agent_ui: Add CLI support for sending agent prompts

[jutaz]

Objective

Zed's agent is driven from inside the app. Making it reachable from the command
line opens up a category of "kick off the agent when X happens" workflows —
scripts, git hooks, CI, webhooks.

There's an existing zed://agent?prompt=… URL, but it's built for a different
purpose: zed:// is an OS-registered scheme (osx_url_schemes = ["zed"],
x-scheme-handler/zed), so any web page can navigate to it. It pre-fills the
message editor and shows a "Review Before Sending" callout, deliberately leaving
the final keystroke to a human. That's the right behavior for an untrusted
caller.

Requested in #58262, which asks for exactly these two capabilities — enumerating
threads and sending a user message to an existing one — for automation and
orchestration workflows. Adjacent asks: #54221 (scheduled prompts, which needs a
way to deliver a prompt later) and #49313 (controlling agent runs remotely).
Distinct from #59146, which proposes running the agent as a standalone headless
binary; this drives the agent already running in your editor.

Solution

Add a --agent flag group to the zed CLI. It travels over the existing
zed-cli:// IPC channel, which is not OS-registered and is only reachable from a
local process, so it can submit directly. The zed://agent path is not touched
by this change.

zed --agent "Summarize the changes on this branch"
zed --agent --agent-new "Investigate the flaky CI job"   # prints the new thread id
zed --agent --agent-thread 567079b5 "Now check the tests"
zed --agent --agent-wait "Run the tests and fix what breaks"
zed --agent-list
curl -s https://example.com/task | zed --agent -

Highlights

• New agent_ui::agent_cli module resolves a thread and dispatches a prompt into
  it. Threads are addressed by ThreadId (in full or as a unique fragment, with
  hyphens optional), by ACP session id, or implicitly as the most recently
  updated thread for the working directory.
• Prompts take the same route as pressing Enter in the panel:
  ThreadView::send_content when the thread is idle, ThreadView::add_to_queue
  when it's generating. A queued message shows up as a real queue entry that you
  can steer or delete, and it drains on AcpThreadEvent::Stopped.
• Resolution checks in-memory threads first. AgentPanel is per-Workspace
  rather than per-window, so this iterates MultiWorkspace::workspaces() instead
  of MultiWorkspace::panel(), which only sees the active workspace. A thread
  that's open in a background project is reachable without stealing focus or
  reloading it from disk.
• --agent-list prints id age title (open), and
  --agent-list-format json emits an array with id, title, updated_at,
  interacted_at, is_open, and the thread's worktree paths. The array shape
  holds even when nothing matches, so it pipes straight into jq. The prompt
  paths keep stdout to the bare thread id so id=$(zed --agent …) works.
• New CliRequest::Agent variant, appended after SetOpenBehavior so the
  existing Open and SetOpenBehavior variants keep their wire representation.
• ThreadId gains Display/FromStr. Prompts run through the existing
  ExternalSourcePrompt sanitizer, which strips bidi and control characters —
  worth keeping for a path that's often fed webhook payloads.
• --agent-profile and --agent-model configure a thread at creation and
  require --agent-new. A thread persists both, and changing the profile also
  swaps the model and cascades into any running subagents, so applying them to an
  existing thread would silently reconfigure something the user already owns.
  Both are validated before anything is dispatched — an unknown profile lists the
  configured ones, an unconfigured model is an error — so a typo can't leave a
  thread running under the wrong tool set.
• Tool permissions are untouched: a CLI-triggered turn still asks before running
  tools, and --agent-wait blocks until you answer, matching what happens when
  you send the same prompt by hand.
• docs/src/reference/cli.md documents the new flags alongside the existing
  ones.

Testing

Verified manually against a debug bundle (script/bundle-mac -d -i), driving a
running instance over the real IPC path:

• --agent-new creates a thread and prints its id; --agent-list then shows it
  marked (open).
• --agent-thread resolves a full id, an 8-character prefix, and a hyphenless
  fragment to the same thread.
• Sending while the agent was generating printed agent is busy; message queued
  on stderr with the bare id on stdout, and the message appeared in the panel as
  a queue entry.
• With two projects open in one window, each directory's --agent-list showed
  only its own threads. --agent-project reached across to the other project,
  and a mismatched scope failed with exit code 1.
• Running from a subdirectory of a project still resolves that project's threads.
• In a repo with a linked git worktree, --agent-list --agent-list-format json
  reported the linked worktree's path, and resolving a branch to its worktree
  then piping the id into --agent-thread delivered a message to the right
  thread. --agent-list-format is rejected outside --agent-list, and an
  unmatched project still yields [].
• echo … | zed --agent - reads the prompt from stdin.
• --agent-wait blocked for the duration of the turn, then exited 0.
• --agent-profile and --agent-model are rejected without --agent-new, and
  with it a bad profile reports unknown agent profile "nope"; available profiles: ask, coding-agent, ..., write while --agent-model bogus/model and
  --agent-model notaslash each fail with their own message. A valid profile
  creates the thread, and sending to an existing thread still works.
• Argument validation rejects --agent-project without --agent, --agent
  together with --agent-list, and file-opening flags (--diff, --wait,
  --new, …) in agent mode.

A user's unsent draft is left alone: the CLI calls send_content, which unlike
send_impl never touches the message editor. There's a test pinning this, and I
confirmed it fails if the editor is cleared.

Automated coverage: 21 unit tests in agent_cli (id parsing and normalization,
prefix ambiguity, most-recent tie-breaking, path matching, prompt sanitization),
28 in cli (argument validation), and 4 in open_listener (list formatting).
The dispatch functions themselves are covered by the manual runs above rather
than by tests. ./script/clippy and cargo fmt --check are clean.

Reviewers can try this on macOS with script/bundle-mac -d -i, then use the CLI
inside the bundle at /Applications/Zed Dev.app/Contents/MacOS/cli so requests
attach to the running instance.

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:

• Added zed --agent for starting an agent turn from the command line

[cla-bot]

We require contributors to sign our Contributor License Agreement, and we don't have @jutaz on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

[jutaz]

@cla-bot check

[cla-bot]

The cla-bot has been summoned, and re-checked this pull request!