Build hangs forever at "Planning" — xcodebuild pipe deadlock on stdout write(), amplified by stale MCP server instances

[WolfTheDesigner]

Summary

Local builds via XcodeBuildMCP (build_sim / build_device / test_sim, etc.) intermittently hang forever at the "Planning" phase and never reach compilation. The MCP tool call has no self-timeout, so the agent session wedges until the process is killed or the machine is restarted.

Root cause (stack-confirmed): pipe deadlock — xcodebuild blocks in write() on stdout while flooding destination-resolution warnings during planning. This is amplified by stale / multiple xcodebuildmcp MCP server processes left alive by previous agent sessions.

This is not a compile, SPM, or project-configuration failure. Healthy -describeAllArchivableProducts normally returns in ~2s; hung processes sat at 0% CPU for 15+ minutes with no progress.

Environment

• macOS (Apple Silicon)
• Xcode 26.x (observed with iPhoneOS 26.5 SDK)
• XcodeBuildMCP via npm exec xcodebuildmcp@latest mcp / Homebrew install (observed on 2.6.x; please confirm whether 2.7.0 changes this path)
• Multiple AI coding agents / terminal sessions starting independent MCP servers against the same project over a workday

Reproduction (intermittent)

1. Use XcodeBuildMCP from an agent client for device or simulator builds over several sessions (or leave old sessions open).
2. Observe multiple xcodebuildmcp / node … xcodebuildmcp mcp processes remaining alive from prior sessions (ps shows several start times).
3. Invoke build_device or build_sim (or equivalent CLI path that goes through the same executor).
4. Build log freezes at planning, often right after target listing / CreateBuildDescription / external tool probes (e.g. clang -v -E -dM …).
5. MCP tool call never returns.

Harder to reproduce cleanly on a fresh reboot with a single MCP instance; much easier when stale wrappers accumulate.

Evidence

1. Stack sample of deadlocked xcodebuild

Main thread blocked writing planning / destination-resolution log output to stdout:

_describeAllArchivableProductsAndExit()
 └ AllSchemesDescriptor.describe
   └ ProjectDescriptor._representativeDestinationInfos
     └ IDERunDestinationManager runDestinationsForScheme…
       └ …runDestinationsForScheme….cold.5
         └ DVTConsoleLogger logMessage
           └ fprintf → __swrite → __write_nocancel   ← BLOCKED writing to stdout

A secondary thread sits in DTDeviceKit startServiceBrowsers (physical-device Bonjour / device discovery), which is a plausible source of the destination-resolution warning flood that fills the pipe.

The hung process was effectively xcodebuild -describeAllArchivableProducts (or an equivalent planning path), sleeping for minutes with no child processes. A healthy run of that path returns in ~2s.

2. Frozen build log (same freeze point on retries)

Log freezes at planning after target listing, for example:

    Target 'App_App' in project 'App' (no dependencies)
    Target '… Notifications' in project '…' (no dependencies)
GatherProvisioningInputs
CreateBuildDescription
ExecuteExternalTool …/swiftc --version
ExecuteExternalTool …/actool --version --output-format xml1
ExecuteExternalTool …/clang -v -E -dM -arch arm64 -isysroot …/iPhoneOS….sdk -x objective-c -c /dev/null
   ← FROZEN HERE

Child clang -v -E -dM … capability probes can appear hung at 0% CPU. That is a symptom, not the root cause: probes stall because the parent xcodebuild main thread is blocked in write(). Killing only the probes does not clear the hang; the parent stays wedged. Re-launching freezes at the same line.

3. Live process snapshot pattern (multi-wrapper amplifier)

Typical bad state (times illustrative):

PID    STARTED     ELAPSED   %CPU  COMMAND
…      13:57       ~1:50     0.0   npm exec xcodebuildmcp@latest mcp     ← stale wrapper gen 1
…      13:57       ~1:50     0.0   node …/xcodebuildmcp mcp              ← stale wrapper gen 1
…      15:31       ~16 min   0.0   xcodebuild -project … -scheme …
                                     -destination generic/platform=iOS
                                     -collect-test-diagnostics never
                                     -derivedDataPath … build            ← deadlocked MCP-spawned build
…      15:32       ~15 min   0.0   npm exec xcodebuildmcp@latest mcp     ← wrapper gen 2
…      15:32       ~15 min   0.0   node …/xcodebuildmcp mcp              ← wrapper gen 2

Multiple MCP server generations coexist and compete over the same project / DerivedData / device discovery. Full Mac restart is the most reliable clear when several wrappers + a deadlocked xcodebuild have accumulated.

