What it does

AgentWatch brings a small set of actions from real coding-agent sessions to the wrist: discover work, read recent turns, send a follow-up, respond to an approval or question, stop an eligible run, and see which files changed. Run groups recent Codex and Claude Code sessions by project. The provider’s session history remains the source of truth.

The Watch talks to an iPhone relay through WatchConnectivity. The relay reaches a paired Mac bridge over HTTP and finds the Mac through Bonjour. The bridge integrates with Codex’s app-server and Claude Code’s CLI. This is a local remote-control path that depends on the Mac being awake and reachable.

Session capabilities are explicit. A session owned by another Mac application is observed rather than given a second writer; some instructions can wait for that application to release it. Approvals belonging to a desktop application may still require the Mac. Changed-file reporting shows file names, not a full diff review on the wrist.

The interface gives priority to Your Turn, then Active, Latest, and Ask. The ambient dog is a companion layer: it never replaces the text needed to understand a session. Ambient can preserve the screen underneath and return to it through the Crown. A separate filming loop shows the dog and its mess; that recording mode is not evidence of a live agent session.

What I worked on

I worked on the product definition, Watch interaction design, and engineering across the Watch interface, iPhone relay, and Mac bridge. This is part of my AI-assisted building practice. The work includes deciding which actions belong on the wrist, preserving session identity, making offline state legible, and keeping simulator, integration, and hardware evidence separate.

What has been checked

  • Real local integrations, 24 September 2026: the Remote MVP document records 23 checks each for authenticated Codex and Claude Code. These cover discovery, history, same-session continuation, approvals, changed files, stopping, stale-command refusal, and restart reconciliation through the bridge. They are integration tests on a Mac, not 46 physical-Watch checks.
  • Physical Watch and iPhone, 24 September 2026: the audit records dictation, follow-up into the same Codex session, an approval followed by a real file edit, and changed-file reporting. Later checks on a Series 12 Watch with an iPhone 17 Pro Max record navigation, footer scrolling, approval haptics while the app is visible, stale state, and offline Retry/recovery without duplicate sessions. These supersede the earlier recorded dictation failure.
  • Simulator and automated checks: fixtures and model tests cover capability gating, offline controls, screen hierarchy, and request handling. They do not establish hardware behavior.
  • Physical Ambient checks, 25 September 2026: the documented runs include Crown return, broom interactions, persistence across relaunch, and the real three-minute timeout. Several interaction checks used accelerated debug timing; the product-timeout check did not.

What remains unproven

These checks do not establish broad off-LAN reliability, general battery performance, or reliable approval notifications with the wrist down. The observed approval haptic depends on the app receiving state while active; wake/recovery checks are not an always-on notification guarantee. Personal device installation is not consumer-release availability.

Keep simulator evidence apart from what the real Watch proved. The factual sources for this record are the project’s Remote MVP, Validation, and Ambient documents; their checks describe particular builds and environments.