feat(tui): navigate the transcript by prompt, landmark and block - #53333

Note: I'm aware v1 is now the maintenance branch and new features go to v2 only. This PR is here for others who want to use it on v1. I will submit a v2 version once I've migrated my plugin setup to v2.

Issue for this PR

Closes #53331

Type of change

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

What does this PR do?

Adds three levels of transcript navigation, plus ctrl+home / ctrl+end for the existing first/last commands. Every binding is a default that tui.json can override.

Prompt navigation fix. messages_previous / messages_next used to skip any prompt within 10 rows of the viewport top, because they compared prompt positions against a fixed ±10 threshold rather than the row they land on. With no prompt left they scrolled a page. They now use the same picker as the other two levels, so every prompt is a stop, including short back-to-back turns. Past either end the view goes to the top or back to the bottom.

Landmark rule. The final response of a turn is the last non-empty text part before the next user prompt. While a turn is still streaming, its latest text part counts. The landmark tools are:

• question: what you were asked and what you answered;
• todowrite: the plan, as it changed;
• task: where work was handed to a subagent.

Edits, patches, reads, greps and shell calls are not landmarks. A turn is mostly made of them, so stopping on each would make landmark navigation the same as block navigation.

Skipped reasoning. A reasoning part with no text after [REDACTED] is stripped is skipped. That covers encrypted blocks and GPT's empty "Thinking" placeholders, which render as a bare Thought · 1.0s line. It is the same test ReasoningPart uses to decide what it shows as content.

How it finds blocks. User messages were already rendered with id={message.id}. This PR gives text, reasoning, inline-tool and block-tool boxes id={part.id}, so the route can match a part to its scrollbox child. Selection lives in a small pure module, src/routes/session/navigation.ts:

• navigationTargets(messages, parts, scope) returns the set of IDs to stop on for prompt, landmark or block.
• pickNavigationTarget(blocks, anchor, direction) picks the nearest one above or below the anchor row. A part rendered as several boxes, such as apply_patch with one box per file, counts once, at its first box.

The anchor is one row below the viewport top, which is where messages_next already landed a prompt. A block that was just navigated to counts as current, so repeated presses walk one block at a time.

The keymap.test.tsx change only updates the count of default session.first bindings from 2 to 3.

How did you verify your code works?

• test/routes/session/navigation.test.ts covers prompt, block and landmark selection:

  • synthetic and ignored prompts, empty and [REDACTED] reasoning, and whitespace-only text are excluded;
  • the final response of each turn, including the last turn while it streams;
  • nearest-above/below picking, multi-box parts, and both ends.

  bun test and bun typecheck pass in packages/tui.

• I imported a 28-message session into isolated XDG_* dirs. It has 6 full turns, each with an empty GPT-style reasoning placeholder, reasoning with content, text, edit, bash, todowrite, alternating question/task, and a multi-paragraph final response, plus two one-line turns whose prompts are fewer than 10 rows apart. I drove it in a throwaway tmux socket with extended-keys on, sending the keys with tmux send-keys.

Screenshots / recordings

The first viewport row after each key, from tmux capture-pane, starting at the bottom of the session.

ctrl+up, before (messages_previous bound to ctrl+up in tui.json) and after:

before                               after
PROMPT 6: please do step 6           PROMPT 6: please do step 6
PROMPT 5: please do step 5           PROMPT 5: please do step 5
PROMPT 4: please do step 4           PROMPT 4: please do step 4
PROMPT SHORT A   <- skips SHORT B    PROMPT SHORT B
PROMPT 3: please do step 3           PROMPT SHORT A
                                     PROMPT 3: please do step 3

ctrl+shift+up and then ctrl+alt+shift+up:

C-S-Up   | PROMPT 6: please do step 6
C-S-Up   | FINAL 5: step 5 is done.
C-S-Up   | ✓ Explore Task — TASK 5
C-S-Up   | # Todos

C-M-S-Up | $ echo step 5
C-M-S-Up | ← Edit /src/step5.ts
C-M-S-Up | WORKING 5: starting on step 5.
C-M-S-Up | + Thought: 1.0s              <- reasoning with content; the empty placeholder before it is skipped
C-M-S-Up | PROMPT 5: please do step 5
C-M-S-Up | FINAL 4: step 4 is done.

On dev (907b3bc518), ctrl+up, ctrl+shift+up, ctrl+alt+shift+up, ctrl+home and ctrl+end don't change the viewport. home / end work as before. A draft in the prompt is left untouched; the prompt doesn't bind any of these keys.

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.