Skip to content

comput-agent.md

Latest commit

 

History

History
81 lines (55 loc) · 4.99 KB

comput-agent.md

File metadata and controls

81 lines (55 loc) · 4.99 KB
Raw

Comput AI Agent Guide

The Comput AI Agent (ComputAgent / ComputAgentSheet) is an interactive, step-by-step assistant integrated into the Comput console. Unlike one-shot command generation in Commandium, the Agent plans a multi-step sequence of privileged shell commands and ADB module service executions, runs them with configurable approval gates, and summarizes the results in your language.

Wiki mirror: https://github.com/HmnDev-Tech/shevery/wiki/Comput-Agent

How It Works

  1. Goal Input: Open the Comput tab, tap the Agent button (robot icon), and describe your goal in natural language (for example: "Check which apps wake up the device the most and list disabled system packages").
  2. Planning Phase (PLANNING):
    • The agent queries your active AI provider (configured in Settings → AI → AI provider).
    • Installed ADB modules that include a service.sh script are passed as context (AgentModuleInfo), allowing the agent to trigger module services when relevant.
    • The AI returns a structured JSON plan ({"steps": [...]}) containing up to 10 steps (MAX_STEPS_HARD = 10).
  3. Execution Phase (EXECUTING):
    • Each step has a kind (SHELL or MODULE_SERVICE), a target (the shell command or module ID), and a localized one-sentence comment explaining what the step does.
    • Depending on your Safety Mode, Shevery either executes the step immediately or pauses to show an interactive approval card.
    • Live stdout / stderr output and step status (PENDING, RUNNING, DONE, FAILED, DENIED, SKIPPED) are displayed in real time.
  4. Summary Phase (DONE):
    • Once all steps finish (or if you stop the run early), the agent sends the execution transcript back to the active model to generate a concise verdict and key findings in your current device locale.

Step Types

Step Kind JSON type Target Execution Path
Shell Command shell Raw Android toybox shell command (pm, am, dumpsys, settings, cmd, service, getprop, wm, soc, etc.) Executed via IShizukuService.newProcess(["sh", "-c", cmd])
Module Service module_service Installed ADB module id (must have service.sh) Executed via AdbModuleManager service runner

On-device constraints: The system prompt strictly forbids desktop Linux tools (apt, dpkg, systemctl, journalctl, ifconfig, iptables) that do not exist on Android's toybox environment, and restricts module_service steps to installed modules that actually provide a service.sh script.

Safety Modes & Approval Controls

Configure the agent's behavior under Settings → AI → Comput agent:

Safety Mode Behavior
Secure (secure, default) Prompts for confirmation before every step (both shell commands and module services).
Turbo (turbo) Executes the entire plan automatically without pausing for confirmation.
Custom (custom) Lets you choose the maximum number of steps and toggle confirmation per step type.

Custom Mode Options

When Safety mode is set to Custom, three additional settings appear in Settings → AI:

  • Max steps: 3, 5 (default), or 10 steps per plan.
  • Ask before shell commands: Require approval before running SHELL steps.
  • Ask before module services: Require approval before running MODULE_SERVICE steps.

Per-Step Approval Actions

When a step requires approval, the Agent sheet displays an approval card with four choices:

  • Allow: Run this step once.
  • Edit & Allow: Edit the shell command or module ID inline before executing the step (overrideTarget).
  • Deny: Skip this step (DENIED) and move to the next step in the plan.
  • Allow always: Run this step and all remaining steps in the current plan without asking again.

Chat History

The Agent sheet keeps a history of its chats so a task can be continued later:

  • The history button in the sheet header opens the list of saved chats (title, time, message count), with New chat to start over and a delete button per chat.
  • Opening a chat restores its transcript (your goals and the agent's answers) above the input; the next goal is sent with the previous turns as context, so the agent sees what was already discussed.
  • Chats are stored on device as one JSON string in the same SharedPreferences file as the Comput macros (ModuleSettings.getAgentChats / setAgentChats, key comput_agent_chats, handled by AgentChats). Up to 30 chats and 200 messages per chat are kept; the oldest are dropped.
  • A finished turn is committed to the chat when the next run starts, when you switch chats, or when the sheet is closed.

Source Files

  • manager/src/main/java/moe/shizuku/manager/comput/agent/ComputAgent.kt
  • manager/src/main/java/moe/shizuku/manager/comput/agent/AgentChats.kt
  • manager/src/main/java/moe/shizuku/manager/logs/ComputAgentSheet.kt
  • manager/src/main/java/moe/shizuku/manager/logs/ComputScreen.kt