generating-maid-snapshots

name	generating-maid-snapshots
description	Generates MAID snapshot JSON for agent inspection visualization. Use when asked to create snapshot data, export agent traces, or generate JSON for the MAID inspection panel.

MAID Snapshot Generator

Generate a snapshot JSON that can be imported by MAID for visualization in the InspectionPanel and ContextViewer.

Critical Output Rules

1. Create a file named maid-snapshot-<id>.json in the current working directory, where <id> is a unique identifier (use epoch milliseconds from the command below)
2. Write ONLY valid JSON to the file (no markdown fences, no explanations)
3. Strict JSON: no trailing commas, no NaN/Infinity, no comments
4. Prefer OMISSION over fabrication:
   • Do NOT guess values
   • If a field is unknown, either omit it or set to null only when explicitly allowed
5. After creating the file, confirm with the file path

Timestamps - CRITICAL

BEFORE generating the snapshot, run these commands:

# Get current epoch milliseconds (for exportedAt and file naming)
date +%s000

# Get current ISO 8601 UTC timestamp (for exportedAt)
date -u +"%Y-%m-%dT%H:%M:%S.000Z"

Export Timestamp (exportedAt)

Use the ISO command above for the current time.

Invocation Timestamps (events[].ts)

Priority 1 — Extract from Amp thread data:

Amp stores thread data at ~/.ampcode/threads/. Each thread is a directory named by thread ID (e.g., T-019b6f89-...).

# List recent Amp threads (sorted by modification time)
ls -lt ~/.ampcode/threads/ | head -20

# Get mtime of thread directory in epoch ms (macOS)
stat -f %m ~/.ampcode/threads/<thread-id> | awk '{print $1 * 1000}'

# Check for any JSON files in a thread
ls ~/.ampcode/threads/<thread-id>/

If files were touched during the session, use file mtimes:

# Get mtime of a file in epoch ms (macOS)
stat -f %m "<file>" | awk '{print $1 * 1000}'

# Get mtime of a file in epoch ms (Linux)
stat -c %Y "<file>" | awk '{print $1 * 1000}'

Priority 2 — Estimate from typical durations: If no source timestamps available, estimate based on realistic operation times:

• User message → Assistant thinking: +500-2000ms
• Tool call start: +100-500ms after assistant message
• File read operation: +50-200ms
• File write/edit operation: +100-500ms
• Bash command: +200-5000ms (depends on command)
• Web search/fetch: [phone]ms
• Assistant final response: +500-2000ms after last tool

Rules:

• events[].ts MUST be monotonic non-decreasing
• Use REAL timestamps when available, estimates only as fallback
• When estimating, ask user about approximate session duration if unclear

Required Root Shape

{
  "version": "1.0",
  "exportedAt": "<ISO timestamp UTC>",
  "model": "<model id or 'unknown'>",
  "tokenUsage": {
    "totalTokens": 0
  },
  "tools": [],
  "context": [],
  "events": []
}

IMPORTANT: tokenUsage.totalTokens is REQUIRED and must be a number (use 0 if unknown).

Optional tokenUsage fields (include only if known): promptTokens, modelOutputTokens, modelReasoningTokens, contextLimit, remainingTokens

Model Name - REQUIRED

Extract the model name from these sources (in priority order):

1. System prompt: Look for model identifier in the agent's system prompt (e.g., "You are Claude", "powered by GPT-4")
2. Amp thread URL: The thread URL or thread metadata may reference the model
3. Ask the user: If not found, ask "Which model was used for this session?"

Common model names: claude-sonnet-4-20250514, claude-3-5-sonnet, gpt-4o, gpt-4-turbo, o1, o3

Do NOT use "unknown" without first attempting to extract or ask.

Tools Array - REQUIRED

Extract tools from these sources:

1. Tool calls in the conversation: Look at every tool/function call made during the session. Each unique tool name should be in the tools array.
2. System prompt: The system prompt often lists available tools with descriptions.
3. Amp's standard tools: If this is an Amp session, common tools include:
   • Read - Read files or directories
   • Grep - Search for patterns in files
   • glob - Find files by pattern
   • edit_file - Edit existing files
   • create_file - Create new files
   • Bash - Execute shell commands
   • finder - Semantic code search
   • web_search - Search the web
   • read_web_page - Read web page content
   • Task - Spawn sub-agents
   • oracle - Consult reasoning model
   • mermaid - Render diagrams
   • todo_write / todo_read - Task management

For each tool used in the conversation, add an entry:

{
  "type": "function",
  "function": {
    "name": "<tool name>",
    "description": "<brief description of what it does>",
    "parameters": {
      "type": "object",
      "properties": {},
      "required": []
    }
  }
}

IMPORTANT: Populate tools with at least the tools that appear in tool_calls. Do NOT leave tools: [] if tool calls exist in the conversation.

Context Array (Chat Transcript)

Chronological list of messages. The first message MUST be the system prompt if available.

System prompt (REQUIRED as first message if available):

{ "role": "system", "content": "<full system prompt text>" }

The system prompt is the initial instruction set given to the agent. It typically includes:

• Agent identity and capabilities
• Available tools and their usage guidelines
• Behavioral rules and constraints
• Response formatting instructions

If the system prompt is not explicitly provided: Extract it from any visible agent configuration/logs like CLAUDE.md or AGENTS.md. If no system prompt is found, use the default system prompt provided by the agent.

User/Assistant message:

{ "role": "user" | "assistant", "content": "<string>" }

Assistant with tool calls:

{
  "role": "assistant",
  "content": "<string>",
  "tool_calls": [
    {
      "id": "<string>",
      "type": "function",
      "function": {
        "name": "<tool name>",
        "arguments": "<JSON string>"
      }
    }
  ]
}

Note: Do NOT include index field in tool_calls.

Tool result:

{
  "role": "tool",
  "tool_call_id": "<matches tool call id>",
  "content": "<string>"
}

Rules:

• Keep exact order: assistant(tool_calls) → tool(result) → assistant(final)
• Tool arguments MUST be a JSON STRING, not an object

Events Array (Inspection Timeline)

Each event:

{
  "ts": 1767103680000,
  "data": "<one-line summary>",
  "inspectionEvent": {
    "message": "<can be long>",
    "invocationId": "<uuid for grouping>",
    "children": [
      { "label": "<label>", "data": "<string>" }
    ]
  }
}

IMPORTANT: ts MUST be a NUMBER (epoch milliseconds), NOT a string.

Allowed Labels

• __INVOCATION_START__
• __INVOCATION_END__
• Content
• Reasoning
• Tool Calls
• Timing
• Token Usage
• Error
• Custom

Recommended Event Pattern Per Invocation

1. Start: children: [{ "label": "__INVOCATION_START__", "data": "" }]
2. Tool execution: Include Reasoning (short), Tool Calls (stringified JSON)
3. Tool completed: Include Timing, Tool Calls with results, Token Usage
4. Final message: Include Content, optionally Reasoning
5. End: children: [{ "label": "__INVOCATION_END__", "data": "" }]

Important

• Do NOT include private chain-of-thought
• If including "Reasoning", keep it short and high-level
• Omit unknown data rather than guessing
• Output ONLY the JSON object—no markdown, no extra text