[Nowaker]

Note: I'm aware v1 is now the maintenance branch and new features go to v2 only. This PR is here for others to use, and I'll submit the v2 version once I've migrated my whole plugin system to it.

Issue for this PR

Closes #56279

The v1 version of #53660 (closed; it targeted dev). Two commits: the first is #53660's change rebased onto v1, the second the three-row bar and the costed load all.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

The TUI keeps a session's newest 100 messages loaded and silently drops the oldest one for every new message past that. The limit counts messages only, and nothing else limits what the transcript renders. This PR makes the hidden part visible and loadable, and the limit configurable.

Bar. Wherever messages are hidden, the transcript shows three rows:

                 load 50 above   load 100 above
──── 3,250 messages hidden:  load all (slow / memory heavy) ────
                 load 50 below   load 100 below

• "Above" fills the gap from its older edge (right after the pinned first prompt, or the session's first messages when none is pinned); "below" from its newer edge, extending the recent messages backwards.
• A page is offered only while it is smaller than the gap: 51-100 hidden shows only the 50 pages, 50 or fewer shows only the rule.
• When the count and load all don't fit on the rule, the rule shows N hidden and load all gets its own row under it.
• While a load runs its label reads loading… and the other actions ignore clicks.
• Loading grows the window by what it loaded, so new output doesn't drop it again straight away; load all stops dropping messages for that session. A scrolled-up reader keeps their place.
• Every action is clickable and a palette command ("Load 50/100 hidden messages above/below the divider", "Load all hidden messages", /load-hidden), with keybinds messages_hidden_above, messages_hidden_above_100, messages_hidden_below, messages_hidden_below_100, messages_hidden_all, unbound by default.

load all label. Computed per gap from a measured 10 ms and 0.5 MB per hidden message: load all below 300 hidden, load all (slow) at 300-499 (an estimated freeze of 3 s or more), load all (slow / memory heavy) from 500 (an estimated 250 MB or more, about a third of an idle TUI). The palette title carries the same note.

Settings (tui.json):

• transcript.max_messages - newest messages kept loaded, default 100 (the old constant).
• transcript.keep_first_prompt - keep the session's first prompt above the bar when older messages are hidden, default true. This changes what a long session looks like today; happy to flip the default to false.

API. GET /session/:id/message only paged backwards, so it gains after (a cursor or message ID; newer messages, oldest first), order=asc (page from the oldest message), message IDs for before as well as cursors, and X-Total-Count. Conflicting combinations are 400; unpaged requests are unchanged. The TUI anchors on message IDs it already has, because a cursor goes stale once its message is dropped, and asks for the first prompt only when the newest page doesn't cover the session.

Memory and time of load all

Ten real sessions of 2,900-13,300 messages, copied into a scratch database, opened in an isolated TUI (160x50, scratch XDG_*), one run at a time on a busy 18-core Linux desktop. RSS is the TUI process once idle. "Half" is transcript.max_messages: 50 against the default 100. "Transcript shown" runs from pressing load all until the bar leaves the screen; "back to idle" until a 2 s window uses under 15% of a core.

Messages	Parts	DB bytes	RSS half (50)	RSS default (100)	RSS after load all	Load all: transcript shown	Load all: back to idle	CPU-s to idle
13,261	45,052	48 MB	665 MB	828 MB	5,162 MB	164 s	248 s	120
7,152	33,798	107 MB	711 MB	747 MB	3,902 MB	80 s	255 s	115
4,310	20,262	126 MB	882 MB	867 MB	3,215 MB	61 s	433 s	178
6,417	24,231	67 MB	785 MB	831 MB	3,039 MB	58 s	62 s	28
4,227	22,585	63 MB	774 MB	706 MB	3,404 MB	47 s	136 s	53
3,438	13,396	22 MB	777 MB	856 MB	2,182 MB	29 s	32 s	16
4,777	20,603	73 MB	780 MB	722 MB	3,109 MB	44 s	51 s	27
4,155	20,252	58 MB	782 MB	864 MB	2,977 MB	32 s	not idle in 900 s	390
2,912	12,608	47 MB	796 MB	729 MB	2,826 MB	19 s	25 s	19
3,224	16,141	54 MB	778 MB	714 MB	2,395 MB	15 s	712 s	244

• Half the window against the default is within run-to-run noise (mean 773 vs 786 MB): an open TUI is 700-900 MB whatever the window holds.
• Load all grows RSS by 1.3-4.3 GB (median 0.52 MB per hidden message) and takes 15-164 s before the transcript shows (median 9.3 ms per message), which is where the label's constants come from.
• In 6 of 10 sessions the TUI stays busy for minutes after load all with nothing changing on screen. That is outside this PR.

How did you verify your code works?

• packages/tui: bun typecheck; bun test. test/routes/session/hidden-divider.test.tsx renders the bar at 40, 80 and 160 columns, the 51-100 and single-message gaps, the label thresholds, clicks and the loading state. test/cli/cmd/tui/sync-hidden-messages.test.tsx covers the pinned first prompt, keep_first_prompt: false, an uncropped session making one request, 50- and 100-message pages from both edges, live messages hiding the oldest recent message but never the pinned prompt, and load all.
• packages/opencode: test/session/messages-pagination.test.ts and test/server/session-messages.test.ts (forward pages, message-ID anchors, X-Total-Count, Link, the 400s).
• My fork's build in a real TUI in tmux with isolated XDG dirs, on a copy of a 2,912-message session: the three rows above with 2,811 messages hidden: load all (slow / memory heavy); clicking load 100 below -> 2,711 messages hidden.

Screenshots / recordings

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

This PR is AI generated, but under direct supervision and on request of @Nowaker.

[github-actions]

The following comment was made by an LLM, it may be inaccurate: