@elizaos/prompts 2.0.0-alpha.43 → 2.0.0-alpha.430

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@elizaos/prompts",
   "private": false,
-  "version": "2.0.0-alpha.43",
+  "version": "2.0.0-alpha.430",
   "description": "Shared prompt templates for elizaOS across TypeScript, Python, and Rust",
   "type": "module",
   "main": "./dist/typescript/index.ts",
@@ -15,18 +15,17 @@
     "./rust": "./dist/rust/prompts.rs"
   },
   "scripts": {
-    "build": "node scripts/generate.js && node scripts/generate-plugin-action-spec.js && node scripts/generate-action-docs.js",
+    "build": "node scripts/generate.js && node scripts/generate-plugin-action-spec.js && bunx @biomejs/biome format --write specs/actions/plugins.generated.json && node scripts/generate-action-docs.js",
     "build:typescript": "node scripts/generate.js --target typescript",
     "build:python": "node scripts/generate.js --target python",
     "build:rust": "node scripts/generate.js --target rust",
     "build:action-docs": "node scripts/generate-action-docs.js",
-    "build:plugin-action-spec": "node scripts/generate-plugin-action-spec.js",
+    "build:plugin-action-spec": "node scripts/generate-plugin-action-spec.js && bunx @biomejs/biome format --write specs/actions/plugins.generated.json",
     "check:secrets": "node scripts/check-secrets.js",
     "clean": "rm -rf dist",
-    "lint": "bunx @biomejs/biome check --write --unsafe .",
+    "lint": "bunx @biomejs/biome check --write .",
     "lint:check": "bunx @biomejs/biome check .",
-    "typecheck": "echo 'No TypeScript source files to check'",
-    "test": "node --test test/*.test.js"
+    "typecheck": "echo 'No TypeScript source files to check'"
   },
   "files": [
     "prompts/",
@@ -44,5 +43,5 @@
   "publishConfig": {
     "access": "public"
   },
-  "gitHead": "921845fc70406f93b8649731260e3254228cb21a"
+  "gitHead": "c3c6281e6b43f1bdf8b54fcec1a00986c5828cf8"
 }

package/prompts/add_contact.txt

@@ -0,0 +1,29 @@
+task: Extract contact information to add to the relationships.
+
+context:
+{{providers}}
+
+recent_messages:
+{{recentMessages}}
+
+current_message:
+{{message}}
+
+instructions[5]:
+- identify the contact name being added
+- include entityId only if it is explicitly known from context
+- return categories as a comma-separated list
+- include notes, timezone, and language only when clearly present
+- include a short reason for why this contact should be saved
+
+output:
+TOON only. Return exactly one TOON document. No prose before or after it. No <think>.
+
+Example:
+contactName: Jane Doe
+entityId:
+categories: vip,colleague
+notes: Met at the design summit
+timezone: America/New_York
+language: English
+reason: Important collaborator to remember

package/prompts/choose_option.txt

@@ -9,14 +9,11 @@
 Analyze the options and select the most appropriate one based on the current context.
 Provide your reasoning and the selected option ID.
 
-Respond using XML format like this:
-<response>
-    <thought>Your reasoning for the selection</thought>
-    <selected_id>The ID of the selected option</selected_id>
-</response>
-
-IMPORTANT: Your response must ONLY contain the <response></response> XML block above.
+Respond using TOON like this:
+thought: Your reasoning for the selection
+selected_id: The ID of the selected option
 
+IMPORTANT: Your response must ONLY contain the TOON document above.
 
 
 

package/prompts/extract_secret_operation.txt

@@ -0,0 +1,20 @@
+You are helping manage secrets for an AI agent.
+
+Determine what operation the user wants to perform:
+- get: Retrieve a secret value
+- set: Store a new secret
+- delete: Remove a secret
+- list: Show all available secrets (without values)
+- check: Check if a secret exists
+
+Common patterns:
+- "What is my OpenAI key?" -> operation: get, key: OPENAI_API_KEY
+- "Do I have a Discord token set?" -> operation: check, key: DISCORD_BOT_TOKEN
+- "Show me my secrets" -> operation: list
+- "Delete my old API key" -> operation: delete
+- "Remove TWITTER_API_KEY" -> operation: delete, key: TWITTER_API_KEY
+- "Set my key to sk-..." -> operation: set, key: <infer>, value: sk-...
+
+{{recentMessages}}
+
+Extract the operation, key (if applicable), value (if applicable), and level from the user's message.

