OpenPocket

Appearance

Prompting and Decision Model ​

This page explains how OpenPocket builds prompts across onboarding, task execution, and user-facing narration.

Task Prompt Stack ​

System prompt ​

buildSystemPrompt(skillsSummary, workspaceContext, { mode }) builds policy instructions for the task loop.

Core layers in full mode:

  • complete tool catalog (Android actions + script + coding + memory + human auth)
  • mandatory planning loop and anti-loop strategy switch
  • execution policy for deterministic one-step actions
  • permission/auth policy
  • completion policy (result-first finish.message)
  • skill-selection protocol
  • memory-recall protocol (memory_search -> memory_get)
  • self-learning/reuse guidance

Modes:

  • full (default)
  • minimal (used by cron runs)
  • none (minimal safety shell)

User prompt (per step) ​

buildUserPrompt(task, step, snapshot, history) includes:

  • task + step
  • screen metadata and scaled coordinate bounds
  • recent history
  • stuck indicators (action streak, app streak, unknown-app streak, focus-loop risk)
  • explicit decision checklist

The screenshot is attached as image input in the same model request.

Workspace Context Injection ​

Runtime injects workspace prompt files into system prompt context:

  • AGENTS.md, BOOTSTRAP.md, SOUL.md, USER.md, IDENTITY.md
  • TOOLS.md, HEARTBEAT.md, MEMORY.md
  • TASK_PROGRESS_REPORTER.md, TASK_OUTCOME_REPORTER.md

Optional pre-hook:

  • .openpocket/bootstrap-context-hook.md

Budget strategy:

  • per file: up to 20,000 chars
  • total: agent.contextBudgetChars (default 150,000)
  • truncation uses head+tail keep with explicit marker

Prompt Observability ​

AgentRuntime generates a prompt report (system size, per-file status, skill/tool budget usage).

Gateway command /context exposes this report in:

  • summary list
  • detailed breakdown
  • raw JSON

Onboarding Prompting ​

Chat onboarding is model-driven when profile is incomplete.

Inputs to onboarding conductor prompt:

  • BOOTSTRAP.md
  • locale hint
  • profile snapshot from IDENTITY.md + USER.md
  • recent onboarding turns
  • SOUL.md

Model returns strict JSON (reply, profile patch, writeProfile, onboardingComplete).

Template support:

  • PROFILE_ONBOARDING.json provides locale questions, persona presets, and confirmations
  • BARE_SESSION_RESET_PROMPT.md controls post-/reset startup guidance when onboarding is already complete

Progress and Outcome Prompting ​

During task execution, gateway does not hardcode per-step text. It asks model narrators with structured context.

Progress narrator ​

  • prompt guide: TASK_PROGRESS_REPORTER.md
  • output JSON: { notify, message, reason }
  • notify only on meaningful user-visible progress
  • avoid step counters unless user asked for telemetry

Outcome narrator ​

  • prompt guide: TASK_OUTCOME_REPORTER.md
  • output JSON: { message }
  • lead with concrete results
  • avoid generic "task completed" boilerplate when result data exists

Gateway adds extra suppression to avoid repetitive low-signal updates on the same screen.

Output Contract and Safety ​

  • model must call exactly one tool per step
  • tool args are normalized to AgentAction
  • unknown/malformed actions degrade to safe wait
  • in-emulator Android permission dialogs are auto-approved locally
  • request_human_auth is reserved for real-device checkpoints and sensitive authorization flows

Chat Routing ​

ChatAssistant.decide(chatId, inputText) routing order:

  1. bootstrap/profile onboarding gates
  2. profile update intent check (name/persona/address)
  3. model classification (task vs chat)
  4. fallback to task if classifier fails

Task mode delegates to AgentRuntime.runTask. Chat mode replies conversationally.