tui: scrolling up through a reply jumps the view into the thinking block above it #58666

Activity

1. opencode-agent commented on Oct 24, 2026

   

   opencode-agentBot

   on Oct 24, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the clear report! I could reproduce this on 2.1.4, and it still happens on the latest dev (f20760b8).

   Steps: set session.thinking to "show", create a session with 20 turns where each reply has a thinking block before it, then scroll up from the bottom one wheel step at a time. For the newest ~10 turns, each step moves ~3 rows. Further back, a single step goes from the first lines of a reply straight to the start of the thinking block above it, skipping the rest of the block, and the next step shows the "Thought" header. The recording below shows the full run, and the script below reproduces it.

   The jumps start where older messages begin to be loaded into the view as you scroll up, so the code that adds those rows above the current view is the likely cause. #54158 touches the same code and may be related.

   recording.mp4

   Reproduction script (drive.ts)

   // Repro for anomalyco/opencode#58666: scrolling up one wheel step at a time
   // jumps over the thinking block above a reply.
   import { Effect } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   const TURNS = 20
   const thought = (n: number) =>
     Array.from({ length: 6 }, (_, k) => `Turn ${n} thought line ${k + 1}: weighing the options carefully.`).join("\n\n")
   const reply = (n: number) =>
     Array.from({ length: 10 }, (_, k) => `Turn ${n} reply line ${k + 1}.`).join("\n\n")
   
   // Text of the transcript area's rows, to see which line is at the top after each step.
   const frameText = (frame: any): string[] =>
     frame.lines.map((l: any) => l.spans.map((s: any) => s.text).join(""))
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tuiConfig: { session: { thinking: "show" } },
       tui: { recording: true, viewport: { cols: 120, rows: 40 } },
     } as any,
     ({ ui, llm }) =>
       Effect.gen(function* () {
         for (let n = 1; n <= TURNS; n++) {
           yield* ui.submit(`Question ${n}`)
           yield* llm.send(Llm.reasoning(thought(n)), Llm.text(reply(n)))
           yield* Effect.sleep(700)
         }
         yield* ui.screenshot("bottom")
         yield* Effect.sleep(1000)
         let prevTop = ""
         for (let step = 1; step <= 330; step++) {
           yield* ui.mouse({ action: "scroll", x: 40, y: 15, direction: "up" })
           yield* Effect.sleep(120)
           const lines = frameText(yield* ui.capture())
           const top = lines.slice(1, 4).map((l) => l.trim()).filter(Boolean).join(" | ")
           yield* Effect.log(`step ${step}: ${top.slice(0, 200)}`)
           prevTop = top
         }
         void prevTop
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
2. opencode-agent

   added

   reproduced

   on Oct 24, 2026
3. Anthony14FR commented on Nov 12, 2026

   

   Anthony14FR

   on Nov 12, 2026

   More actions

   I'd like to work on this. Following the triage, the jump starts where older rows get loaded above the viewport while scrolling up, so I'll look at how the scroll offset is kept when those rows are prepended (and check #54158, which touches the same code). I'll add a regression test that scrolls up one step at a time through replies with thinking blocks.
4. Anthony14FR commented on Nov 12, 2026

   

   Anthony14FR

   on Nov 12, 2026

   More actions

   Update: I can't reproduce this on current dev. I wrote an OpenTUI renderer test (20 and 30 turns, each with a reasoning part and a text part, session.thinking: "show", 100x30 and 120x40) that scrolls up one wheel step at a time and checks each step moves the view by exactly 3 rows. Every step does, including across the point where older rows get mounted.

   The triage repro was on f20760b8. Since then #59236 ("stop older history from flashing when scrolling up") changed pinAbove, the older-row reveal and rows.ts, which is likely what fixed it. The default verbosity is also low now (#63207), which folds thoughts into a collapsed group.

   My test loads the history in one fetch rather than streaming it live, so it doesn't cover everything the original script does. @ganyicz could you check whether it still happens on the latest release? If it does, I'll keep digging; if not, this can probably be closed.