Skip to content

macOS: decide the public surface for "do not activate this app" (the --foreground axis collision) #3338

Description

@thymikee

Purpose

Decide the public surface for "do not bring this app forward" on macOS. Two different axes currently
want the same word: --foreground already means "include an initial interactive snapshot in a fresh
open response"
(packages/command-registry/src/flag-definitions-action.ts, key foreground:
recorded: false, projectConfig: false, and foreground: req.flags?.foreground === true && openTarget === undefined
in src/daemon/session-lifecycle/internal/session-open-execution.ts:524), while #3254 proposal 1 asks
for "do not activate the app unless --foreground is passed". #3254 says nothing about which
reading wins; #3323 implements a third answer (open -g keyed off the daemon's app backend, no flag);
nothing yet states the contract. This issue is that decision. It does not ask for an implementation to
be guessed ahead of it.

Why a runner-side "skip activation" is not free

Measured on a macOS host against the built runner (System Settings as the app under test):

  • A backgrounded macOS app can answer a stale or sparser tree: the same snapshot answered 146
    nodes in the foreground and 143 in the background. A control (foreground → background →
    foreground again) showed 38 name-matched rows with a moved rect for the foreground-vs-background
    pair and 39 for foreground-vs-later-foreground, so the rect drift is time and relayout, not
    foreground state — an earlier "79 of 143" figure in this body was an index-alignment artifact and
    is corrected. Content staleness remains possible (AppKit defers work in the background), and a
    background read carries no foreground disclosure. fix(apple-runner): serve a macOS read of a background app without raising it #3339 ships background reads on this evidence
    with a live check; treat the residual risk as accepted, not unknown.
  • An app-targeted screenshot on the XCTest backend is a screen-region grab, so a background
    window comes back with the occluding app's windows in it; activate() first returns the app's own
    content. This is why the window-level capture in RunnerTests+CommandExecution.swift keeps its
    raise (measured, not assumed) and why the helper's SCContentFilter(desktopIndependentWindow:)
    path exists. Note the backend split: on the default XCTest backend the helper's window capture
    never runs (macOsSurfaceBackend in packages/contracts/src/session-surface.ts routes the app
    surface to the helper only under appBackend === 'native'), so --fullscreen is the only
    non-raising capture fix(apple-runner): serve a macOS read of a background app without raising it #3339 adds there.
  • What the host can do without the runner is the native backend (AGENT_DEVICE_MACOS_APP_BACKEND=native,
    ADR 0031), which already acts through accessibility actions. So the flag question is mostly about
    open plus how one asks the XCTest backend to stay out of the user's screen, if it ever can.

Options, with their costs

A. Overload --foreground. Recommended against (see the recommendation comment for the full
case). Today open without it means "no initial snapshot"; src/mcp/server-guide.ts:20 and
skills/agent-device/SKILL.md:11 both instruct agents to always pass it, so "the agent's default
spelling" of an overloaded flag would mean "activate" — preserving exactly the behavior #3254 asks
to escape. The flag is recorded: false, so .ad replay cannot carry it, and
parseReplayOpenFlags (packages/ad-script/src/internal/open-script.ts) knows only --surface,
--relaunch, and --test-ime: any other token falls through into positionals, so an overloaded
--foreground in a recorded open would silently become an app-argument on replay. The flag is also
platform-neutral: iOS simctl launch and Android am start have no non-activating form, so the
overload would have to mean "activate only if the platform can", putting two unrelated promises in
one usageDescription. Do not pick this without accepting the replay default flip.

