Windsurf Trajectory Extractor

English | 简体中文

Deep extraction of Windsurf Cascade conversation history via protobuf decoding.

What This Does

Extracts complete trajectory data from Windsurf's internal storage, including:

• ✅ Thinking content — Internal reasoning (only in thinking mode, not visible in UI)
• ✅ Visible responses — User-facing text
• ✅ Tool calls — Complete parameters and results
• ✅ Microsecond timestamps — High-precision timing data
• ✅ Provider info — Model provider (e.g., "anthropic")

Why This Tool?

Windsurf stores conversation history in a protobuf-encoded format inside state.vscdb. This tool reverse-engineers that format to extract data that's not accessible through the standard UI export.

Key differentiator: Unlike JSON-based extraction tools, this performs deep protobuf decoding to access thinking content and precise timestamps.

Installation

# No dependencies required — pure Python standard library
git clone https://github.com/jijiamoer/windsurf-trajectory-extractor.git
cd windsurf-trajectory-extractor

# Using uv (recommended)
uv sync
uv run windsurf-trajectory --help


# Or using pip
pip install -e .
windsurf-trajectory --help

Usage

# List all workspaces with trajectory data
windsurf-trajectory --list

# Search trajectories by keyword (auto-discover workspace)
windsurf-trajectory --find cascade_solver

# List conversation summaries for a workspace
windsurf-trajectory --summaries WORKSPACE_ID

# Extract trajectory to JSONL
windsurf-trajectory -w WORKSPACE_ID -o trajectory.jsonl

Output Format (JSONL)

Each line is a JSON object representing one step:

{
  "step_id": 15,
  "step_type": 3,
  "timestamp": "2026-03-12T14:30:45.123456+08:00",
  "timestamp_unix_ms": 1741761045123,
  "thinking": "Let me analyze this code...",
  "visible": "Here's my analysis of the code...",
  "tool_calls": [
    {
      "tool_id": "abc123",
      "tool_name": "read_file",
      "params": {"file_path": "/path/to/file.py"}
    }
  ],
  "provider": "anthropic",
  "content_preview": "Here's my analysis..."
}

Field Reference

Field	Description
step_id	Step opcode: 14=user message, 15=AI response, 21=tool execution
step_type	Internal step type (usually 3)
timestamp	ISO 8601 timestamp with microsecond precision
timestamp_unix_ms	Unix timestamp in milliseconds
thinking	Internal reasoning (only in thinking mode)
visible	User-visible response text
tool_calls	Array of tool invocations with parameters
provider	Model provider (e.g., "anthropic")

Technical Details

Data Location

~/Library/Application Support/Windsurf - Next/User/globalStorage/state.vscdb

Supports both Windsurf (stable) and Windsurf - Next (preview) on macOS, Linux, and Windows.

Protobuf Structure (Reverse-Engineered)

Top-level:
  f1 (string): Trajectory UUID
  f2 (message): Steps container
    repeated Step:
      f1  (varint): step_id
      f4  (varint): step_type
      f5  (message): metadata {f1: Timestamp {f1=seconds, f2=nanos}}
      f19 (message): user message
      f20 (message): AI response
        f3  (string): thinking content ← only in thinking mode
        f7  (message): tool call {f1=tool_id, f2=tool_name, f3=params_json}
        f8  (string): visible response
        f12 (string): provider
      f28 (message): tool result

Limitations

• Only extracts the active trajectory (most recently selected conversation) per workspace
• To extract a specific conversation, first select it in Windsurf's sidebar
• Protobuf structure may change with Windsurf updates

Use Cases

• Import to Nowledge Mem — Extract conversations for personal knowledge management
• Research — Analyze AI coding assistant behavior and timing
• Backup — Export conversation history not available through UI
• Debugging — Access thinking content for troubleshooting

License

MIT

Related Projects

• nowledge-mem — Personal memory system for AI agents
• antigravity-trajectory-extractor — Live RPC extraction for Antigravity
• 0xSero/ai-data-extraction — Multi-tool extraction (JSON-based)

---

Built for the Nowledge Community