[agent-cli] Move the command-socket reload rung into its own module by vonovak · Pull Request #217 · expo/expo-agent-cli

vonovak · 2026-10-27T22:46:32Z

Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.

This PR moves the reload broadcast into its own module, so the change in #207 to what a reload reaches reads as a behavior diff.

Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 5/8).

Problem

The command-socket rung (the dev server's reload broadcast and the peer-churn wait after it) is inside src/runtime/reload/reloadAsync.ts. That file also holds the reload policy, the evidence wait and the relaunch. #207 changes this rung. A move inside the same diff would hide that change.

Fix

A pure move, with no change in behavior. reloadOverDevServerAsync, the CommandSocketChurn type, PEER_CHURN_TIMEOUT_MS, PEER_POLL_INTERVAL_MS and the private waitForPeerChurnAsync move to the new src/runtime/reload/rungOne.ts. The two constants are exported because reloadAsync.ts still reads them. reloadAsync.ts, smoke/smokeAsync.ts and reloadAsync-test.ts import them from rungOne.ts. The reloadAsync.ts code-line cap goes down to 922.

Red test: none, because the move changes no behavior. The unit suite passes with only import changes.

Live check: run at the stack top only. This layer moves code and changes no behavior. At the top, rung 1 broadcast on a local iOS simulator and was proved by message-socket-peers on 3 of 3 reloads.

Watch: reloadAsync.ts and rungOne.ts share the two churn constants until the next split.

Not in this PR: the split of reloadAsync.ts into evidence, relaunch and report modules. #207 starts it with evidence.ts.

🤖 Generated with Claude Code

Kudo · 2026-10-29T18:18:03Z

+ * `appsReconnected: 0` [observed — live tier, 2026-08-27, twice], and a reader reconciling those two
+ * numbers was reconciling a label against a **different signal's** count: `appsReconnected` is what
+ * `fresh-debugger-target` rests on. Every label now carries its own count, and a label whose count is
+ * zero cannot be named ({@link hasEvidenceFor}).


{@link hasEvidenceFor} does not resolve from this file. hasEvidenceFor stayed in reloadAsync.ts and is not imported here. Importing it would cycle, because reloadAsync.ts already imports this module. Name the function and its file in the sentence instead.

Fixed in f3efa73

A pure move, with no change in behavior. The command-socket rung moves from src/runtime/reload/reloadAsync.ts to the new src/runtime/reload/rungOne.ts: reloadOverDevServerAsync, the CommandSocketChurn type, PEER_CHURN_TIMEOUT_MS, PEER_POLL_INTERVAL_MS and the private waitForPeerChurnAsync. The two constants are exported because reloadAsync.ts still reads them. reloadAsync.ts, smoke/smokeAsync.ts and reloadAsync-test.ts import them from rungOne.ts. The reloadAsync.ts code-line cap goes down to 922. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…since the link did not resolve (llp/0021 §The rules) hasEvidenceFor lives in reloadAsync.ts, which imports rungOne.ts, so an import for the link would cycle. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…onnection (#202) **Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR gives every runtime command one decision about which apps a run reads, so each layer above changes one place. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 1/8). ## Problem On main, each runtime caller (reads, waits, the bundle check, smoke) builds its own scope from the flags, and the device facts from simctl and adb are mixed with decisions. smoke also keeps a failed device lookup for the rest of the run, so a later phase cannot recover. ## Fix `TargetSelection` is the run's scope, `DeviceInventory` holds the device tools' facts, and one resolver classifies the targets. `RuntimeConnection` carries the classified targets, and every reader, wait and deferred command reads `classifiedTargets.read`. An attachment wait keeps the facts for an unchanged list and reads them again, within its budget, for a changed list. smoke keeps a successful device choice until a boot or a session start replaces it, and does not keep a failed lookup. The rules are in llp/0005 §Which app a run reads. Red test: smoke targetSelection-test "asks the device tools once for a list that does not change" (4 simctl calls before the fix, 1 after). Also: "retries a failed device lookup, shares concurrent discovery, and retains the successful choice". Live check: run at the stack top only, not at this layer. The local iOS simulator runs (#207, #218) and the EAS session runs (#203, #200) go through this layer's scope, classifier and inventory refresh. None of them counted device-tool calls or forced a failed lookup, so this layer's claim rests on its red tests. Watch: an unscoped wait now asks the device tools once for each changed list that holds an unplaced target. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

