OpenPocket

Appearance

Prompt Templates ​

This page documents runtime prompt templates used by both task loop and chat gateway flows.

Task System Prompt (buildSystemPrompt) ​

buildSystemPrompt(skillsSummary, workspaceContext, { mode }) supports three modes.

full (default) ​

Includes complete policy stack:

  • Tooling catalog (all task tools):
    • Android actions: tap, swipe, drag, long_press_drag, type_text, keyevent, launch_app, shell, wait, finish
    • script fallback: run_script
    • coding tools: read, write, edit, apply_patch, exec, process
    • memory tools: memory_search, memory_get
    • auth tool: request_human_auth
  • Mandatory planning loop (sub-goal, screen inference, deterministic next action, anti-loop switch)
  • Execution policy (focus anti-loop, avoid repetitive taps, no internal-log typing, practical reproducibility)
  • Human-authorization policy (in-emulator permission dialogs handled locally; real-device checkpoints use request_human_auth)
  • Completion policy (lead finish.message with concrete result values)
  • Output discipline (exactly one tool per step, thought in English)
  • Skill selection protocol
  • Memory recall protocol (search first, then targeted read)
  • Messaging and runtime discipline
  • Self-learning/reuse guidance for reusable flows
  • Available skills summary and workspace context block

minimal ​

Condensed prompt for lower-noise runs (used by cron tasks):

  • one-tool-per-step
  • deterministic progress
  • permission and human-auth basics
  • finish with key outputs
  • includes skills + workspace context when available

none ​

Minimal safety skeleton:

  • one-step tool call
  • permission dialog local handling
  • real auth via request_human_auth
  • finish with concise result

Task User Prompt (buildUserPrompt) ​

Per step, runtime builds a decision prompt with:

  • task text + step number
  • structured screen metadata (currentApp, scaled width/height, device id, timestamp)
  • recent execution history (last 8 lines)
  • stuck signals:
    • trailing action streak
    • trailing app streak
    • unknown-app streak
    • focus-loop risk
  • decision checklist:
    • active sub-goal
    • evidence-based next action
    • anti-loop alternative path
    • text-entry policy (max focus taps, then type and submit)
    • forbid typing logs/history/JSON
    • permission/auth policy reminder
    • finish criteria

The screenshot is attached as base64 image in model payload.

Workspace Context Injection ​

Runtime injects workspace files into system prompt context (subject to char budgets):

  • 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 hook (injected first):

  • .openpocket/bootstrap-context-hook.md

Budget policy:

  • per file max: 20,000 chars
  • total max: agent.contextBudgetChars (default 150,000)
  • truncation strategy: head+tail with explicit middle-truncation marker

Prompt Observability ​

AgentRuntime generates a prompt-context report containing:

  • prompt mode
  • system prompt total chars
  • workspace/non-workspace char split
  • per-file inclusion/truncation/missing/budget-exhausted status
  • skill summary char usage
  • tool list/schema char usage

Gateway exposes this via /context [list|detail|json].

Onboarding Prompt Templates (Chat Assistant) ​

Bootstrap onboarding conductor ​

ChatAssistant builds a strict-JSON prompt that includes:

  • BOOTSTRAP.md
  • SOUL.md
  • current IDENTITY.md and USER.md
  • recent onboarding conversation turns
  • locale hint and current profile snapshot

Model output contract:

json
{
  "reply": "...",
  "profile": {
    "userPreferredAddress": "...",
    "assistantName": "...",
    "assistantPersona": "...",
    "userName": "...",
    "timezone": "...",
    "languagePreference": "..."
  },
  "writeProfile": true,
  "onboardingComplete": false
}

Completion requires all required fields:

  • userPreferredAddress
  • assistantName
  • assistantPersona

Locale onboarding template ​

PROFILE_ONBOARDING.json provides locale text for:

  • step questions
  • empty-answer prompt
  • saved/update/no-change confirmations
  • persona presets and aliases
  • default fallback values

Session Reset Prompt Template ​

BARE_SESSION_RESET_PROMPT.md provides reset/startup guidance text used after /reset when onboarding is not pending.

Task Narration Prompt Templates ​

Progress narrator ​

Prompt source:

  • TASK_PROGRESS_REPORTER.md
  • compact recent progress context
  • locale hint + profile context (SOUL.md, IDENTITY.md, USER.md)

Model output contract:

json
{
  "notify": true,
  "message": "...",
  "reason": "..."
}

Rules include:

  • silence on low-signal/no-visible-progress loops
  • notify on meaningful checkpoints/errors/auth blockers
  • avoid step counters unless user requested telemetry
  • natural conversational tone

Outcome narrator ​

Prompt source:

  • TASK_OUTCOME_REPORTER.md
  • task result, recent progress summary, and artifact flags
  • profile context (SOUL.md, IDENTITY.md, USER.md)

Model output contract:

json
{
  "message": "..."
}

Rules include:

  • lead with concrete findings
  • avoid boilerplate status-first phrasing
  • for failures, include practical next move
  • mention reusable artifacts briefly when generated