agent_ui: Surface saved terminal tool output after restart

[dandv]

Objective

Fixes #57230: after restarting Zed, terminal tool calls in a reloaded thread showed only the command, and the output persisted in threads.db was unreachable (no expand/collapse controls).

Solution & Showcase

Before: when a thread was reloaded, Thread::replay recreated each terminal tool call without its live Terminal entity, so rendering fell through to the generic tool-call path (render_tool_call), which never offered a way to open the output that acp_thread had already reconstructed from the persisted raw_output.

After: once such a call has a result, the PR renders a header row above the command block that mirrors the UX of the live terminal card (behavior, label and icon position):

Expand.replayed.terminal.tool.output.webm

• Working directory on the left (instead of the static "Run Command" label). Calls with no cd (e.g. external ACP agents) show "Ran command" instead.
• Expand/collapse chevron revealed on hover.
• Failure indicator (red ✕ with an "Exited with code N" tooltip). Nonzero exit codes are recovered from the persisted output string with a new new agent::parse_failed_exit_code helper.

Differences from the live cards:

1. Replayed cards intentionally start collapsed regardless of the expand_terminal_card setting, to mirror the current UX of agent threads after Zed restarts. Later we might add auto-expansion during in-thread search, or an UX affordance to expand all.
2. There's a thin border around replayed output, coming from the markdown fenced code block. Signals a static snapshot rather than a live terminal, so I decided to keep it.

The PR also serves two more paths that had the same unreachable-output bug:

1. Execute-kind calls that never get a terminal entity, i.e. commands denied by the user or auto-denied by the permission layer (due to containing $VAR substitutions, for example)

User-denied.command.webm

2. External ACP agents that report command output as plain content blocks.

Testing

Automated tests:

• acp_thread: regression test for the update_fields raw-output fallback the UI depends on (test_raw_output_update_populates_content_when_empty).
• agent: round-trip test for parse_failed_exit_code against process_content.

Manual test plan, all pass

Summary: replayed cards start collapsed; chevron and copy button appear on hover; output expands/collapses on click; failed commands show the red ✕ with the exit-code tooltip.

Setup

1. In the agent panel, start a thread and send this prompt:

We're testing Zed's terminal output handling. Please run the following commands using the termnal tool:
- A successful command with output (e.g. `ls -al`)
- A successful command without output (e.g. `true`)
- A failing command (e.g. `false` → exit 1, or a `git add nonexistent.xyz` → exit 128)
- A command containing a shell variable (e.g. run `echo $HOME` verbatim)
- A long multi-line command (heredoc-style or `&&`-chained across lines)
- A command the user is likely to dent (e.g. `rm ~/nonexistent-file`).

2. Restart Zed (same --user-data-dir), reopen the same thread.

Replayed cards (after restart)

T-001 — Replayed cards start collapsed

Status: ✅ PASS pre-restructure 2026-06-11; ✅ re-verify

Repro: Reload the thread. Every terminal card shows the command but no output. This holds even with agent.expand_terminal_card: true (default) in settings (that setting only auto-expands live terminals as they are created).

T-002 — Header shows working dir, not "Run Command"

Status: ✅ PASS 2026-06-12

Repro: Each replayed terminal card has a header row above the command, showing the working directory (the tool's cd input) on the left — matching live cards. "Run Command" must not appear on any card that has a result.

T-003 — Chevron appears on hover, top-right of the header row

Status: ✅ PASS 2026-06-12

Repro: Mouse over a replayed card. The expand/collapse chevron fades in at the top-right of the header row — same position as live cards. The copy button appears on the command row below it.

T-004 — Click chevron: expand and collapse

Status: ✅ PASS 2026-06-12

Repro: Click the chevron → saved output renders below the command (markdown code block). Click again → collapses. State persists while scrolling the thread (virtualized list re-renders).

T-005 — Copy button copies unfenced command

Status: ✅ PASS 2026-06-12

Repro: Hover the command row, click copy. Pasted text is the bare command — no ``` fences.

T-006 — Failed command: red ✕ with exit-code tooltip

Status: ✅ PASS 2026-06-12

Repro: On the replayed card of the failing command, a red ✕ shows at the right of the header row (next to the chevron, same position as live). Hovering it shows "Exited with code N". Expanding shows the saved output including the failure sentinel text.

T-007 — Auto-denied $VAR command after restart

Status: ✅ PASS 2026-06-12

Repro: The auto-denied command's card shows a red ✕ (status Failed; no exit-code tooltip — it never ran). Expanding shows the denial message ("…does not allow shell substitutions…").

