pi-modus

Amp style mode switching and explicit named subagents for Pi.

Installation

pi install git:github.com/ofrades/pi-modus

Concepts

Modes are named behavioral profiles (rush, smart, deep) each bound to a specific model and thinking level. The prompt injected for each mode matches Amp's exact system prompts — they are behavioral contracts, not just model presets.

Mode	Intent
rush	Speed first: small focused changes, minimal explanation, ultra-concise, verify narrowly
smart	Unconstrained state-of-the-art: end-to-end agency, simple-first, parallel by default
deep	Pragmatic senior engineer: build broad context, smallest correct change, YAGNI/KISS

Subagents are named, explicitly requested assistants with their own model and thinking level. They are available as tools but are never proactively recommended by the mode prompt. The intended pattern is user-directed delegation:

"Ask the oracle to inspect these files first, then implement the smallest patch."

Available subagents:

Subagent	Purpose
search	Fast parallel codebase retrieval (8+ parallel calls, 3-turn limit)
vision	Image and screenshot analysis
review	Code review with git diff focus, abstraction-fit evaluation
oracle	Planning, tradeoffs, 6-part structured response (TL;DR → risks → effort signal)
librarian	External docs, dependencies, API research

Subagent system prompts are copied verbatim from Amp's binary.

Commands

/modus

Opens the interactive modus selector for both modes and subagents.

/modus              # open selector UI
/modus rush         # switch directly to rush
/modus smart        # switch directly to smart
/modus deep         # switch directly to deep
/modus cost         # show session and 7-day cost summary
/modus subagents    # list configured subagents
/modus subagents set oracle openai/gpt-5.5:high
/modus subagents unset oracle

Inside the selector:

• ↑↓ / j / k — navigate modes and subagents
• Enter — apply selected mode, or choose a model for selected subagent
• t — change thinking level for selected mode/subagent
• c — change model for selected mode/subagent
• Esc — cancel

Ctrl+Shift+M

Cycles forward through rush → smart → deep → rush.

Amp models

Modes:
- Rush: gpt-5.5, no reasoning
- Smart: claude-opus-4.7
- Deep: gpt-5.5, medium reasoning

Subagents:
 - Review: gemini 3.1-pro
 - Search: gemini 3-flash
 - Oracle: gpt-5.4, medium reasoning
 - Librarian: claude-sonnet-4.6

Tool:
 - Vision: gemini-3-flash

Configuration

Config is stored in settings.json under the agent directory:

These are cost-effective opencode-go models. Swap to the Amp equivalents if you prefer.

{
  "modus": {
    "activeMode": "smart",
    "modes": {
      "rush":  { "provider": "opencode-go", "model": "deepseek-v4-flash", "thinkingLevel": "off" },
      "smart": { "provider": "opencode-go", "model": "kimi-k2.6", "thinkingLevel": "medium" },
      "deep":  { "provider": "opencode-go", "model": "deepseek-v4-pro", "thinkingLevel": "medium" }
    },
    "subagents": {
      "review":    { "provider": "opencode-go", "model": "deepseek-v4-pro" },
      "search":    { "provider": "opencode-go", "model": "deepseek-v4-flash", "thinkingLevel": "off" },
      "oracle":    { "provider": "opencode-go", "model": "kimi-k2.5", "thinkingLevel": "medium" },
      "librarian": { "provider": "opencode-go", "model": "kimi-k2.5" },
      "vision":    { "provider": "opencode-go", "model": "kimi-k2.5" } 
    }
  }
}

Notes

• thinkingLevel defaults to the highest level supported by the configured model if not explicitly set.
• Subagent costs are tracked separately in the cost log and appear in /modus cost.
• Subagents run read-only. They cannot edit files.