[rekram1-node]

Fixes #34839

Problem

When a run fails before any step starts (an unavailable model, a variant that doesn't exist, instructions that can't load), the TUI shows a "Session failed" toast and then nothing. Once the toast fades, the transcript ends at the user's prompt.

The server already stores the error on the idle marker that ends the run. #53826 rendered it on desktop and in the TUI, but the TUI row change was reverted in #54061. That version moved the error below a prompt the run never sent, and the reordering broke after /model, after a resync, and for live rows. Without it, the TUI's ExecutionError view never gets a row.

Change

reduceSessionRows now emits a row for a failed idle marker, in its stored position. Nothing is reordered, so a prompt the failed run left unsent still shows below it in its usual pending style.

 second message, are you there?
+Error: Model unavailable: opencode/gpt-5.5

The error is skipped only when the turn's last step or compaction already shows the same error (same type and message), so a step failure isn't shown twice:

on idle marker
  if it has an error and it differs from the last step/compaction error
    add a row for it
  forget the last error

A different error still shows. For example, a manual /compact that failed, followed by a run that failed for another reason, shows both.

Out of scope

• The "Session failed" toast stays as is.
• Desktop keeps its own rule, which skips the error after any failed step.

Verification

• bun test test/cli/tui in packages/tui: 428 pass, including 4 new cases in session-rows.test.ts (failure before any step, same error not repeated, different error shown, unsent prompt and resync order)
• bun typecheck in packages/tui: clean
• Ran the issue's repro script against this branch in a live TUI: the error stays in the transcript after the toast fades

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none">
  <style>
    .err { animation: show 4s ease-out infinite; }
    @keyframes show { 0%, 25% { opacity: 0; transform: translateY(4px); } 40%, 100% { opacity: 1; transform: none; } }
    @media (prefers-reduced-motion: reduce) { .err { animation: none; } }
  </style>
  <rect x="8" y="8" width="224" height="24" rx="3" fill="currentColor" opacity="0.08"/>
  <rect x="8" y="8" width="2" height="24" fill="currentColor" opacity="0.5"/>
  <rect x="20" y="17" width="120" height="6" rx="3" fill="currentColor" opacity="0.55"/>
  <g class="err">
    <circle cx="24" cy="50" r="5" stroke="currentColor" stroke-width="1.5" opacity="0.8"/>
    <path d="M24 47v3.5M24 52.5v.5" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" opacity="0.8"/>
    <rect x="36" y="47" width="150" height="6" rx="3" fill="currentColor" opacity="0.7"/>
  </g>
</svg>

---

Added by the opencode bot (AI-generated).

Related: #59177

[opencodejr]

The fix works. With the #34839 repro script (switch to a model that isn't available, then prompt), the base transcript ends at the user's prompt once the toast fades. With this PR, Error: Model unavailable: opencode/gpt-5.5 stays under the prompt (screenshots below). bun test test/cli/tui passes (428), and tsc -b in packages/tui is clean.

The duplicate check looks right: a step or compaction error that matches the run's error is shown once, a different error gets its own row, and rows keep their stored order, so a prompt the failed run never sent still shows below the error, including after a resync.

Public API is unchanged; only the TUI's row building changes.

No concerns.

Related: #59177