fix/fill input

[phanen]

• feat(session): refill prompt input after /undo
• refactor(history): derive prompt history from active session messages

Not sure why all prompt history is stored in ~/.local/share/nvim/opencode/history.txt.

Ideally different session should have separate prompt history.

Maybe history can be deprecated in favor of timeline picker?

[sudo-tee]

@phanen

The history file was a remnant from the original repo this one was forked from (goose.nvim)

But just scope it from the session would limit what you can do with it.

Sometimes I start new sessions but want to send a prompt I already sent in another.

But they should be scoped per project at least.

[phanen]

But they should be scoped per project at least.

This pr currently make history picker per-session since it's just wrapper on timeline picker. Maybe we can have a action for history to toggle session/project/...

Also maybe each session should have persistent draft message if tui support this (i don't often use tui so i'm not sure)

[sudo-tee]

@phanen

Can we split the 2 things that this PR addresses ?

I really like the /undo fix.

But for the history I would need to play with it and think about what is the best solution.

Because opencode tui does the same history navigation as we do with arrows in the input. and it does not scope it per session either. The history picker in this plugin was just a view on the message history the user sent over time.

[phanen]

Found another bug... /undo say "No user message to undo" but it actually have.

msg is often truncated from the end

opencode.nvim/lua/opencode/ui/renderer.lua

Lines 368 to 370 in eebda20

	if lazy_limit and #visible_messages > lazy_limit then
	visible_messages = vim.list_slice(visible_messages, #visible_messages - lazy_limit + 1)
	end

opencode.nvim/lua/opencode/ui/renderer.lua

Lines 87 to 101 in eebda20

	local function get_visible_session_messages(messages)
	local real_messages = get_real_session_messages(messages)
	local revert_index = get_revert_index(messages)
	if revert_index then
	real_messages = vim.list_slice(real_messages, 1, revert_index - 1)
	end
	local limit = get_max_rendered_messages()
	if not limit or #real_messages <= limit then
	return real_messages, 0
	end
	local start_index = #real_messages - limit + 1
	return vim.list_slice(real_messages, start_index, #real_messages), start_index - 1
	end

set_last_user_message is only called in on_message_updated (only on truncated messages)

opencode.nvim/lua/opencode/ui/renderer.lua

Lines 375 to 377 in eebda20

	if hidden_count > 0 then
	local hidden_notice = build_hidden_messages_notice(hidden_count)
	events.on_message_updated(hidden_notice)

and only when role is user

opencode.nvim/lua/opencode/ui/renderer/events.lua

Lines 291 to 292 in eebda20

	if msg.info.role == 'user' then
	state.renderer.set_last_user_message(msg)

but a canceled user messages sequence can be:

user -> assistant (tool call 1) -> assistant (tool call 2) -> ...

so when there's many tool call, truncated msg won't contain user msg -> last_user_message will be never set.

so sometimes /undo cannot found last_user_message (usually reproduce on startup or when use max_messages, and haven't load_all_message/load_more_message).

[phanen]

on_message_updated seems correct to call set_last_user_message when it's call from opencode event. But it have have side effect on render job... so _render_full_session_data lazily use it... but set_last_user_message is check/updated on every message.

Not sure if it's proper to add another set_last_user_message here... or make anther version of on_message_updated without side effect.. i'll think about it later.

EDIT: llm suggest a another option to remove set_last_user_message totally and get last userr msg from store.messages when require..

[phanen]

EDIT: llm suggest a another option to remove set_last_user_message totally and get last userr msg from store.messages when require..

I finally choose this direction since it feels proper and easy to go...

[sudo-tee]

Since we can restore running sessions, and the output is sliced, it might be a good isea to reconstruct from the messages instead set_last_user_messsage. it's worth a try at least

[sudo-tee]

Looks good to me.

Thanks for the PR