Alice desktop request observations, 2026-11-04

The original Alice batch1681 ETIMEDOUT cause remains unknown. This change adds diagnostic evidence for a future authorized execution; injected controls do not reproduce or repair the original failure. Input diagnosis: af727325512faa8855efba86adefd6a11c1b4f9b and the retained timeout analysis.

Opt-in trace and interpretation

Set WINE_PROCESS_DESKTOP_OBSERVE to an existing, private, writable, local run directory, or pass observeDirectory to ProcessDesktop. Each main client, authority and bridge writes its own mode0600 fixed-size file. No logging, preload hashing, request ID generation or additional wire metadata occurs with observation disabled. The enabled trace is synchronous diagnostic I/O and can perturb timing; use a local filesystem, never a network mount.

Each file is 536576 bytes: three reserved 4096-byte slots for identity, first error and exit, followed by 128 rotating phase slots. Memory per write is bounded; there is no accumulated event array in a worker. A root desktop has three files; each additional client adds two. Use a fresh directory per bounded run; file count follows participant count, not poll count. Normal client close closes its trace descriptor; worker descriptors close on thread exit. Run files remain for evidence recovery. This is a bounded recent history, not a complete audit trail. Each JSON slot has a SHA256 checksum; the reader skips incomplete or checksum-mismatched slots during concurrent writes; missing/overwritten/torn records cannot prove non-execution. It returns sorted records for offline analysis and should not be called on every guest poll.

Records use process.hrtime.bigint() nanoseconds, PID, Node threadId, role, writer serial and a random client ID plus per-client request sequence and known operation name. An authority shared by several OS processes therefore retains unambiguous request correlation. No request/window/message payload, endpoint path, authentication token, raw execArgv or raw NODE_OPTIONS is recorded. The namespace token is also redacted from captured error stacks. Stack strings are bounded to2400 characters (further truncated if necessary for a slot).

Phase names are main-submission, bridge-receipt, socket-write, authority-receipt, handler-completion, reply-write, bridge-frame-receipt, shared-notification, main-return, timeout and request-error. Poll submission includes requested limit and local pending count; authority receipt/completion include queue before/after and returned message count. Timeout captures the two response words before terminating the bridge. Word0 is the response-ready flag; word1 can still be the previous reply length when word0 is zero. The words are separate atomic loads, not a coherent transactional snapshot if a response races the deadline.

Write markers follow socket.write() returning, meaning accepted into Node's socket write path, not kernel transmission or peer receipt. The notification marker follows the response stores and Atomics.notify; the main thread may wake before that marker is written. Only the normal response path acknowledges the request. Handler completion, queue removal, reply-write or trace presence never changes the return value, timeout, delivery status or unknown outcome. The3000ms timeout and fail-closed/no-retry policy are unchanged. A timed-out client cannot call again or reuse its response buffer. firstError and the reserved error slot preserve the first request error when observation is enabled.

Worker identity is written from the worker itself, before serving requests: Linux /proc/thread-self/stat TID (null elsewhere), Node version, hashes of inherited execArgv and NODE_OPTIONS, and a bounded loaded CommonJS module inventory (preloads, including entry/transitive modules). At most eight paths of200 characters are recorded; files up to1MiB have a SHA256, larger entries retain size with a null hash. Preserve the separately pinned launch/preload manifest for the complete inventory. An uncaughtExceptionMonitor writes the first worker error without handling it; worker exit writes its code when Node runs exit hooks. Forced Worker.terminate() may skip those hooks. Main records creation, eventual error and exit callbacks with the original Node threadId; those callbacks still require a main event-loop turn. An inherited preload can fail or block before this module starts: no worker identity in that case does not diagnose the cause. Retain inherited preload stderr and supervisor exit/error evidence as well as the trace.