…ps on local devices (#203) **Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR makes an EAS Simulator scope read the session's app and skip the apps that local devices run. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 2/8). ## Problem The app on an EAS Simulator iOS session registers under the cloud simulator's model name, for example `iPhone 17`. No local tool reports that name, so nothing placed the target (DF3, N2-06, N2-09). `runtime:tree --ios` refused with "No app connected", and smoke reported `app` unknown after 192 s. ## Fix On an EAS Simulator scope for iOS, or with no platform flag, a target that a local device tool names is skipped. The one target that nothing places is the session's app. While the scope's device tools give no answer, the targets stay undetermined. `--eas` alone is a cloud scope. An attachment wait reads the facts again for a changed list. It asks a tool that gave no answer again at most once a second, and its first read counts as the first ask. Red tests: runtime targetSelection-test "holds every unclaimed target while the local simulators are unknown"; smoke targetSelection-test "asks a failing simctl again at most once a second while the wait polls"; preflight-test "should refuse two apps nothing places, and say how to name the session's platform"; e2e cloud-simulator-test "reloads the session's app, not the app on a booted local simulator" and "reads neither unclaimed app with no platform". Live check: at the stack top, on an EAS iOS session (Expo Go, v2 tunnel) with a local simulator app on the same dev server: `runtime:reload --eas` skipped the local app as `other-device` and proved 2 of 2 reloads by `debugger-context-replaced`; `runtime:reload --ios` skipped the cloud app and proved 3 of 3 the same way. The `live-cloud` suite passed 17 of 17 at the stack top. Watch: a session whose model name equals a booted local simulator's name is skipped as a local app. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

**Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR makes runtime reads and smoke skip an app on a simulator that another worktree bound. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 3/8). ## Problem A booted simulator that another worktree bound loads its JavaScript from `127.0.0.1:8081`. When this worktree's dev server takes port 8081, that app connects to it, and runtime reads read it as this worktree's app (DF22–DF24, root cause DF29). In worktree `mapfilter`, bound to an Android EAS session, `runtime:tree --json` returned `bundle.platform: "ios"`. It read worktree `legend`'s simulator, with old JavaScript. ## Fix A target is another worktree's app when every booted simulator with its device name has a udid that only other worktrees bind with a live lease. When the simulator list cannot be read, a target whose name only other worktrees' bindings hold is skipped. Runtime reads, the debugger client and smoke skip that app, with or without a platform flag, and name the device and the worktree on stderr. `runtime:reload` lists it in `skippedApps` and does not check its platform's bundle. A released or expired binding does not count. smoke names the skipped apps once per run, in its report and on stderr. A runtime read whose list empties after the preflight refuses like the preflight: `NO_APP_CONNECTED`, exit 22. An `--android` run does not ask `simctl` because another worktree holds an iOS lease. The no-app refusals live in `scopeRefusal.ts`, so the debugger client refuses the same way without importing the preflight. Red test: targetSelection-test "skips a target with a foreign binding's name when the simulator list could not be read". Also: e2e runtime-target-selection-test "runtime:eval skips a renamed foreign simulator and names its worktree". Live check: run at the stack top only, not at this layer. The local iOS simulator and EAS session runs (#207, #203) go through this layer's binding lookup, but no app from another worktree was attached, so this layer's claim rests on its red tests. Watch: at this layer, a broadcast reload still reaches the skipped app. #207 stops that. Not in this PR: one `EmptyScopeReason` model for the refusal texts. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

