns.yaml
│   ├── enrichment-rules.yaml
│   ├── timeout-rules.yaml
│   └── dispatch-kb.md
├── templates/
│   ├── execution-plan-tmpl.yaml
│   ├── wave-manifest-tmpl.yaml
│   ├── task-prompt-tmpl.md
│   ├── execution-report-tmpl.md
│   ├── cost-report-tmpl.md
│   └── dispatch-state-tmpl.json
├── scripts/
│   ├── wave-optimizer.py
│   ├── build-command-registry.py
│   ├── cost-tracker.py
│   ├── validate-dispatch-gate.sh
│   ├── validate-wave-results.py
│   ├── enrich-task.py
│   ├── dispatch-health-score.py
│   ├── log-task-start.py
│   ├── check-wave-complete.py
│   ├── save-dispatch-state.py
│   ├── reload-dispatch-context.py
│   └── log-task-failure.py
├── lib/
│   ├── __init__.py
│   ├── pipeline_state.py
│   ├── progress.py
│   └── event_log.py
└── docs/
    ├── ARCHITECTURE.md
    └── MIGRATION.md

Cost Economy (2026)

Model Input/MTok Output/MTok Cache Read
Haiku 4.5 $1.00 $5.00 $0.10
Sonnet 4.5 $3.00 $15.00 $0.30
Opus 4.6 $5.00 $25.00 $0.50

Main context (50k tokens): ~$0.30/interaction Haiku subagent (2k tokens): ~$0.007/task = 43x cheaper Haiku with cache: ~$0.005/task = 58x cheaper

Integration

  • Source: Self-contained in squads/dispatch/ (legacy .aios-core/ dispatcher files have been removed)
  • Auto-discovery: *discover scans ecosystem and rebuilds registries automatically
  • Backward compat: wf-dispatch-free.yaml supports legacy free-text input

Dispatch Squad v1.0.0 — AIOS Framework