T-008 — Long multi-line command layout

Status: ✅ PASS 2026-06-12

Repro: The multi-line command card renders without layout breakage; long working-dir paths scroll horizontally in the header instead of wrapping or pushing the chevron/✕ out of view.

Live cards (same session, no restart needed)

T-009 — Live terminal cards unaffected

Status: ✅ PASS 2026-06-12

Repro: In the reloaded thread, ask the agent to run another command. The new card uses the live terminal rendering: working-dir header, real terminal view on expand, stop button while running, elapsed time. Compare side by side with a replayed card.

T-010 — Awaiting-confirmation preview unchanged

Status: ✅ PASS 2026-06-12

Repro: With confirm-mode permissions, trigger a command needing approval. While waiting, the card shows the "Run Command" preview label (no header row, no chevron). After approval it becomes a live terminal card.

T-011 — Auto-denied $VAR command in a live session

Status: ✅ PASS 2026-06-12

Repro: Same as T-007 but observed live, without restarting: the denied call renders as a static card (no terminal) with a red ✕.

Edge cases

T-012 — User-stopped command after restart

Status: ✅ PASS 2026-06-12

Repro: Stop a long-running command via the stop button. After restart, its card shows output text ("the user intentionally interrupted this command…" / "…timed out…") but no ✕ — the persisted result is not an error and carries no exit code. Document, don't try to fix.

T-013 — User-timed-out command after restart

Status: ✅ PASS 2026-06-12

Repro: Let a long-running command hit a timeout_ms. After restart, its card shows output text ("Command …timed out…") but no ✕ — the persisted result is not an error and carries no exit code. Document, don't try to fix.

T-014 — Empty-output success after restart

Status: ✅ PASS 2026-06-12

Repro: Run true (no output). After restart the card's saved output is "Command executed successfully." — chevron present, expanding shows that text.

T-015 — Themes

Status: ✅ PASS 2026-06-12

Repro: Quick pass in light + dark themes: header bg, chevron hover, red ✕ legible; card border visible but not jarring.

T-016 — External ACP agent: "Ran command" fallback label

Status: ✅ PASS 2026-06-12 (Claude Code)

Repro: In a thread driven by an external ACP agent (e.g. Claude Code / Gemini CLI) that reports Execute-kind tool calls as plain content blocks (no embedded terminal), completed command cards show a "Ran command" header label (past tense — no cd input to display) with the chevron, and the output expands on click.

Notes: Claude Code also produces genuine live-terminal cards (it creates embedded ACP terminals for some commands without passing a cwd), which render via the unchanged live path and show its pre-existing "current directory" fallback label — out of this PR's scope. The label "jumping" between "Ran command" and "current directory" within one thread reflects those two mechanisms, not a bug in this change.

T-017 — ✕ tooltip fallback when no exit code is recoverable

Status: ✅ PASS 2026-06-13 (fix shipped 2026-06-12; regression-guarded by test_terminal_failure_tooltip)

Repro: Hover the red ✕ on failed cards whose saved output is not the native terminal tool's failure sentinel:

1. An auto-denied $VAR command (live or replayed) → tooltip "Failed".
2. A failed command in an external ACP agent thread (e.g. Claude Code's "Exit code 2" format) → tooltip "Failed".
3. Reject a command at the confirmation prompt → tooltip "Failed". (Not "Rejected": the deny click sets a transient Rejected status, but the denial becomes an error tool result whose final Failed status update overwrites it — that's what is recorded and replayed. The expanded output still shows "Permission to run tool denied by user".)

A tooltip must appear in every case; "Exited with code N" still wins whenever the native sentinel is present (T-006). For comparison, release v1.7.2 shows no ✕ and no tooltip at all on these cards, live or replayed.

• Are there any parts that need more testing? Can't really think of any.
• How can other people (reviewers) test your changes? Is there anything specific they need to know? See the manual test plan above.
• If relevant, what platforms did you test these changes on, and are there any important ones you can't test? Verified on Linux (dev build). Haven't tested on macOS or Windows.

Self-Review Checklist:

• I've reviewed my own diff for quality, security, and reliability
• Unsafe blocks (if any) have justifying comments
• The content adheres to Zed's UI standards (UX/UI and icon guidelines)
• Tests cover the new/changed behavior
• Performance impact has been considered and is acceptable

---

Release Notes:

• Fixed showing terminal tool output in agent threads after restart