…bound to it (#206) **Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR makes the device bound to this worktree the scope of every runtime read and of smoke, with or without a platform flag. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 4/8). ## Problem N2-33: a worktree bound to an iOS EAS session ran `runtime:tree`, which read a local `iPhone 17 Pro` app with the same bundle id and named the session as the device it read. N2-30: on the same kind of worktree, `runtime:tree --ios` refused with "No app connected" while the session's app was connected. ## Fix The scope is this worktree's local binding on the run's platform while its lease is live, else the one live cloud binding, else the flags. A shut-down bound simulator has no app, so a read refuses and names it, and smoke boots it. On Android, the run reads the app whose device name carries the bound serial's model, in the spelling `adb devices -l` uses (`Pixel 9` is `model:Pixel_9`). With two devices of that model, only the one that alone reverses the port is read. An EAS Simulator Android session skips an app on an attached emulator of its model. A wait asks `adb reverse` again for a changed list, so a twin that attaches mid-run cannot authorize a read. One function, `silentTool`, names the device tool that gave no answer, for placement and for the wait's retry. A command with no dev server refuses `NO_DEV_SERVER` before it reads a binding; an unreadable EAS binding refuses `NO_BOUND_DEVICE`. `AGENT_CLI_NO_DEVICE=1` skips only this machine's bindings, so an EAS session still scopes the run. `runtime:reload --json` names the scope under `scope`: platform, how it was chosen, whether it is bound, and the device's backend, id and name. Red tests: targetPlatform-test "matches %s to the attached Android device adb names model:Pixel_9"; e2e runtime-bound-selection-test "eval with android-cloud, flag none, reads androidSession"; targetSelection-test "classifyTargets on a bound phone whose model adb spells differently". Also: e2e "runtime:tree refuses with NO_APP_CONNECTED and does not name the session as read". Live check (adb, on an API 29 emulator whose `Build.MODEL` is `Android SDK built for arm64`): `adb devices -l` printed `model:Android_SDK_built_for_arm64`, and the match accepted the React Native name `Android SDK built for arm64 - 10 - API 29`. `adb reverse --list` printed `host-14 tcp:8197 tcp:8197`; the parser read `tcp:8197` / `tcp:8197`. At the stack top, `runtime:reload --android` read the bound emulator by that model, with an EAS Android app on the same dev server skipped, on 3 of 3 runs. Watch: no physical phone was checked. Not in this PR: one device type for `SelectedDevice` and `NavigateDevice`; one `EmptyScopeReason` model for the refusal texts. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

…ext (#218) **Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR gives #207 a proof, per debugger target, that a reload replaced that app's JavaScript context. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 6/8). ## Problem A `Page.reload` keeps the target's page id, so a fresh target cannot prove it. On main, nothing proves that one target's context was replaced. An acknowledged `Page.reload` does not show that the app reloaded. ## Fix `jsContextReplacedAsync` sets a global with a per-check name, takes the action, and reads the global on the same target id until it is gone. One debugger session owns the marker, the action, the reads and the cleanup. A context that cannot keep the marker is unmarkable, and the action is not taken. An action that the app answers with a CDP error is thrown. An action whose answer is lost may have run, so the marker is read anyway. Each phase has its own bound, and an absolute deadline bounds the whole proof. The open handshake counts against a request's budget, and a request that times out names the whole budget the caller gave. The phase table is in llp/0005 §What proves a reload. Red tests: contextReplaced-test "does not mistake another concurrent check for a replaced context"; "the proof of a Page.reload" rows "reads the marker when the socket drops before Page.reload answers" and "does not prove a reload whose answer was lost and whose context stayed"; cdpClient-test "should name the caller's budget when the socket opened late". Live check: at the stack top, on an SDK 58.0.21 development build on a local iOS 27.0 simulator, with a second app on another simulator so that the run skips one app: `runtime:reload --ios --json` reported `verifiedBy: debugger-context-replaced` (`outcome: context-replaced`) on 4 of 4 reloads, exit 0. On an EAS iOS session (Expo Go, v2 tunnel) with a local simulator app on the same dev server: `runtime:reload --eas` skipped the local app as `other-device` and proved 2 of 2 reloads by `debugger-context-replaced`; `runtime:reload --ios` skipped the cloud app and proved 3 of 3 the same way. On Android at the stack top, with Expo Go on a local API 29 emulator and on an EAS Android session on one dev server: `runtime:reload --android` skipped the cloud app and proved 3 of 3 by `debugger-context-replaced`; `runtime:reload --android --eas` skipped the local app and proved 3 of 3 the same way. Watch: on Android the proof was seen with Expo Go only, not yet with a development build. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

