TUI cockpit shows stale low-signal state and clipped recent-tool failures

[Hmbown]

Problem

The current TUI cockpit/sidebar can surface stale or low-signal state in ways that make the app feel busy without being useful. A recent screenshot showed the right-side Tasks panel still highlighting a failed gh issue create ... command even though the main flow had moved on, while other panels were clipped/truncated enough that the visible text did not explain what mattered.

This is not just visual polish. It weakens trust in the state surface: if the UI foregrounds an old failed command without explaining whether it is still actionable, users have to mentally diff the interface against the actual task.

Observed UX issues

• Recent tools can keep a failed CLI command visible after it is no longer the most useful status.
• The failure text is truncated (Command failed (exit cod...) and does not explain whether the user needs to act.
• Sidebar panels use heavy boxes and repeated separators, which compete with the transcript and composer.
• Work items and agent rows are clipped (in_prog..., st...) so the UI shows activity without enough meaning.
• The composer takes a large fixed area even when empty, reducing space for the main transcript.
• The bottom status line is dense and partly redundant with the sidebar (agents, step, elapsed time, model, live state).
• The main transcript can show verbose patch/diff output while the side panels simultaneously show stale task status, making it hard to know the current center of attention.

Desired direction

Treat the right-side panels as a truth surface, not a log sampler.

• Show current actionable state first: active wait, blocked reason, running command, pending approval, failed action that still needs attention, or clear idle/done.
• Demote old failures once acknowledged, superseded, or no longer relevant.
• If a failure remains visible, show the actionability: retry, inspect, ignore, linked detail, or already handled.
• Prefer detail handles/pagers for verbose command output instead of clipping important meaning.
• Keep typography and borders quiet enough that the active task, composer, and transcript have a clear hierarchy.

Acceptance criteria

• A failed tool/CLI item does not remain foregrounded as the active status after the task has moved on unless it is still actionable.
• Recent tool failures have a concise reason and a way to inspect details.
• Sidebar rows avoid meaningless truncation for status-critical text.
• The UI has one obvious center of attention across transcript, sidebar, composer, and bottom status line.
• Regression coverage or a screenshot/manual smoke path verifies stale failed-command state is cleared, demoted, or marked handled.

Related roadmap

• v0.8.43 truth surface: active state, last event, stall/failure reason, evidence/provenance.
• v0.8.48 presence pass: visual hierarchy, first-run/main-loop polish, task finale, and noise budget.

[Hmbown]

Additional pain point from the v0.8.41/v0.8.42 work loop: subagent and command output truncation is currently hiding exactly the evidence the parent needs to continue safely.

Specific rough edges seen in the TUI:

• delegate done can collapse the child result into N lines omitted; Alt+V for details, so after minutes of delegated work the parent still lacks the classification table, findings, or handoff details.
• Recent tool cards over-index on raw CLI failure text and clipped command output. The visible state says a command failed, but not what action is now useful.
• Filtered test commands can show misleading summaries like 0 passed; 0 failed from unrelated test binaries while the useful matching test lines are omitted.
• Full reasoning in Ctrl+O appears repeatedly, but there is no equally obvious affordance for “open the full subagent transcript / tool output”.
• Diff and write-file surfaces are syntax-heavy and hard to scan when the important question is whether the change is complete and verified.

The code-side mitigation should be that terminal subagent results expose a direct full-transcript transcript_handle, with the compact result treated only as a summary. The UX pass should make that recovery path discoverable rather than forcing users or parent agents to infer it from clipped text.

[Hmbown]

Follow-up from a live v0.8.43 screenshot of the Work / Tasks / Agents panels: make these rows as inspectable as possible, and make mouse click support a progressive enhancement where the terminal backend can deliver mouse events.

Desired behavior:

• Keyboard-first remains required: focused row + Enter inspect, Space expand/collapse, y copy id, Y copy receipt/details.
• Mouse support, when enabled, should map naturally: click focuses/inspects a row, double-click or Enter opens details, maybe click on a chevron toggles expand/collapse.
• No mouse dependency: every clickable action must have a visible keyboard equivalent.
• Rows need real identity behind rendering; a clicked/selected Agents/Tasks/Recent Tool row should resolve to the same object that y, Y, detail view, or copy actions use.

This fits #1975, but #1975 was closed as the v0.8.43 inspectability pass. Keeping this note here for the remaining cockpit/truth-surface polish, especially the Work/Agents panels shown in the current UI.