echospace-export

name	echospace-export
description	Convert conversation files (OpenAI, Anthropic, Google, Vercel/Helicone, raw text) into .echo format. Triggers: echo export, convert to echo, json to echo, helicone to echo, openai to echo, anthropic to echo, convert conversation

Echo Export Skill

Convert any conversation file into .echo NDJSON format for use with EchoSpace.

Usage

/echo-export <file-or-glob>

Examples:

• /echo-export helicone.json
• /echo-export exports/*.json
• /echo-export . (convert all supported files in current directory)

Supported Input Formats

Format	Detection
OpenAI	messages[] with role + content, or { messages: [...] }
Anthropic	content[] with type blocks, or { system, messages }
Google Gemini	{ contents: [{ parts }] }
Vercel AI / Helicone	{ request: { prompt: [...] }, response: { ... } }
Raw text	Lines starting with User: / Assistant: / System:
Echo (NDJSON)	Already .echo — skip or rewrite

Procedure

1. Resolve input files. If the argument is a glob or directory, expand it. Filter to .json, .jsonl, .txt, and .echo files.

2. For each file: a. Read the file content. b. Detect the format using these rules (in order):

   • Try JSON.parse(). If it succeeds:
     • Has request.prompt array → Vercel
     • Has request.messages array → OpenAI (Helicone wrapper)
     • Is array with role + content objects → OpenAI
     • Has messages array → OpenAI
     • Has content array with type: "text" blocks → Anthropic
     • Has contents array with parts → Google
   • If JSON parse fails, try NDJSON: parse first line, check for kind: "meta" or kind: "message" → Echo
   • Check for User: / Assistant: / System: line patterns → Raw text
   • Otherwise → Unknown (skip with warning)

   c. Parse the messages according to the detected format:

   OpenAI:

   • Unwrap Helicone { request, response } wrapper if present
   • Map role directly (system, user, assistant, tool)
   • String content → single text part; array content → map text and image_url blocks
   • reasoning_content string → prepend as thinking part
   • tool_calls[] → tool_call parts (function.name, function.arguments)
   • role: "tool" with tool_call_id → tool_result part
   • Append response choices/messages as additional messages

   Anthropic:

   • Extract top-level system field as first role: "system" message
   • Map content blocks: text → text, thinking → thinking, tool_use → tool_call, tool_result → tool_result, image (base64) → image

   Google Gemini:

   • systemInstruction.parts → system message
   • role: "model" → assistant
   • functionCall → tool_call, functionResponse → tool_result

   Vercel AI SDK:

   • Part type mapping: tool-call → tool_call (toolCallId → id, toolName → name), tool-result → tool_result (toolCallId → tool_call_id, unwrap { type: "json", value } wrappers), reasoning → thinking
   • Append response choices as final assistant message with model/usage metadata

   Raw text:

   • Split on User: / Human: / Assistant: / System: line prefixes
   • Each section becomes a message with a single text part

   d. Build the echo conversation:

   meta: { kind: "meta", v: 1, id: <nanoid>, title: <filename without extension>, created_at: <now> }
   messages: <parsed messages, each with kind: "message">
   

   e. Serialize to NDJSON: one JSON.stringify() per record, joined by \n, trailing newline.

   f. Write to <original-name>.echo in the same directory (e.g., helicone.json → helicone.echo). If the output file already exists, warn the user before overwriting.

3. Report results. Print a summary: how many files converted, any skipped files, output paths.

Output Format

The .echo file is NDJSON. Line 1 is always the meta record. Subsequent lines are message records.

{"kind":"meta","v":1,"id":"abc123","title":"My Chat","created_at":"2025-06-01T12:00:00Z"}
{"kind":"message","id":"msg_1","role":"user","created_at":"2025-06-01T12:00:01Z","parts":[{"type":"text","text":"Hello"}]}
{"kind":"message","id":"msg_2","role":"assistant","created_at":"2025-06-01T12:00:02Z","parts":[{"type":"text","text":"Hi!"}]}

Important Notes

• Generate unique IDs with nanoid(8) for any missing id fields
• Always use ISO 8601 timestamps for created_at
• The parts field is always an array, even for single text content
• Do NOT pretty-print JSON lines — each line must be a single compact JSON object
• File ends with a trailing newline
• When the input is already .echo format, skip it (no need to re-convert)