…skips an app (#207) **Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR makes a reload reach only this worktree's apps when the run skips an app, and prove the reload with this worktree's apps only. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 7/8). ## Problem The dev server sends its reload broadcast to every app on its port. N2-17: a worktree bound to an iOS EAS session ran `runtime:reload`. It printed "Skipped 1 app on local simulators … iPhone 17 Pro", then "Reloaded yes · via the dev server's client command socket … 2 client(s)". The broadcast also reloaded the skipped app. ## Fix Rung 1 is decided on one target list read after the bundle gate. While the scope skips an app, rung 1 sends no broadcast. It sends CDP `Page.reload` to each own target and proves each with the context marker, under one shared deadline. When nothing is skipped, rung 1 stays the broadcast, and the attempt says when it also reached an app outside the scope. Peer churn beside an app outside the scope makes the attempt `ok: false, delivered: true`: the reload then needs a fresh own target or a bundle. Nothing of an app outside the scope proves the reload. The failure text says whether nothing was watched, the churn was discounted, or the watched observations did not happen. When no own app could keep the marker, it says no `Page.reload` was sent. It names `--eas` and the tunnel only on a cloud run. `Page.reload` has 8 s to prove itself, and the first proof stops the other checks (`stopped`). A `Page.reload` whose answer was lost is not delivered. When the read before rung 1 fails, the run keeps its first placement and sends nothing (mechanism `none`). After a relaunch that nothing can prove, local or cloud, the run exits 22 at once instead of waiting out `--timeout`. CI runs the reload cases that need a local simulator on the macOS job. Red tests: rungOne-test "a broadcast beside an app nothing places: no bundle proves it"; e2e runtime-target-selection-test "sends Page.reload to its own app only, and a page of the skipped app proves nothing ($pageReload)", "sends no Page.reload when its own app cannot keep the marker" and "runtime:reload does not take another platform's churn as its proof". Live check (before the 8 s window and the relaunch exit; not rerun since): at the stack top, on an SDK 58 development build on a local iOS simulator. With a second app on another simulator, the run skipped it as `other-device`, sent `Page.reload` to its own app only, and proved 4 of 4 reloads by `debugger-context-replaced`. With one app, it broadcast and proved 3 of 3 by `message-socket-peers`. On an EAS iOS session (Expo Go, v2 tunnel) with a local simulator app on the same dev server: `runtime:reload --eas` skipped the local app as `other-device` and proved 2 of 2 reloads by `debugger-context-replaced`; `runtime:reload --ios` skipped the cloud app and proved 3 of 3 the same way. Watch: when rung 1 broadcasts, an app that attaches after the last target list also gets the broadcast. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

**Every runtime command acts on the device bound to this worktree, and nothing else attached to the same dev server, and says which device it acted on.** This PR makes `navigate` decide "attached" from the device it opens. Stack: #202 → #203 → #204 → #206 → #217 → #218 → #207 → #200 (this is 8/8). ## Problem `navigate` counted the attached apps of every platform. N3-25 (also DF22): after `dev --android --eas`, with only a local `iPhone 17 Pro` app on port 8081, `navigate "/street/…"` with no flag counted the iOS app as attached. It skipped the launcher, opened only the route link, and exited 22 after 120 s. On Android, the stuck-app recovery could also stop an app that another device ran on the same dev server. ## Fix `navigate` validates the route, resolves the device, and then reads the attached apps of that device's scope to decide the app kind, the route URL and the launcher. One reader serves the launcher wait and the route wait. It reads the device facts again only for a changed list, within the budget of the wait that polls. So a cloud app that attaches after the accepted Expo Go dialog counts. When an Android twin of the bound device is attached, `navigate` reads `adb reverse` before the launcher decision and again after this run's own `adb reverse`. When nothing attached on Android, the recovery stops `--app-id`, else the app config's app, else Expo Go. smoke hands its device to the route open, so the open and the waits are about one device. `--print-url` reads the scope as `navigate` does, `adb reverse` included, once the dev server lists an app. The recovery stops Expo Go when this run opened it, and says the app was stopped only when `pidof` saw it running. The first commit is a pure move of route URL resolution to `src/navigate/routeResolution.ts`. The second commit holds the behavior. Red tests: navigate-target-selection-test "no flag, bound Android emulator, only an iOS app and another model's Android app attached"; navigateAsync-test "asks adb for the device list once while the attach wait polls an unchanged list"; e2e cloud-simulator-test "counts the app that attaches after the dialog is accepted"; navigate-target-selection-test adds three Android twin rows, one of them "no flag, bound Android emulator beside a twin, neither reverses the port before this run". Live check: the `live-cloud` iOS suite passed 17 of 17 at the stack top on an EAS session: `navigate / --eas` exit 0, attached after 3 ms. Plain main gives the same results. Watch: after a wait's budget ends, its later polls keep the last device facts. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