4. Workaround that avoids the hang

Running the same build from a terminal with an active drain does not deadlock:

xcodebuild … 2>&1 | tee /path/to/build.log

tee keeps reading the pipe. Direct CLI without a non-draining parent is also much safer than a wedged MCP capture path.

Other mitigations that reduce incidence:

• Kill stale xcodebuildmcp / related node processes before building.
• Prefer a single MCP client session / single server instance per machine when possible.
• Full reboot clears deadlocked xcodebuild + orphaned wrappers.

Root cause analysis (from runtime stacks + package source)

A. Pipe deadlock

During planning, xcodebuild enumerates schemes / run destinations and can flood stdout with destination-resolution warnings (device discovery / DTDeviceKit path). If the process that owns the stdout pipe stops draining (or drains too slowly relative to the flood), the kernel pipe buffer fills (~64 KB) and xcodebuild's write() blocks forever. Planning never completes.

B. How XcodeBuildMCP captures output (relevant implementation details)

In the installed package’s command executor (defaultExecutor):

• Spawns with stdio: ["ignore", "pipe", "pipe"].
• Attaches data listeners that:
  • append to in-memory strings (stdout += chunk / stderr += chunk) with no max size cap in the released code path we inspected;
  • forward chunks to pipeline handlers.
• Build pipeline log capture uses fs.writeSync per chunk (sync disk I/O on the event loop).
• There is no build-level timeout / watchdog that kills a stalled xcodebuild and returns an error to the MCP client.

If the Node process crashes mid-stream (e.g. huge output → string length limits — see also unmerged direction around large-output handling), freezes the event loop, or is one of several competing/orphaned MCP servers, the child can be left with an undrained or unhealthy pipe and block forever.

C. Amplifier: orphaned / multiple MCP servers

Separate issue class (orphaned MCP after parent agent exit, especially under npm exec) means multiple xcodebuildmcp processes stay alive for hours. That matches the multi-generation snapshot above and makes pipe + DerivedData + device-discovery contention much more likely. Idle-timeout on the MCP server (when enabled) does not help a hung in-flight build tool call, because that request is still considered active.

D. Misleading secondary symptom

Hung clang -v -E -dM probes look like known SwiftBuild planning deadlocks, but here parent write() blockage explains the freeze: children wait on a wedged parent.

Impact

• Agent tool calls (build_* / test_*) block indefinitely → whole coding session unusable.
• Users learn to avoid XcodeBuildMCP and fall back to raw xcodebuild | tee.
• System can accumulate multiple multi‑GB orphaned MCP servers + stuck xcodebuild / SWBBuildService processes until manual kill or reboot.

Proposed fixes

Ordered by leverage:

1. Build / exec watchdog (highest priority)
   If xcodebuild produces no new stdout/stderr (or no parsed stage progress) for N minutes, kill the process group and return a structured MCP error. Never leave the tool call open forever. Optionally make N configurable (related prior ask: custom build timeout).

2. Always-safe drain

   • Prefer async file stream (or OS-level redirect) for full logs instead of sync writeSync on every chunk.
   • Cap in-memory accumulation (maxOutputBytes / Buffer[] + truncate) so huge floods cannot crash the MCP process mid-read (related PR discussion around preventing RangeError: Invalid string length).
   • Consider piping planning-heavy discovery through a path that always drains even if the MCP client disconnects.

3. Single-instance / peer hygiene

   • Detect other live xcodebuildmcp mcp processes (lifecycle already samples peers / peer-count-high).
   • On start: warn hard, refuse second instance per user/workspace, or offer kill-stale policy.
   • Parent/ppid watchdog so npm exec wrappers don’t leave orphaned MCP servers after the agent dies.

