multi-team-sdd

pi package for multi-agent orchestration with SDD-aware routing.

Includes

• Orchestrator policy injected into system prompt
• subagent tool with single, parallel, and chain modes
• Specialized agents:
  • orchestrator
  • explorer
  • documentator
  • planner
  • implementer
  • tester-reviewer
  • hacker
• Horizontal subagent cards widget on startup (name + short description + color)
• Commands:
  • /subagents
  • /orchestrator-status
  • /security-mode passive|active

Install (local path)

pi install /absolute/path/to/multi-team-sdd

Temporary run:

pi -e /absolute/path/to/multi-team-sdd

Setup Codex

This repo also ships Codex-native agent configuration.

Install globally:

./setup.sh codex --global

Install into a specific project:

./setup.sh codex --project /absolute/path/to/project

Install both global and project config:

./setup.sh codex --global --project /absolute/path/to/project

The Codex setup installs:

• ~/.codex/agents/*.toml or <project>/.codex/agents/*.toml
• pipeline.json
• AGENTS.md instructions with a managed multi-sdd-team block
• service_tier = "fast" and [features].fast_mode = true
• [agents].max_threads = 6 and [agents].max_depth = 1

Codex demo-fast behavior:

• the main session acts as orchestrator
• SDD_SUBAGENTS runs sequentially: explorer -> documentator -> planner -> implementer -> tester-reviewer -> main integration
• hacker is skipped unless explicitly requested or security-sensitive
• if tester-reviewer asks for small demo-blocking fixes, the main session applies them directly instead of calling implementer again

Quick usage

• Ask normally; the orchestrator policy is injected automatically.
• Force delegation explicitly with tool usage, for example:
  • single: explorer
  • chain: explorer -> planner -> implementer -> tester-reviewer
• Show active catalog: /subagents
• Check mode/status: /orchestrator-status
• Change security mode: /security-mode passive|active

Notes

• documentator is constrained to ./docs/** write/edit operations.
• tester-reviewer is report-only in v1.
• hacker supports passive and active modes. Active mode allows high-risk commands and requires explicit opt-in (/security-mode active).
• Startup UI renders horizontal cards with each specialized subagent (name + short description + role color).