package/prompts/extract_secret_request.txt

@@ -0,0 +1,16 @@
+You are helping an AI agent request a missing secret.
+Determine what secret the agent needs and why based on the recent conversation.
+
+Common patterns:
+- "I need an API key for OpenAI" -> key: OPENAI_API_KEY
+- "Missing TWITTER_TOKEN" -> key: TWITTER_TOKEN
+- "I cannot proceed without a Discord token" -> key: DISCORD_TOKEN
+
+Recent Messages:
+{{recentMessages}}
+
+Output JSON with:
+- key: The name of the secret needed (e.g. OPENAI_API_KEY)
+- reason: Why it is needed (optional)
+
+If no specific secret is requested, return null json.

package/prompts/extract_secrets.txt

@@ -0,0 +1,17 @@
+You are extracting secret/configuration values from the user's message.
+
+The user wants to set one or more secrets. Extract:
+1. The secret key (should be UPPERCASE_WITH_UNDERSCORES format)
+2. The secret value
+3. Optional description
+4. Secret type (api_key, secret, credential, url, or config)
+
+Common patterns:
+- "Set my OpenAI key to sk-..." -> key: OPENAI_API_KEY, value: sk-...
+- "My Anthropic API key is sk-ant-..." -> key: ANTHROPIC_API_KEY, value: sk-ant-...
+- "Use this Discord token: ..." -> key: DISCORD_BOT_TOKEN, value: ...
+- "Set DATABASE_URL to postgres://..." -> key: DATABASE_URL, value: postgres://...
+
+{{recentMessages}}
+
+Extract the secrets from the user's message. If the key name isn't explicitly specified, infer an appropriate UPPERCASE_WITH_UNDERSCORES name based on the context.

package/prompts/image_description.txt

@@ -1,28 +1,21 @@
-<task>Analyze the provided image and generate a comprehensive description with multiple levels of detail.</task>
+Task: Analyze the provided image and generate a comprehensive description with multiple levels of detail.
 
-<instructions>
+Instructions:
 Carefully examine the image and provide:
 1. A concise, descriptive title that captures the main subject or scene
 2. A brief summary description (1-2 sentences) highlighting the key elements
 3. An extensive, detailed description that covers all visible elements, composition, lighting, colors, mood, and any other relevant details
 
 Be objective and descriptive. Focus on what you can actually see in the image rather than making assumptions about context or meaning.
-</instructions>
 
-<output>
-Do NOT include any thinking, reasoning, or <think> sections in your response. 
-Go directly to the XML response format without any preamble or explanation.
+Output:
 
-Respond using XML format like this:
-<response>
-  <title>A concise, descriptive title for the image</title>
-  <description>A brief 1-2 sentence summary of the key elements in the image</description>
-  <text>An extensive, detailed description covering all visible elements, composition, lighting, colors, mood, setting, objects, people, activities, and any other relevant details you can observe in the image</text>
-</response>
-
-IMPORTANT: Your response must ONLY contain the <response></response> XML block above. Do not include any text, thinking, or reasoning before or after this XML block. Start your response immediately with <response> and end with </response>.
-</output>
+Respond using TOON like this:
+title: A concise, descriptive title for the image
+description: A brief 1-2 sentence summary of the key elements in the image
+text: An extensive, detailed description covering all visible elements, composition, lighting, colors, mood, setting, objects, people, activities, and any other relevant details you can observe in the image
 
+IMPORTANT: Your response must ONLY contain the TOON document above. Do not include any text, thinking, or reasoning before or after it.
 
 
 

package/prompts/image_generation.txt

@@ -9,14 +9,11 @@ The prompt should be specific, descriptive, and suitable for AI image generation
 # Recent conversation:
 {{recentMessages}}
 
-Respond using XML format like this:
-<response>
-    <thought>Your reasoning for the image prompt</thought>
-    <prompt>Detailed image generation prompt</prompt>
-</response>
-
-IMPORTANT: Your response must ONLY contain the <response></response> XML block above.
+Respond using TOON like this:
+thought: Your reasoning for the image prompt
+prompt: Detailed image generation prompt
 
+IMPORTANT: Your response must ONLY contain the TOON document above.
 