4. Reduce the warning flood that fills the pipe

   • Prefer explicit destinations when known; avoid broad device destination enumeration when not needed.
   • Where safe, quiet / filter destination-resolution spam during planning.
   • Document that device builds with many paired devices increase flood risk.

5. Client-visible hang recovery

   • Document recommended client tool_timeout and ensure server-side kill still runs so orphans don’t outlive the client timeout.
   • Doctor / diagnostics: report multi-instance peers, long-running xcodebuild children, and “possible pipe stall” guidance.

Expected outcome

• A stalled planning/xcodebuild cannot wedge an agent forever.
• Multi-session days don’t leave N competing MCP servers + deadlocked builds.
• Even if Apple’s planning path floods stdout, the wrapper keeps draining (or times out cleanly).

Related upstream issues / PRs (for context, not duplicates)

• Orphaned MCP servers / lifecycle: MCP server processes orphaned when parent (Claude CLI) dies — 4GB memory leak #273, fix(mcp): Harden lifecycle teardown and observability #274, MCP server mode lacks idle timeout (daemon mode has it) #394 / feat(mcp): Add opt-in idle timeout shutdown #398 (idle timeout — doesn’t cancel active hung builds)
• Build timeout feature ask: [Feature]: build-sim custom timeout #174
• Large command output / string length: PR fix(utils): Prevent RangeError from large command output #433 (if still relevant)
• Concurrent sessions / DerivedData isolation: Isolate DerivedData per workspace to avoid conflicts with concurrent sessions #340

Happy to provide more stack samples, sample/spindump artifacts, or a minimal project repro if useful. I can also test patches against a multi-wrapper stress scenario.

[cameroncooke]

Thanks for the detailed report. I investigated this reproduction-first before changing code, but could not trigger the reported hang.

I repeated the following paths:

• v2.7.0 through the real command executor and build pipeline
• deterministic high-output stress through that same capture, parsing, and logging path
• the actual build_device flow and full build_sim calls over MCP
• published v2.6.2 with multiple concurrent npm exec wrapper/server generations
• paired A/B controls using the same generated xcodebuild command with 2>&1 | tee

All runs continued draining and returned normally. The tee control did not materially change liveness in this environment. Because the issue was not reproduced, I did not attempt a fix, and the working tree remains clean.

Could you share the smallest set of inputs needed to recreate the missing condition?

1. A minimal reproducible project or archive.
2. The exact MCP tool name and payload/parameters, plus macOS, Xcode/SDK, XcodeBuildMCP version/install method, launch command, and MCP client.
3. The exact generated xcodebuild invocation, if available.
4. Deterministic steps that produce the destination-resolution warning flood, or a complete captured build log and corresponding sample/spindump from a hung run.

That should let us verify the same failure rather than a nearby build, client-timeout, or DerivedData-contention issue.

[allenkaplan]

This happens on codex, many issues in the repo -- here is my cleanup script https://gist.github.com/allenkaplan/e475095658acbc0130b8b1aea7786f28

❯ codex-cleanup --help
Usage: codex-clean-zombies [options]

Inspect and optionally terminate stale Codex helper processes.

Default behavior is a dry run. Use --kill to send SIGTERM.

Options:
  --dry-run          Show matching processes without killing them (default)
  --kill            Terminate stale matching helper processes
  --force           After SIGTERM, SIGKILL helpers that are still alive
  --all             Include plugin/MCP helpers such as codegraph, lsp-daemon,
                    gcloud-mcp, and xcodebuildmcp
  --age MINUTES     Only target helpers at least this old (default: 30)
  --keep COUNT      Keep newest COUNT helpers per category (default: 2)
  --signal SIGNAL   Signal to send in --kill mode (default: TERM)
  --help            Show this help

Examples:
  codex-clean-zombies
  codex-clean-zombies --all
  codex-clean-zombies --all --age 60 --keep 3 --kill
  codex-clean-zombies --all --kill --force

Notes:
  Zombie processes cannot be killed directly. This command reports them, but
  only restarting Codex can make the parent process reap them.
  turn-ended helper processes are completion notifiers; once they are older
  than --age, they are always treated as stale.