rysweet/amplihack-rs/amplihack-expert

amplihack Expert Knowledge

What is amplihack?

Engineering system for coding CLIs (Claude, Copilot, Amplifier): 5 mechanisms, 23-step workflow, 30+ agents, 80+ skills. Core entry point: /dev <task> — unified task orchestrator with auto-classification, parallel workstream detection, and goal-seeking execution loop.

Quick Reference

Top Commands

Top Agents

Workflows

Navigation Guide

Quick: SKILL.md | Architecture: reference.md | How-To: examples.md | Deep: all

Core Concepts

5 Mechanisms: Workflow (process), Command (entry), Skill (auto), Agent (delegate), Hook (runtime)

Composition: Commands → Workflows → Agents → Skills

Execution: /dev orchestrates — classifies task, detects parallel workstreams, executes via smart-orchestrator recipe, goal-seeking loop (3 rounds max)

Entry point: /dev <task> → dev-orchestrator skill → smart-orchestrator recipe → default-workflow recipe

dev-orchestrator Architecture

Entry point: /dev <task> or any non-trivial prompt

Routing:

• Q&A → amplihack:core:analyzer responds directly
• Operations → bash (simple) or amplihack:core:builder (complex)
• Development/Investigation → smart-orchestrator recipe (full orchestration)

Execution flow:

1. Classify task (architect agent → JSON decomposition)
2. Detect workstreams (1 = single session, 2-5 = parallel via multitask)
3. Register session in tree (depth/capacity enforcement)
4. Execute rounds (up to 3, goal-seeking reflection loop)
5. Summarize results with PR links and GOAL_STATUS

Session tree: Prevents infinite recursion

• AMPLIHACK_MAX_DEPTH=3 (default) — increase to 5 for deep orchestration
• AMPLIHACK_MAX_SESSIONS=10 (default) — max concurrent sessions per tree

Status signals:

• GOAL_STATUS: ACHIEVED — all criteria met
• GOAL_STATUS: PARTIAL -- [gaps] — another round will run
• GOAL_STATUS: NOT_ACHIEVED -- [reason] — final failure status

Related Docs

• reference.md: Architecture (5 mechanisms, 5 layers, hooks)
• examples.md: Scenarios (5+ real examples)
• @~/.amplihack/.claude/context/PHILOSOPHY.md: Core principles
• default-workflow skill/recipe: 23 steps
• @~/.amplihack/.claude/agents/amplihack/: All agents
• @~/.amplihack/.claude/commands/amplihack/: All commands
• @~/.amplihack/.claude/tools/amplihack/hooks/: Hook system
• amplifier-bundle/recipes/smart-orchestrator.yaml: Core orchestration recipe
• Native recursion guard via AMPLIHACK_MAX_DEPTH env var
• amplihack orch helper: Native Rust JSON extraction (source: crates/amplihack-cli/src/commands/orch.rs)

Known Failure Points

Rate-limit resilience

When an upstream API (Copilot, Anthropic) returns a rate-limit error during an agent step, the recipe runner treats it as a hard failure by default. Non-critical steps such as step-06c-documentation-refinement set continue_on_error: true so that documentation polish does not abort the entire recipe. For critical steps, the agent runtime retries with exponential back-off (configured by the SDK adapter). If a rate-limit error aborts your workflow, check which step failed — if it is a polish/review step, adding continue_on_error: true is the recommended fix.