add adapters for more local AI coding logs

[ohad6k]

Ditto currently reads Codex and Claude Code JSONL logs.

Good next adapters:

• Cursor
• Cline
• Continue
• Windsurf

Best contribution:

1. document where the tool stores local sessions;
2. add a tiny parser path in ditto.py;
3. include a redacted/minimal fixture shape if possible;
4. never commit private logs.

If you know one of these formats, comment with the local path and sample schema first.

[balaredde]

Hi, I'd like to pick this up as a good first issue. I will start by documenting where Cursor and Windsurf store their local sessions

[ohad6k]

It's yours, and starting with documenting where the sessions live is exactly the right first step. Most useful shape: for each tool, the storage path per OS, plus a small sanitized sample of the jsonl format (strip anything real, never post actual logs).

One heads up: there's a ditto.py rewrite in flight (bounded mining, stable segments), so hold off on wiring the adapter code until it lands, the docs work is valuable either way and won't get invalidated. If you want to sync on it, the Discord is here: https://discord.gg/VNnMq2U5r

[BullX92]

Hermes Agent stores local conversation sessions as JSON files rather than JSONL.

In our Linux installation the files are stored at /opt/data/sessions/session_*.json. Each session has top-level metadata such as session_id, session_start, last_updated, platform, model, messages, system_prompt, and tools.

messages is an array of objects with {role, content}, where roles include user, assistant, and tool. A Ditto adapter should extract only user-authored messages and exclude system_prompt, tool messages, and injected context.

Sanitized minimal shape:

{
  "session_id": "<redacted>",
  "session_start": "<timestamp>",
  "messages": [
    {"role": "user", "content": "<redacted>"},
    {"role": "assistant", "content": "<redacted>"},
    {"role": "tool", "content": "<redacted>"}
  ]
}

No real session contents should be committed. A small parser path or a temporary JSON-to-JSONL adapter would be sufficient.

[BullX92]

Hermes Agent stores local conversation sessions as JSON files rather than JSONL.

In our Linux installation the files are stored at /opt/data/sessions/session_*.json. Each session has top-level metadata such as session_id, session_start, last_updated, platform, model, messages, system_prompt, and tools.

messages is an array of objects with {role, content}, where roles include user, assistant, and tool. A Ditto adapter should extract only user-authored messages and exclude system_prompt, tool messages, and injected context.

Sanitized minimal shape:

{
  "session_id": "<redacted>",
  "session_start": "<timestamp>",
  "messages": [
    {"role": "user", "content": "<redacted>"},
    {"role": "assistant", "content": "<redacted>"},
    {"role": "tool", "content": "<redacted>"}
  ]
}

No real session contents should be committed. A small parser path or a temporary JSON-to-JSONL adapter would be sufficient.

[ohad6k]

This is great, thanks for digging into Hermes. It's a strong target for ditto, provider-agnostic with persistent memory is exactly the kind of tool the profile should read from.

One thing before anyone wires an adapter: can you confirm the path on a standard (non-container) install? What I've seen points to Hermes keeping sessions in a sqlite/fts5 store, and /opt/data/... looks like a Docker layout, so I want to make sure we point at where a normal user's sessions actually land. If you can drop the default path per OS, and whether it's loose json files vs the sqlite db, that's basically the adapter spec right there.

Heads up that adapter code is on hold until the ditto.py rewrite lands (bounded mining, stable segments), so no rush, but nailing the Hermes format now is real progress and won't get invalidated. If you want to own the Hermes adapter once the rewrite ships, it's yours.

Discord if you want to sync: https://discord.gg/VNnMq2U5r

[JaSchW-jpg]

We established that Hermes Agent sessions are stored primarily in a SQLite database called state.db, not as loose JSON files.

Default paths discussed:

Linux/macOS/WSL: ~/.hermes/state.db

Windows native: %LOCALAPPDATA%\hermes\state.db

Root install: /root/.hermes/state.db

Docker: typically /opt/data/state.db

We also noted that:

Named profiles may use <HERMES_HOME>/profiles//state.db

HERMES_HOME overrides the default base directory

JSON session snapshots are optional/legacy, not the primary storage format

An adapter should preferably read SQLite directly in read-only mode and account for WAL files rather than assuming JSON files exist

[ohad6k]

This nails it, thanks. So the adapter spec is: open state.db read-only (accounting for WAL), resolve the base dir as HERMES_HOME override -> ~/.hermes / %LOCALAPPDATA%\hermes, and check profiles/<profile>/state.db for named profiles. Treat the JSON snapshots as legacy, not the source of truth.

That plus BullX92's message shape (user messages only, skip system_prompt and tool output) is enough to build against. Code is still on hold until the ditto.py rewrite lands, but this thread now has everything an adapter needs. If either of you wants to own the Hermes adapter when it opens up, say the word. Discord: https://discord.gg/VNnMq2U5r