feat(chat): export + rehydrate conversation history for lossless resume

[localai-bot]

What & why

Adds two capabilities to chat.Session so a downstream app can persist a conversation to disk and later resume it losslessly — meaning the model actually remembers the prior conversation, not a summary. The motivating consumer is Dante Desktop, which keeps multiple persisted chats the user can close and reopen.

Today this is impossible from outside the package: GetMessages() strips messages down to {Role, Content}, and NewSession always starts from an empty fragment with no way to seed prior history.

The two additions

1. Full-fidelity export

func (s *Session) ExportHistory() []openai.ChatCompletionMessage

Returns a copy of the full conversation messages (JSON-serializable, suitable for persistence). Mutating the result never touches the live session. It excludes the system prompt — s.messages only ever records user/assistant turns; the system prompt lives on the fragment and is regenerated per model/locale each turn. Safe to call from another goroutine while a turn runs: it takes the same lock SendMessage holds while appending.

2. Seed-on-construct (rehydration)

// types.Config
InitialHistory []openai.ChatCompletionMessage

NewSession seeds it into both s.messages (the UI log) and s.fragment (the real model context), so the very next SendMessage continues with full memory — identical to a session that organically reached that state. Round-trips with ExportHistory.

System-prompt correctness

SendMessage prepends the current system prompt at the start of every turn (s.fragment.AddMessage("system", …)), and cogito consolidates all system messages to the front of the request. So the seed is loaded into the fragment WITHOUT a system message, letting that per-turn add supply exactly one. This:

• mirrors a fresh multi-turn session (whose fragment likewise carries no leading system message before the turn's own add), and
• guarantees the first resumed request carries the system prompt exactly once — no duplication, no omission. Baking a system message into the seed would have doubled it.

Config.InitialHistory is documented to exclude the system prompt (and ExportHistory never emits one), keeping the contract airtight. A new historyMu guards the parallel fragment/messages state so ExportHistory can snapshot consistently mid-turn.

Tests

New chat/session_history_test.go (drives a real Session against a fake OpenAI endpoint, race-clean):

• export is a full copy — returns the whole history; mutating/appending to the result leaves the session untouched.
• seeded history reaches the model — the first resumed request carries the prior turns in order, followed by the new message.
• system prompt exactly once — with a prompt configured and history seeded, the first resumed request has exactly one system message, base prompt un-duplicated, at the front, with prior turns following.

go build ./..., go vet ./..., and go test ./chat/... ./types/... all pass; history tests verified under -race.

🤖 Generated with Claude Code