Sessions: keep, fork and join

A harness that runs several branches from one conversation can ask the server to keep that conversation's state under a name and start new requests from it. Set HALOGEN_FORK=1 to turn it on. It is off by default.

# keep the state where this reply ends, under the name plan-7f3a
curl -s localhost:8731/v1/chat/completions -H 'X-Halogen-Keep: plan-7f3a' -d '{"messages": [...]}'
# three workers start from it at once, each sending the history, the reply and its own message
curl -s localhost:8731/v1/chat/completions -H 'X-Halogen-From: plan-7f3a' -d '{"messages": [..., "do part A"]}'
# join: continue from one name and keep the result under another
curl -s localhost:8731/v1/chat/completions -H 'X-Halogen-From: plan-7f3a' -H 'X-Halogen-Keep: plan-7f3b' -d '{...}'

Every request is an ordinary request, and a server without the setting answers it the same way. The response header X-Halogen-From says used, unmatched (the history does not repeat the kept one) or unknown, and the last two are answered as usual. To branch from inside a reply, send part of it as the last assistant message (continue_final_message on Chat Completions, a final assistant message on Messages and Responses). The same headers work on /v1/messages and /v1/responses, for harnesses that can set request headers. Names are released after 10 minutes unused, at most 8 live at once (HALOGEN_FORK_TTL_S, HALOGEN_FORK_HANDLES), and GET /halogen/handles lists them. With HALOGEN_CACHE_DIR set they survive a restart. Every client of the server shares one set of names. Keep the port private or set HALOGEN_API_KEY, and use names other clients cannot guess.

Claude Code. Claude Code cannot send these headers, but it names its own threads on every request. With HALOGEN_FORK=1 and HALOGEN_FORK_MAP=claude-code, the server keeps each thread's state (the main conversation and each subagent) under a name of its own and starts that thread's next request from it, with nothing to set in Claude Code. A subagent's first request starts from its parent's state when it repeats it. Each name is made from the session, the thread and the API key, so it cannot be guessed and two keys never share one. A side request on a session (a title, a summary) never replaces the conversation's state. Off by default.