 
 

package/prompts/initial_summarization.txt

@@ -0,0 +1,28 @@
+# Task: Summarize Conversation
+
+You are analyzing a conversation to create a concise summary that captures the key points, topics, and important details.
+
+# Recent Messages
+{{recentMessages}}
+
+# Instructions
+Generate a summary that:
+1. Captures the main topics discussed
+2. Highlights key information shared
+3. Notes any decisions made or questions asked
+4. Maintains context for future reference
+5. Is concise but comprehensive
+
+**IMPORTANT**: Keep the summary under 2500 tokens. Be comprehensive but concise.
+
+Also extract:
+- **Topics**: List of main topics discussed (comma-separated)
+- **Key Points**: Important facts or decisions (bullet points)
+
+Respond in TOON:
+text: Your comprehensive summary here
+topics[0]: topic1
+topics[1]: topic2
+topics[2]: topic3
+keyPoints[0]: First key point
+keyPoints[1]: Second key point

package/prompts/long_term_extraction.txt

@@ -0,0 +1,128 @@
+# Task: Extract Long-Term Memory (Strict Criteria)
+
+You are analyzing a conversation to extract ONLY the most critical, persistent information about the user using cognitive science memory categories.
+
+# Recent Messages
+{{recentMessages}}
+
+# Current Long-Term Memories
+{{existingMemories}}
+
+# Memory Categories (Based on Cognitive Science)
+
+## 1. EPISODIC Memory
+Personal experiences and specific events with temporal/spatial context.
+**Examples:**
+- "User completed migration project from MongoDB to PostgreSQL in Q2 2024"
+- "User encountered authentication bug in production on March 15th"
+- "User had a negative experience with Docker networking in previous job"
+
+**Requirements:**
+- Must include WHO did WHAT, WHEN/WHERE
+- Must be a specific, concrete event (not a pattern)
+- Must have significant impact or relevance to future work
+
+## 2. SEMANTIC Memory
+General facts, concepts, knowledge, and established truths about the user.
+**Examples:**
+- "User is a senior backend engineer with 8 years experience"
+- "User specializes in distributed systems and microservices architecture"
+- "User's primary programming language is TypeScript"
+- "User works at Acme Corp as technical lead"
+
+**Requirements:**
+- Must be factual, timeless information
+- Must be explicitly stated or demonstrated conclusively
+- No speculation or inference from single instances
+- Core identity, expertise, or knowledge only
+
+## 3. PROCEDURAL Memory
+Skills, workflows, methodologies, and how-to knowledge.
+**Examples:**
+- "User follows strict TDD workflow: write tests first, then implementation"
+- "User prefers git rebase over merge to maintain linear history"
+- "User's debugging process: check logs → reproduce locally → binary search"
+- "User always writes JSDoc comments before implementing functions"
+
+**Requirements:**
+- Must describe HOW user does something
+- Must be a repeated, consistent pattern (seen 3+ times or explicitly stated as standard practice)
+- Must be a workflow, methodology, or skill application
+- Not one-off preferences
+
+# ULTRA-STRICT EXTRACTION CRITERIA
+
+## DO EXTRACT (Only These):
+
+**EPISODIC:**
+- Significant completed projects or milestones
+- Important bugs, incidents, or problems encountered
+- Major decisions made with lasting impact
+- Formative experiences that shape future work
+
+**SEMANTIC:**
+- Professional identity (role, title, company)
+- Core expertise and specializations (stated explicitly or demonstrated conclusively)
+- Primary languages, frameworks, or tools (not exploratory use)
+- Established facts about their work context
+
+**PROCEDURAL:**
+- Consistent workflows demonstrated 3+ times or explicitly stated
+- Standard practices user always follows
+- Methodology preferences with clear rationale
+- Debugging, testing, or development processes
+
+## NEVER EXTRACT:
+
+- **One-time requests or tasks** (e.g., "can you generate an image", "help me debug this")
+- **Casual conversations** without lasting significance
+- **Exploratory questions** (e.g., "how does X work?")
+- **Temporary context** (current bug, today's task)
+- **Preferences from single occurrence** (e.g., user asked for code once)
+- **Social pleasantries** (thank you, greetings)
+- **Testing or experimentation** (trying out a feature)
+- **Common patterns everyone has** (likes clear explanations)
+- **Situational information** (working on feature X today)
+- **Opinions without persistence** (single complaint, isolated praise)
+- **General knowledge** (not specific to user)
+
+# Quality Gates (ALL Must Pass)
+
+1. **Significance Test**: Will this matter in 3+ months?
+2. **Specificity Test**: Is this concrete and actionable?
+3. **Evidence Test**: Is there strong evidence (3+ instances OR explicit self-identification)?
+4. **Uniqueness Test**: Is this specific to THIS user (not generic)?
+5. **Confidence Test**: Confidence must be >= 0.85 (be VERY conservative)
+6. **Non-Redundancy Test**: Does this add NEW information not in existing memories?
+
+# Confidence Scoring (Be Conservative)
+
+- **0.95-1.0**: User explicitly stated as core identity/practice AND demonstrated multiple times
+- **0.85-0.94**: User explicitly stated OR consistently demonstrated 5+ times
+- **0.75-0.84**: Strong pattern (3-4 instances) with supporting context
+- **Below 0.75**: DO NOT EXTRACT (insufficient evidence)
+
+# Critical Instructions
+
+1. **Default to NOT extracting** - When in doubt, skip it
+2. **Require overwhelming evidence** - One or two mentions is NOT enough
+3. **Focus on what's PERSISTENT** - Not what's temporary or situational
+4. **Verify against existing memories** - Don't duplicate or contradict
+5. **Maximum 2-3 extractions per run** - Quality over quantity
+
+**If there are no qualifying facts (which is common), return no memories entries.**
+
+# Response Format
+
+memories[0]:
+  category: semantic
+  content: User is a senior TypeScript developer with 8 years of backend experience
+  confidence: 0.95
+memories[1]:
+  category: procedural
+  content: User follows TDD workflow: writes tests before implementation, runs tests after each change
+  confidence: 0.88
+memories[2]:
+  category: episodic
+  content: User led database migration from MongoDB to PostgreSQL for payment system in Q2 2024
+  confidence: 0.92