Commands for the next authorized diagnostic execution

This task ran no Alice route and touched no remote box. A future execution requires coordinator RUNTIME RELEASE Alice timeout, a newly pinned input manifest containing this change and the external scheduling observer below. Do not resume the stopped original box or retry its unknown pending command. After provisioning the existing pinned local-entry/module-observer preloads, the launch from the runtime checkout is:

alice_diag_run="$PWD/scratch/runs/alice-desktop-next-diagnostic"
alice_preloads="$PWD/scratch/runs/20261103-alice-provisioned-dialogue"
mkdir -m 700 -p "$alice_diag_run/desktop"
ALICE_DIAG_DIR="$alice_diag_run" WINE_PROCESS_DESKTOP_OBSERVE="$alice_diag_run/desktop" \
node -r "$alice_preloads/module-observer.cjs" -r "$alice_preloads/local-entry.cjs" \
  test/run.js --app=alice_original_local --no-build \
  --wasm="$PWD/build/wine-assembly.wasm" --quiet-api --quiet-blocks \
  --no-threads --gl-renderer=software --frozen --control=8123 \
  --batch-size=100000 --max-seconds=1800 --trace-input \
  >"$alice_diag_run/original.log" 2>&1

Use the existing bounded supervisor/driver with unchanged first-error and unknown-command handling to launch that command; this bare command specifies the diagnostic environment and exact guest options, and does not supply the external sampler. Record the new PID, boot ID, launch/preload hashes, Git SHA and relevant binary diff before issuing input. The earlier path/app ID is a run-specific local registration, not a new public app registration.

Read current or recovered traces from another Node process, even while main waits synchronously:

node -e 'const d=require("./lib/process-desktop"); console.log(JSON.stringify(d.readProcessDesktopObservations(process.argv[1]),null,2))' \
  "$alice_diag_run/desktop" > "$alice_diag_run/desktop-phases.json"

The next runtime still needs an external supervisor sampling continuously from before worker creation through the wait/deadline, before normal cleanup removes workers. Keep a bounded recent window with monotonic and wall times:

A dead bridge with its worker-written error distinguishes a worker failure. Authority receipt without completion locates the last observed execution phase. Queue removal/reply-write without bridge receipt distinguishes the injected lost-reply case. None by itself proves an OS scheduling cause. A late marker requires the external scheduling/throttling deltas to explain delay. Pre-module preload failures still require supervisor evidence.

Focused validation and evidence

Run: scratch/runs/20261104-alice-desktop-phase-observation/. Small test command (no guest module, media or browser):

node test/test-process-desktop-observation.js
flock /tmp/wine-build.lock bash tools/build.sh

The dedicated test is automatically in the unit tier through tools/test-tiers.js; this checkout has no tools/test-tiers.json. The real-worker cases cover disabled observation, real publish/post/one-time poll transport, full phases and queue counts, checksum rejection of a parseable torn slot, bounded history wrap and first error retention, token/payload exclusion, terminated authority, dead bridge, and a real authority behind a reply-dropping socket proxy. The latter proves one destructive poll and a discarded reply with the consumed event while the caller still throws and rejects further calls. A real worker with an inherited preload exercises worker-written uncaught-error/exit and preload identity.

One attempted final fixture run exceeded the Linux Unix-socket path limit because the evidence path was long. Its original error is preserved in focused-path-limit-failure.log; the fixture now creates only its proxy socket under a short temporary directory and removes it afterwards. This was a test setup error, not a production desktop timeout.

Node v24.18.1 focused results are retained as focused-final.log; the required locked build is in build-checksum.log. input-head.txt, relevant.diff and publication.json record exact source provenance and publication. No source tarball or copied game media was made. The WAT/game integration test was not executed; source build validation is not a gameplay claim. No box was assigned to this worker. Stronghold's separately owned bx_pmfcmfw2 was reported RUNNING/idle at00:20Z and subsequently STOPPED by coordinator at00:21:20.638Z per board00:21:44 (original expiry01:12:18.560Z). This task neither queried nor changed its state and has no runtime lifecycle ownership to hand back.