Skip to content

[agent-cli] Fix local reload connections and emulator port allocation - #145

Draft
vonovak wants to merge 41 commits into
codex/eas-session-bindingfrom
codex/local-device-reload
Draft

[agent-cli] Fix local reload connections and emulator port allocation#145
vonovak wants to merge 41 commits into
codex/eas-session-bindingfrom
codex/local-device-reload

Conversation

@vonovak

@vonovak vonovak commented Oct 12, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Local runtime:reload --method device previously stopped the app and reopened a bare route link, which could leave a development client at its launcher or connected to its previous Metro. This change opens a development-client connection URL for the selected Metro, even when an old debugger target remains listed, and forwards the port on the selected Android emulator. It resolves the device before stopping, respects the remaining reload budget, and sends an optional route once after the existing reconnection evidence wait.

Android emulator allocation also reserves the console ports of every listed emulator, including offline instances started outside the CLI. A separate inventory supplies port reservation; the existing readiness filter still controls command targets. This prevents starting a new emulator on an occupied port while another instance is booting or disconnected from adb.

Expo Go detection receives only an explicit --app-id as an override. A configured bundle identifier or package name no longer misclassifies a disconnected Expo Go project as a development build. The stop and reopen use Expo Go's platform-specific app ID when Expo Go is selected, and explicit development-build overrides retain their behavior.

Depends on #139.

Validation:

  • Initial reload change: 4,980 unit tests and 22 built-CLI reload integration tests passed; build, typecheck, lint, and formatting checks passed.
  • Offline-port fix: the new regression cases reproduced occupied-port reuse and failure to report exhausted ports before the fix. All 397 device, binding, and device-selection tests now pass, including rejection of an unbound offline emulator as an explicit command target. Typecheck, lint, and formatting checks passed; lint reports existing warnings.
  • Expo Go detection fix: 561 related unit tests passed, including disconnected iOS/Android projects with configured app IDs and explicit development-build overrides. All 22 built-CLI reload integration tests passed, including the configured-bundle-ID regression. Build, typecheck, lint, and formatting checks passed; lint reports existing warnings.
  • Live reload checks on iOS: recovered from the wrong Metro, retained saved data, and delivered the requested route once.
  • Live reload checks on Android: recovered after Metro changed ports, retained saved data, and left the sibling test app unchanged.

Open issues before leaving draft:

  • The Android live test needed a later manual link to deliver the requested route. Debugger reads were unavailable in these test builds; execution and saved data were checked through app-originated reports.

vonovak and others added 30 commits October 8, 2026 16:14
@vonovak @claude
[agent-cli] Verify the project root of a scanned dev server (llp/0004…
712df42 
… §Discovery ladder)

The port scan took the first answering server, which on a machine with parallel worktrees is often another project's. The scan now reads X-React-Native-Project-Root from GET /status and compares it with this project's root. A foreign server is skipped. When only foreign servers answer, discovery is unreachable and reports them in foreignServers, and status prints them. A server with no header is accepted with projectRootVerified: false.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9CwPuoTfcRMrAGrgqGRza
@vonovak @claude
[agent-cli] Discovery rejects a server whose project root it cannot r…
4169ba2 
…ead, and checks the log port (llp/0004 §Discovery ladder)

A /status that times out is no longer the same as no header: it is retried once, then the port is not accepted and the reason names it. The start.log port gets the same project-root check as a scanned port.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9CwPuoTfcRMrAGrgqGRza
@vonovak @claude
[agent-cli] the dev-server lock never publishes a port that serves an…
743203e 
…other project (llp/0004 §Discovery ladder)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9CwPuoTfcRMrAGrgqGRza
@vonovak @claude
[agent-cli] the dev --detach building test runs on its own free --por…
375f261 
…t, so a Metro on 8081 no longer skips it (llp/0004 §Discovery ladder)

The lock's fallback is the step's --port, so pinning one makes 8081 irrelevant. Run with a stand-in
foreign Metro on :::8081 answering /status with another project root: the test runs and passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] discovery's project-root rules live in llp/0004 (§Discove…
5ad026f 
…ry ladder, §A busy port)

The four commits before this one cite the rejected branch's llp/0028 and llp/0030, which main never had.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] changelog for #132
260bd7d 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the /status read gets its own 3 s budget, a named port is…
4c152bc 
… checked before the lock publishes, and the port watch requires isRunning (llp/0004 §Discovery ladder)

Review fixes for #132. A server whose /status is slow is judged within 3 s on two tries, not within the probe's 800 ms. The lock checks a named port for another project's root before it publishes, as it already did for the fallback. The port watch after a foreign fallback has no deadline, so isRunning is required. projectRootVerified and ForeignDevServer.url, which nothing read, are gone, with the duplicate @ref and a cast.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the status-read budget constant sits above discoverDevSer…
b3e6c8b 
…verAsync's doc block, not inside it