package/prompts/message_classifier.txt

@@ -0,0 +1,37 @@
+Analyze this user request and classify it for planning purposes:
+
+"{{text}}"
+
+Classify the request across these dimensions:
+
+1. COMPLEXITY LEVEL:
+- simple: Direct actions that don't require planning
+- medium: Multi-step tasks requiring coordination
+- complex: Strategic initiatives with multiple stakeholders
+- enterprise: Large-scale transformations with full complexity
+
+2. PLANNING TYPE:
+- direct_action: Single action, no planning needed
+- sequential_planning: Multiple steps in sequence
+- strategic_planning: Complex coordination with stakeholders
+
+3. REQUIRED CAPABILITIES:
+- List specific capabilities needed (analysis, communication, project_management, etc.)
+
+4. STAKEHOLDERS:
+- List types of people/groups involved
+
+5. CONSTRAINTS:
+- List limitations or requirements mentioned
+
+6. DEPENDENCIES:
+- List dependencies between tasks or external factors
+
+Respond in this exact format:
+COMPLEXITY: [simple|medium|complex|enterprise]
+PLANNING: [direct_action|sequential_planning|strategic_planning]
+CAPABILITIES: [comma-separated list]
+STAKEHOLDERS: [comma-separated list]
+CONSTRAINTS: [comma-separated list]
+DEPENDENCIES: [comma-separated list]
+CONFIDENCE: [0.0-1.0]

package/prompts/message_handler.txt

@@ -1,75 +1,49 @@
-<task>Generate dialog and actions for the character {{agentName}}.</task>
+task: Generate dialog and actions for {{agentName}}.
 
