Execution activity

Open Activity in GateFox, or run:

gateway activity --json --config /absolute/path/gateway.json

The report belongs to this execution machine, not every connected computer. For a server, run the command there over SSH or open its Gateway app. A client-only inference configuration does not make remote execution local. The default native home is the user's .codex, not an agent's inherited private CODEX_HOME. Use --native-home /absolute/home for a custom home.

Existing service-owned SDKs + saved SDK history
  → short, ownership-verified read connections
  → Gateway Activity: running / waiting / completed / failed / interrupted / unknown

Private Gateway terminal telemetry (last 24 hours)
  → recorded outcome, with its observation time
  → current state remains unknown until verified with the SDK

A state describes a sampled turn, not completion of an entire requested objective or goal. notLoaded alone never means done. Active SDK status takes precedence over an older terminal turn. Metadata is rechecked after history; an observed history change without verified running status becomes unknown. An active turn waiting for approval or user input is explicitly waiting; the report does not imply it can continue unattended or answer it automatically. Socket replacement, conflicting owners, unavailable history and timeout do not manufacture a successful or failed outcome.

The report is a bounded snapshot, not a complete audit. It checks at most 16 existing SDK sources and four loaded-list pages per source, samples up to 64 chats and displays up to 32. Recent saved chats, loaded chats and recorded outcomes are candidates. Larger or incomplete enumerations are marked explicitly. The complete operation has a five-second deadline. Each SDK request has a two-second deadline. Cancellation closes only the observer's borrowed sockets. No engine is started, no chat is resumed and no callback is answered.

The Activity screen refreshes every 15 seconds only while mounted. Snapshots older than 30 seconds are labelled expired and cannot keep a live-looking running badge. Changing configuration clears the previous host's report; cancelled or superseded requests cannot restore it. Failed refreshes clear the report instead of presenting old data as current. No LaunchAgent, notification daemon, scheduler or new background job is installed.

The fallback reads at most the last 2 MiB of each of today's and yesterday's private telemetry files, retaining at most 100 threads within 24 hours. Only thread/turn IDs, fixed status/error classes and timestamps survive projection. Messages, arguments, prompts and credentials are neither returned nor logged. An observation timestamp is not invented as the turn's completion timestamp.

Managed worker models

For an observed SDK chat using an internal managed worker alias, Activity reads its existing private route pin and shows the selected model/account alias. The JSON row includes requestedModel and providerId; the app and CLI show the internal policy followed by the selected model. This is a read-only display: it never selects a provider, creates a pin, changes the saved chat model or restarts a worker. Missing pins leave the policy label unchanged. Unsafe pins produce partial coverage without guessing a model. A pin describes the worker's binding; the separately sampled SDK state still determines whether it is running. The selected-model projection uses this inference host's local pin store. A client-only execution host using remote inference cannot read those remote pins and retains the policy label instead.

Boundary

This is a GateFox feature, not a patch for Codex's own unread dots. Codex apps, signatures, profiles, settings, permissions and read receipts remain untouched. Activity never replays a completion notification, inference request, tool action or approval. The existing shared executor remains responsible for work while a viewing laptop sleeps.

Verification

npm run build
node scripts/run-tests.mjs dist/tests/activity.test.js
npm run app:test
node scripts/e2e-chat-state.mjs --package /absolute/verified/package \
  --root /tmp/task-specific-activity --activity

The real-SDK fixture verifies running and completed snapshots, disconnects both application clients before a synthetic provider failure, then requires that Gateway Activity expose that saved failure and retain it on refresh. It verifies no extra inference and unchanged native configuration. No paid inference or production chat is used. Rendered-window and actual laptop-sleep verification are separate from these protocol and model tests.