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the /status read is retried after a socket error only, an…
ff59297 
…d status reads it within its own budget (llp/0004 §Discovery ladder)

A read that used its budget is unproved at once, so a stalled /status costs one budget, not two. status passes its 1.5 s probe budget as the read's, so the specific reason reaches the report instead of its 3 s deadline cutting the read off. The constant's comment says what the budget does: a slow /status is accepted, a stalled one is unproved.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak
test(agent-cli): wait for telemetry worker exit on Windows
f94f7c1
@vonovak
fix(agent-cli): detached readiness waits for this project's Metro
7a1db8a 
During parallel native builds, another worktree can claim a detached child's initial port before its Metro starts. The parent previously accepted that foreign server's ready response and reported success while its own build continued.

Follow the child's lock throughout the readiness wait, cancel the old status request when the port changes, and reject explicit project-root mismatches. Keep one timeout budget, preserve child-failure reporting and older Metro versions without a root header, and honor --wait-ready when reusing an existing server. Final status checks also reject a foreign root.

Validation: all 4,785 unit tests pass; all 27 detached-start end-to-end tests pass, including real HTTP/lock-socket port changes and cancellation of an obsolete request. Typecheck, build, formatting and lint pass (existing warnings outside changed files).

Place immediately after #132 and before #133. The regression tests reproduce the false success against the original implementation. Native command-connection cleanup is separate Expo work.
@vonovak @claude
[agent-cli] prep for device binding: simulators.ts, adb.ts, eas.ts, R…
1caa5cc 
…unner, digestOf, and one xcrun stub (llp/0030)

Moves the simctl, adb and EAS session-stop parsers and runners into src/device, exports digestOf,
adds the Runner type, and makes installedAppStubs' installStubXcrunAsync the one xcrun stub of
navigate, stop and runtime-reload. This changes no behavior.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the stub simctl's bootstatus -b marks the simulator Boote…
f7985b2 
…d, as the real one boots it (llp/0030)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the device registry: one iOS simulator per worktree, boun…
4e51f5d 
…d under a lock and read by every verb (llp/0031 §The module)

src/deviceBinding/ holds the records, the lock, the lease, the choice, the
inspect table and the rung loop of llp/0030, with the iOS boot through
simctl bootstatus -b. No caller yet. The reap and the Android half arrive
with llp/0032 and llp/0033. The five design files land with it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] dev binds the worktree's iOS simulator, and every verb re…
d9b0409 
…ads the binding (llp/0031 §How `dev` uses it, §Readers, §smoke)

The resolver binds for a serving draft before its presence probe; a
building draft is bound after the refusals; every run and install step
carries --device <udid>, and the open takes the bound device. dev --detach
on a running server reuses the parked simulator or refuses not-reusable,
and the child's DEVICE_* and NO_BOUND_DEVICE refusals reach the caller
unchanged through the verdict's code row. navigate, runtime:*, status and
the installed readers walk the rung loop; the EAS fallback accepts the
dotenv id only. smoke's iOS boot is the registry's, left bound. The
first-booted probe, pickSimulator, the iOS boot and NO_DEVICE go.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] e2e seeds the binding where a verb meets a booted simulat…
881bee8 
…or, two worktrees get two simulators, and the docs name the bound device (llp/0031 §Tests, §Docs)

The stub xcrun binds the booted simulator it lists; the smoke boot cases
assert the registry file and bootstatus -b; a two-worktree dev case pins
distinct E2E-CREATED-* udids on each run:ios --device. The first-booted
sentences of llp/0004, 0005, 0022, 0025 and 0026 read the bound device.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] changelog link for #134
80c28f9 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the review of #134: absent survives a missing adb, the Tr…
3bacb86 
…y line stays the dev command, the relayed refusal keeps its data, and the old simulator reader goes (llp/0031 §How `dev` uses it)

A Mac without an Android SDK reaches `absent` again, because the iOS
inspection answered. A NO_BOUND_DEVICE refusal keeps `dev` as its Try
line; --print-url is an Or: door. The detached child's code row carries
`data`, and only the Try line next to it counts. smoke keeps the How line
in its boot reason and prints the acquire line. The EAS fallback asks the
service only when the dotenv names a session. The booted-only simulator
reader is deleted. The named tests produce the child's log through
logCmdError, forward a --device the pin replaces, and pin the acquire line
before the report. 0031 says the order built and the decisions taken;
0030 gains the boot-failed row.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] a created simulator is a mainstream iPhone on the newest …
591c236 
…runtime that lists one, not the newest runtime's only iPhone (llp/0030 §Choice)

iOS 27.1 lists iPhone Duo alone, and no app rendered on it. The choice now
walks the available runtimes newest first and takes the first with an
`iPhone <n> Pro`, else an `iPhone <n>`; only when no runtime has one does
the newest runtime's first iPhone apply. The RFC's §Choice sentence says
so, and §Risks gains the observation.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] dev --ios binds nothing off macOS, and the module's tests…
d333338 
… pass on Windows (llp/0031 §How `dev` uses it)