-<providers>
+context:
 {{providers}}
-</providers>
 
-<instructions>
-Write a thought and plan for {{agentName}} and decide what actions to take.
-
-IMPORTANT ACTION ORDERING RULES:
-- Actions are executed in the ORDER you list them - the order MATTERS!
-- REPLY should come FIRST to acknowledge the user's request before executing other actions
-- Common patterns:
-  - For requests requiring tool use: REPLY,CALL_MCP_TOOL (acknowledge first, then gather info)
-  - For task execution: REPLY,SEND_MESSAGE or REPLY,EVM_SWAP_TOKENS (acknowledge first, then do the task)
-  - For multi-step operations: REPLY,ACTION1,ACTION2 (acknowledge first, then complete all steps)
-- REPLY is used to acknowledge and inform the user about what you're going to do
-- Follow-up actions execute the actual tasks after acknowledgment
-- Use IGNORE only when you should not respond at all
-- If you use IGNORE, do not include any other actions. IGNORE should be used alone when you should not respond or take any actions.
-
-IMPORTANT ACTION PARAMETERS:
-- Some actions accept input parameters that you should extract from the conversation
-- When an action has parameters listed in its description, include a <params> block for that action
-- Extract parameter values from the user's message and conversation context
-- Required parameters MUST be provided; optional parameters can be omitted if not mentioned
-- If you cannot determine a required parameter value, ask the user for clarification in your <text>
-
-IMPORTANT CODE BLOCK FORMATTING RULES:
-- If {{agentName}} includes code examples, snippets, or multi-line code in the response, ALWAYS wrap the code with ``` fenced code blocks (specify the language if known, e.g., ```python).
-- ONLY use fenced code blocks for actual code. Do NOT wrap non-code text, instructions, or single words in fenced code blocks.
-- If including inline code (short single words or function names), use single backticks (`) as appropriate.
-- This ensures the user sees clearly formatted and copyable code when relevant.
-
-First, think about what you want to do next and plan your actions. Then, write the next message and include the actions you plan to take.
-</instructions>
-
-<keys>
-"thought" should be a short description of what the agent is thinking about and planning.
-"actions" should be a comma-separated list of the actions {{agentName}} plans to take based on the thought, IN THE ORDER THEY SHOULD BE EXECUTED (if none, use IGNORE, if simply responding with text, use REPLY)
-"text" should be the text of the next message for {{agentName}} which they will send to the conversation.
-"params" (optional) should contain action parameters when actions require input. Format as nested XML with action name as wrapper.
-</keys>
-
-<output>
-Do NOT include any thinking, reasoning, or <think> sections in your response. 
-Go directly to the XML response format without any preamble or explanation.
-
-Respond using XML format like this:
+rules[11]:
+- think briefly, then respond
+- always include a <thought> field, even for direct replies
+- actions execute in listed order
+- if replying without another grounded state/action query, REPLY goes first
+- use IGNORE or STOP only by themselves
+- include providers only when needed
+- use provider_hints from context when present instead of restating the same rules
+- if an action needs inputs, include them inside that action's <params> block
+- if a required param is unknown, ask for clarification in text
+- for live status questions or remaining-work queries, do not answer from recent conversation alone; call the relevant action/provider to refresh state, and do not pair it with a speculative REPLY that guesses the result
+- when an action will fetch the state and produce the final grounded answer, do not add REPLY just to say "checking", "let me look", or similar filler; use the action alone and leave text empty
+- for LifeOps create requests with a clear defaultable habit or natural window, such as drinking water, stretch breaks during the day, weekday-after-lunch Invisalign checks, or brushing when waking up and before bed, call LIFE instead of asking for exact clock times unless the user explicitly asks for precise scheduling
+
+control_actions:
+- STOP means the task is done and the agent should end the run without executing more actions
+- STOP is a terminal control action even if it is not listed in available actions
+
+fields[5]{name,meaning}:
+- thought | short plan
+- actions | ordered <action> entries inside <actions>
+- providers | array of provider names, or empty
+- text | next message for {{agentName}}
+- simple | true or false
+
+formatting:
+- wrap multi-line code in fenced code blocks
+- use inline backticks for short code identifiers
+
+output:
+XML only. Return exactly one <response>...</response> document. No prose before or after it. No <think>.
+
+Example:
 <response>
-    <thought>Your thought here</thought>
-    <actions>ACTION1,ACTION2</actions>
-    <text>Your response text here</text>
-    <params>
-        <ACTION1>
-            <paramName1>value1</paramName1>
-            <paramName2>value2</paramName2>
-        </ACTION1>
-        <ACTION2>
-            <paramName1>value1</paramName1>
-        </ACTION2>
-    </params>
+  <thought>Reply briefly. No extra providers needed.</thought>
+  <actions>
+    <action>
+      <name>REPLY</name>
+    </action>
+  </actions>
+  <providers></providers>
+  <text>Your message here</text>
+  <simple>true</simple>
 </response>
-
-The <params> block is optional - only include when actions require input parameters.
-If an action has no parameters or you're only using REPLY/IGNORE, omit <params> entirely.
-
-IMPORTANT: Your response must ONLY contain the <response></response> XML block above. Do not include any text, thinking, or reasoning before or after this XML block. Start your response immediately with <response> and end with </response>.
-</output>
-
-
-
-
-
-

package/prompts/multi_step_decision.txt

@@ -1,6 +1,4 @@
-<task>
 Determine the next step the assistant should take in this conversation to help the user reach their goal.
-</task>
 