B. A dedicated activation flag (e.g. --no-activate). Costs the full threading
docs/agents/cli-flags.md lists: packages/contracts/src/cli-flags.ts, the
packages/command-registry/src/flag-definitions-*.ts owner plus its flag-groups.ts group, the open
family metadata and input schema (src/commands/management/app.ts, which today declares
foreground: optionField('foreground') and no activation key), the src/commands/cli-grammar/*
reader, src/commands/command-projection.ts, src/client/client-types.ts and
src/client/client-normalizers.ts if Node/MCP expose it, src/daemon/context.ts and
src/core/dispatch-context.ts, the handler/platform modules, the
scripts/integration-progress-model.ts classification, and a pass over
packages/contracts/src/interaction-guarantees.ts if it changes interaction semantics. Also needs a
scoping answer: is it an open option, or does it apply to every command that names the app?

C. No flag: the backend is the switch. #3323 already does this for open alone
(const openOptions = { background: hostMacOsAppBackend() === 'native' }), i.e. native backend ⇒
open -g, XCTest backend ⇒ activating open. Cheapest, and consistent with ADR 0031's opt-in model,
but it gives the operator no way to ask for a foreground open on the native backend, and it says
nothing about the commands that still activate on both backends. Note the side effect #3323 carries:
under the native backend, open <app> --surface frontmost-app stops bringing the app forward (its
ADR 0031 update states this) — an undecided behavior change riding along with the backend switch
that this decision should ratify or reject explicitly.

Required behavior

One of A, B, or C is chosen by a maintainer and written where semantics are owned: the flag definition's
usageDescription/inputDescription (versioned CLI help) for A or B, or the ADR 0031 rule plus
website/docs/docs/commands.md for C. The chosen reading must be one sentence that does not mention
the other axis.

Observable completion

  • agent-device help open states the activation contract in one place, and no second prose copy
    restates it in a way that can drift.
  • If a flag is added: its decode and daemon projection are tested, projectConfig/recorded are
    chosen deliberately and justified, and scripts/integration-progress-model.ts classifies it (the
    architecture-progress gate fails CI on an unclassified public flag).
  • If --foreground is overloaded: an existing .ad replay's behavior change is named in the changelog,
    because a recorded: false flag is absent from every existing recording.
  • The macOS docs continue to name which commands can still activate the app or move the real pointer.

Depends on: none for the decision itself. #3254 proposal 1, #3323 rule 7, and #3213 item 13 (which is
what would make non-activating raw-coordinate interaction possible at all) each take this answer as
input.

Activity

  1. thymikee commented on Oct 8, 2026

    @thymikee
    MemberAuthor

    Recommendation: against Option A (overloading --foreground). Maintainer position as of this comment; Option B or C should be chosen.

    1. Overloading is actively harmful here, not merely inelegant. The main skill tells every agent to run open <app> --foreground (skills/agent-device/SKILL.md:11) and the MCP guide says open {app, foreground: true} (src/mcp/server-guide.ts:20), so "the agent's default spelling" of the flag would mean "activate" — the overload would preserve the exact behavior macOS: drive a local app without activating it or moving the real pointer #3254 asks to escape for precisely the callers that matter.
    2. The flag is recorded: false, so .ad replay cannot carry it, and parseReplayOpenFlags (packages/ad-script/src/internal/open-script.ts) knows only --surface, --relaunch, --test-ime, --no-test-ime: any other token falls through into positionals, so an overloaded --foreground in a recorded open would silently become an app-argument on replay.
    3. Correction to the evidence in the body: the "79 of 143 rows differing" figure was an index-alignment artifact. The measured control (foreground → background → foreground) gave 38 name-matched rows with a moved rect for foreground-vs-background and 39 for foreground-vs-later-foreground, so tree drift is time and relayout, not foreground state. That removes the strongest argument that a non-activating XCTest read is structurally unsafe; the remaining risks are content staleness (AppKit defers work for background apps) and the disclosure gap, both real but survivable — fix(apple-runner): serve a macOS read of a background app without raising it #3339 ships background reads on that basis with a live check.

    Side effect worth naming for the decision: #3323's open -g under the native backend also changes open <app> --surface frontmost-app — it no longer brings the app forward (its ADR 0031 update states this), which is an undecided behavior change riding along with the backend switch.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions