The process_output tool result carried only output, exit code, and
running state, so chat transcripts rendered every poll as a generic
"Process output" row with no indication of which process produced it.

Return the process command from the agent's process output endpoint
and propagate it through the chatd tool result. The UI now renders
"Checking <command>" while the process runs and "Checked <command>"
once it exits, mirroring the execute tool's verb-first label grammar.
Results without a command (older agents, historical transcripts) fall
back to the previous generic label.

Also align the process output body with the execute tool: tinted
rounded panel with a scrollable max-h-64 viewport instead of the
bordered box, and drop the bespoke secondary expand/collapse button
in favor of the standard header toggle plus scrolling.

Mirroring the execute tool, process_output now takes an optional
model_intent describing why the agent is checking the process. The
chat header composes it with the command (e.g. "Waiting for the dev
server to be ready on npm start"), reusing the execute intent
sanitizer to strip redundant command references. Calls without an
intent keep the previous Checking/Checked labels.

go generate emits third-party import grouping; make gen applies gci
with the repo's coder/cdr.dev prefix group on top. Run the formatter
so the gen freshness check passes.

Two review fixes:

- process_output results now serialize running:true when the process
  outlived the poll's wait, instead of only embedding it in the note
  string. The chat UI reads the flag so a completed poll over a live
  process renders "Checking <command>" (and keeps its running
  spinner) instead of "Checked <command>".
- Backgrounded execute calls without a model intent now render
  "Started <command> in the background" instead of "Ran <command>",
  restoring a background-specific label for the no-intent case.
- Update the execute schema test for the reworded model_intent
  description from the background framing change.

A process_output poll that timed out with running:true kept the row
labeled "Checking <command>" and suppressed the killed-by-signal icon
even when the transcript recorded a subsequent successful
process_signal for the same process. The snapshot was true at poll
time; the kill is newer information and wins.

The ProcessOutputRunningThenKilled story queried the indicator by
lucide's private CSS class, violating FE10. Wrap the icon in an
aria-labeled span (mirroring the execute tool's existing pattern) and
query it by role and accessible name instead.

background_process_id is set on both intentional run_in_background
launches and foreground commands that exceeded their timeout (so the
caller can re-attach), and the UI derived "backgrounded" from its
presence alone. A timed-out foreground command was therefore labeled
"Started ... in the background" and lost its meaningful
wall_duration_ms suffix.

ExecuteResult now carries an explicit backgrounded flag set only on
the intentional path, and the frontend derives the background label
and duration suppression from it.

Transcripts recorded before the backgrounded result flag existed
carry the launch intent in the persisted call args
(run_in_background: true). Falling back to the args when the result
predates the flag keeps legacy rows labeled as background launches
while foreground timeouts (backgrounded omitted, not false) still
render as failures with their duration.

The execute tool promotes "cmd &" to background mode and strips the
ampersand locally, but the persisted tool-call args keep the original
command with no run_in_background. Those legacy rows have neither the
backgrounded flag nor the args flag, so they lost the background
label and gained a meaningless spawn-duration suffix. Detect the
trailing-ampersand form (excluding && and |&) as another legacy
fallback. Also drop a story comment that restated its assertion.

SIGTERM is catchable: the agent delivers it and returns without
waiting for exit, so a successful terminate must not flip a poll
row's recorded running state to "Checked". Only an uncatchable kill
overrides the snapshot. Also drops story comments that restated
their assertions.

run_in_background in the persisted args records intent, not outcome:
a failed StartProcess returns an error result with neither flag nor
process ID, and the legacy fallback alone labeled it "Started ... in
the background". Gate the args fallback on a nonempty
background_process_id so only launches that produced a process count.

Tool output viewports were plain scrollable divs with no tab stop, so
keyboard-only users in browsers that do not auto-focus overflow
containers could not reach content below the fold. Add
viewportTabIndex={0} to every tool ScrollArea so the regions accept
focus and scroll by keyboard. The shared ScrollArea already forwards
viewportTabIndex; a positive default there would break the one
consumer that opts out with -1, so call sites opt in explicitly.

site/AGENTS.md requires every tabIndex={0} element to carry a
semantic role; the scroll-area sweep created fourteen anonymous
generic focus stops. ScrollArea now accepts viewportAriaLabel and
renders role="region" with it on the viewport when set, and each
tool output region supplies an accessible name (process output,
command output, diffs by path, file contents by name, subagent
prompt/response/report, advisor, models, summary, build log).