vonovak added this pull request to stack #219 October 27, 2026 22:47

vonovak force-pushed the codex/reload-extract-rung branch from 4c148bf to ff58709 Compare October 28, 2026 08:48

vonovak force-pushed the codex/reload-extract-rung branch from ff58709 to ad3416f Compare October 28, 2026 10:12

vonovak force-pushed the codex/reload-extract-rung branch from ad3416f to c6a994b Compare October 29, 2026 03:54

vonovak changed the title ~~[agent-cli] Extract the shared command-socket reload rung~~ [agent-cli] Move the command-socket reload rung into its own module Oct 29, 2026

vonovak force-pushed the codex/reload-extract-rung branch from c6a994b to b8a9ebf Compare October 29, 2026 04:33

vonovak force-pushed the codex/reload-extract-rung branch from b8a9ebf to dea2f41 Compare October 29, 2026 05:18

vonovak force-pushed the codex/reload-extract-rung branch from dea2f41 to 05ecf55 Compare October 29, 2026 12:55

vonovak force-pushed the codex/reload-extract-rung branch from 05ecf55 to f71e57c Compare October 29, 2026 14:05

vonovak force-pushed the codex/reload-extract-rung branch from f71e57c to 69a5b35 Compare October 29, 2026 14:11

vonovak force-pushed the codex/reload-extract-rung branch from 69a5b35 to b6fed77 Compare October 29, 2026 14:13

vonovak force-pushed the codex/reload-extract-rung branch 2 times, most recently from 4a80661 to 2c719ef Compare October 29, 2026 14:19

vonovak force-pushed the codex/reload-extract-rung branch from 24b37d6 to 2fe7eb1 Compare October 29, 2026 15:01

vonovak marked this pull request as ready for review October 29, 2026 15:16

vonovak force-pushed the codex/reload-extract-rung branch from 2fe7eb1 to 0ce65a7 Compare October 29, 2026 18:46

Kudo approved these changes Oct 29, 2026

View reviewed changes

vonovak force-pushed the codex/reload-extract-rung branch 2 times, most recently from 5e642c5 to 34ed6c0 Compare October 29, 2026 19:34

vonovak force-pushed the codex/reload-extract-rung branch from 34ed6c0 to e61caf1 Compare October 29, 2026 19:52

vonovak force-pushed the codex/reload-extract-rung branch from e61caf1 to 830b6d3 Compare October 30, 2026 08:41

vonovak force-pushed the codex/reload-extract-rung branch from 830b6d3 to 28cdbee Compare October 30, 2026 09:09

vonovak force-pushed the codex/reload-extract-rung branch from 28cdbee to 95ffb64 Compare October 30, 2026 09:22

vonovak and others added 2 commits October 30, 2026 10:34

vonovak force-pushed the codex/reload-extract-rung branch from 95ffb64 to f3efa73 Compare October 30, 2026 09:36

vonovak merged commit 2af78d8 into main Oct 30, 2026
13 checks passed

vonovak deleted the codex/reload-extract-rung branch October 30, 2026 10:11

Navigation Menu

Repository navigation

[agent-cli] Move the command-socket reload rung into its own module - #217

[agent-cli] Move the command-socket reload rung into its own module#217
vonovak merged 2 commits into
vonovak/fix-cloud-scope-own-targetfrom
codex/reload-extract-rung

vonovak commented Oct 27, 2026 •

edited

Loading

Uh oh!

Kudo Oct 29, 2026

Uh oh!

vonovak Oct 30, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

vonovak commented Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Uh oh!

Kudo Oct 29, 2026

Choose a reason for hiding this comment

Uh oh!

vonovak Oct 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

vonovak commented Oct 27, 2026 •

edited

Loading

vonovak Oct 30, 2026 •

edited

Loading