The live-cloud suite runs dev --ios --no-open on a Linux worker to serve for an EAS session it opens itself; the acquire ran xcrun there and failed. iOS simulators exist on macOS only, so the callback binds nothing elsewhere, as on main. On Windows the function-length guard split on LF and never saw a lone closing brace in a CRLF checkout, and the binding's project root is the canonical path, which is a drive path there.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] a bound simulator whose runtime is gone counts as gone, s…
4ec8e63 
…o dev creates on a runtime that works (llp/0030 §Choice)

simctl keeps listing such a simulator with isAvailable false. The reuse took it, bootstatus failed, the failed boot kept the binding, and every retry picked it again. The choice and the listing now read an unavailable simulator as missing: the binding is removed and a new simulator is created, and read verbs report gone.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak
test(agent-cli): model the iOS host in device binding cases
019b827
@vonovak @claude
[agent-cli] the registry spawns a read-only emulator instance per wor…
9c83c32 
…ktree on a free console port, and kills only what it spawned (llp/0032 §Android boot)

acquireDeviceAsync takes 'android': adb devices -l and emulator -list-avds before the lock, busyPorts and the first free even port in 5554..5584 under it, the detached spawn and one binding with serial and pid in the same section, sys.boot_completed after it. inspect reads get-state plus the pid for a spawned instance. The let-go of a spawned binding removes the file, then kills the pid only while ps names its -avd and -ports. The injected Android rung stays for its one caller until the next commit.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] every verb drives the bound emulator instance: dev --andr…
e8ae773 
…oid acquires, the readers inspect get-state, smoke binds and keeps it, and run:android gets --device <serial> (llp/0032 §Wiring)

The injected Android rung, the first-adb-device probe, the emulator boot and the emu-kill shutdown go. The presence probe asks the bound device on both platforms and pins nothing, because withDevice pins every run and install step. The installed reader takes the bound serial, and lists only to filter --device. The e2e stub adb answers get-state and sys.boot_completed, a stub emulator stays up until killed, and the dev e2e spawns two read-only instances on 5554 and 5556.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak @claude
[agent-cli] the Android binding's changelog, the llp sentences it cha…
5be6148 
…nges, and what was decided while building (llp/0032 §Decided while building)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@vonovak
fix(agent-cli): reject emulator boot after child exit
0db5f1c
@vonovak
test(agent-cli): isolate local device assumptions in CI
c591768
@vonovak
feat(agent-cli): manage worktree device lifecycle
d440865
@vonovak
fix(agent-cli): keep smoke cleanup scoped to its device
c650900
@vonovak
fix(agent-cli): make device lifecycle fixtures Windows-safe
583a2a1
vonovak and others added 10 commits October 9, 2026 10:17
@vonovak
test(agent-cli): seed worktree bindings for status follow-ups
4be2d71
@vonovak
feat(agent-cli): bind EAS sessions to worktrees
f43a989
@vonovak
fix(agent-cli): recover EAS sessions after stale bindings
800b583
@vonovak
test(agent-cli): use native paths in cloud binding fixture
aa35175
@vonovak
fix(agent-cli): keep the cloud session guard on one command line
15d364d
@vonovak
fix(agent-cli): preserve cloud controller argv through Windows shims
c805ff1
@vonovak @claude
test(agent-cli): decode the base64 guard target in cloud reload and s…
4d37342 
…creenshot tests

guardedCloudArgs carries the controller target as base64, but two unit tests
still parsed it as JSON. guardedCloudTarget decodes it beside the encoder.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@vonovak
fix(agent-cli): reconnect reused EAS development clients
dd96a07
@vonovak
fix(agent-cli): reconnect local device reloads to Metro
204a3fb
@vonovak
fix(agent-cli): reserve offline emulator ports
99b403e
@vonovak vonovak changed the title [agent-cli] Reconnect local device reloads to the selected Metro [agent-cli] Fix local reload connections and emulator port allocation Oct 12, 2026
@vonovak
vonovak force-pushed the codex/eas-session-binding branch 3 times, most recently from aa67ba1 to 09287d2 Compare October 12, 2026 20:20
@vonovak
vonovak force-pushed the codex/eas-session-binding branch from 09287d2 to 7cbeeae Compare October 19, 2026 21:08
@vonovak vonovak mentioned this pull request Oct 20, 2026
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@vonovak