 {{recentMessages}}
 
@@ -26,29 +24,19 @@ These are the actions or data provider calls that have already been used in this
 
 {{actionResults}}
 
-<keys>
+keys:
 "thought" Clearly explain your reasoning for the selected providers and/or action, and how this step contributes to resolving the user's request.
 "action"  Name of the action to execute after providers return (can be empty if no action is needed).
 "providers" List of provider names to call in this step (can be empty if none are needed).
-"params" Optional XML parameters for the selected action. Use nested XML only when the action needs input.
 "isFinish" Set to true only if the task is fully complete.
-</keys>
 
 ⚠️ IMPORTANT: Do **not** mark the task as `isFinish: true` immediately after calling an action. Wait for the action to complete before deciding the task is finished.
 
-<output>
-<response>
-  <thought>Your thought here</thought>
-  <action>ACTION</action>
-  <providers>PROVIDER1,PROVIDER2</providers>
-  <params>
-    <ACTION>
-      <paramName>value</paramName>
-    </ACTION>
-  </params>
-  <isFinish>true | false</isFinish>
-</response>
-</output>
+output:
+thought: Your thought here
+action: ACTION
+providers[2]: PROVIDER1,PROVIDER2
+isFinish: false
 
 
 

package/prompts/multi_step_summary.txt

@@ -1,6 +1,4 @@
-<task>
 Summarize what the assistant has done so far and provide a final response to the user based on the completed steps.
-</task>
 
 # Context Information
 {{bio}}
@@ -28,14 +26,10 @@ Here are the actions taken by the assistant to fulfill the request:
 
  - Review the execution trace and last reasoning step carefully
 
- - Your final output MUST be in this XML format:
-<output>
-<response>
-  <thought>Your thought here</thought>
-  <text>Your final message to the user</text>
-</response>
-</output>
-
+ - Your final output MUST be TOON in this format:
+output:
+thought: Your thought here
+text: Your final message to the user
 
 
 

package/prompts/option_extraction.txt

@@ -12,17 +12,12 @@
 3. Return the task ID (shortened UUID) and selected option name exactly as listed above
 4. If no clear selection is made, return null for both fields
 
-Do NOT include any thinking, reasoning, or <think> sections in your response. 
-Go directly to the XML response format without any preamble or explanation.
 
-Return in XML format:
-<response>
-  <taskId>string_or_null</taskId>
-  <selectedOption>OPTION_NAME_or_null</selectedOption>
-</response>
-
-IMPORTANT: Your response must ONLY contain the <response></response> XML block above. Do not include any text, thinking, or reasoning before or after this XML block. Start your response immediately with <response> and end with </response>.
+Return in TOON format:
+taskId: string_or_null
+selectedOption: OPTION_NAME_or_null
 
+IMPORTANT: Your response must ONLY contain the TOON document above. Do not include any text, thinking, or reasoning before or after it.
 
 
 

package/prompts/post_action_decision.txt

@@ -0,0 +1,35 @@
+Continue helping the user after reviewing the latest action results.
+
+context:
+{{providers}}
+
+recent conversation:
+{{recentMessages}}
+
+recent action results:
+{{actionResults}}
+
+latest reflection task status:
+{{taskCompletionStatus}}
+
+rules[10]:
+- think briefly, then continue the task from the latest action results
+- actions execute in listed order
+- if replying, REPLY goes first
+- use IGNORE or STOP only by themselves
+- include providers only when needed
+- use provider_hints from context when present instead of restating the same rules
+- if an action needs inputs, include them under params keyed by action name
+- if a required param is unknown, ask for clarification in text
+- if reflection says the task is incomplete, keep working or explain the concrete follow-up you still need
+- if the task is complete, either reply to the user or use STOP to end the run
+- STOP is a terminal control action even if it is not listed in available actions
+
+output:
+TOON only. Return exactly one TOON document. No prose before or after it. No <think>.
+
+thought: Your thought here
+actions[1]: ACTION
+providers[0]:
+text: Your message here
+simple: true