v0.2.0 — ChatGPT Adapter MVP

ChatGPT export adapter (F2)

The first production adapter. You can now feed a real ChatGPT
conversations.json export into the library and get back a clean linear
thread of NormalizedMessages.

What works

• ✅ Detects ChatGPT exports by sniffing "mapping" or "conversations" in the file header
• ✅ Reconstructs the active linear thread from the non-linear mapping tree
  (walks current_node → parent chain → root, then yields root-first)
• ✅ Role mapping: user / assistant / system / tool. Unknown roles
  (e.g. custom GPT name) are dropped.
• ✅ Skips whitespace-only and empty messages
• ✅ Extracts text parts only — drops structured payloads (image_url,
  code interpreter, tool calls)
• ✅ FNV-1a 64-bit content hash for downstream dedup
• ✅ Slugifies conversation title into a project_hint

Library API

use loust_llm_mempipe::adapter::chatgpt::ChatGptAdapter;
use loust_llm_mempipe::adapter::Adapter;

let adapter = ChatGptAdapter;
let reader = Box::new(std::fs::File::open("conversations.json")?);
let messages: Vec<_> = adapter.stream_messages(reader)?.collect();

Validation

• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• cargo test — 16/16 pass (7 new ChatGPT tests + 9 library tests)
• cargo build --release — OK (~10s clean build)

Limitations (F2 MVP)

• The full export is materialized in memory (~2× JSON size). For 50 MB
  exports that's ~100 MB peak — fine for a developer laptop. True
  streaming optimization is tracked for F3+ if real exports need it.
• CLI surface still placeholder (F4). Library is usable from Rust code.

Next

• F3: pipeline core (Rule E secret scrubber → Jaccard dedup → signal scoring
  → JSONL + Markdown writer)
• F4: clap CLI with --input, --output, --format, --stats
• F5: CI + smoke E2E