Skip to content

chore(sync): preserve upstream history and repair WebKit consumers - #221

Merged
vincentkoc merged 338 commits into
mainfrom
claude/fix-aot-require-facade-20261102
Nov 15, 2026
Merged

vincentkoc merged 338 commits into
mainfrom
claude/fix-aot-require-facade-20261102

Conversation

@vincentkoc

@vincentkoc vincentkoc commented Nov 2, 2026 •

Copy link
Copy Markdown
Member

Scope and current gate

Integrate the upstream AOT rewrite while preserving fork behavior, signed history and the exact artifacts consumed by OpenClaw and Birdclaw. This includes hook-aware compiled modules and namespaces, writable/recoverable stdio, retained external runtime, cache persistence, TLS/HTTP ownership, filesystem behavior and their original regressions.

Landed with signed merge 9debcc4b. Main has the exact qualified tree; both parent histories are retained. All required native, shipping and consumer gates passed. Cluster regression evidence was handed upstream on the existing proposal.

  • PR head is signed 9c4e2b1f426f7e938c18ed2a7ace5d17d68c55fa, tree 6c2389ca16899c716070fae66ad103986a438147. Normal fast-forwards preserve the imported rewrite and every signed fork commit; no rebase, squash or force-push was used. The two new commits verify signatures. Synthetic PR merge fe6297f8aeb9fb8f9ef1d55a49d054528c68f90c has that exact tree and the expected base/head parents. The canonical manifest pin remains signed 68ff388d45fbc960bdf844225f9f128c9cdeaa23.

  • Current candidate gates — all passed: native qualification: Linux 461/461 reported results, macOS 465/465; cluster suites 43/42, unchanged VM leak, AOT suites 72/36, class guards, four engine IR cases and direct networking pass. Verified text artifacts: Linux SHA-256 2a2bac1357dab79a76035d03b0bb609045f5a4125a953564c61d026e2a3b32db; macOS SHA-256 594d33a44c7f7b927aea47615dde43fcb1535ca7864561a91ae2fea6e23f0bc3. Shipping qualification: all six builds and smoke jobs passed, each smoke report has 364 passing cases; Windows compatibility passed 44/44 per architecture plus 50/50 ARM64 owner-exit stress trials; manifest and both consumer jobs succeeded. OpenClaw passed all 22 phases, including package/standalone acceptance, with frozen source 7c26e8cccd19b74c3443fb3f0ed6fbf9d7d410a3. Its completed owning-job log was independently checked (SHA-256 39a5131d1c9578555c8207385cc7afd804678bc531fa59fe4f32d1052e37e7ab); the large native-output artifact remains remote. Birdclaw text report passed all 11 phases, preserved frozen source 59c9ee27f39238654efc93ff7862b27ad474c044, SHA-256 59c4fa1259fd6bd211ae16f3f926617f29d0756d06898ecb96256200a80351f3. Both consume the same admitted run artifact; its manifest SHA-256 is 3ac4de463647c6e5233b6003596db1ec6b79a3f7df9d5148806f837e7d671810. Rust, Miri and quick checks passed; Publish was skipped. No Bun release or permanent consumer repin. Earlier results below belong to prior candidates and are not relabeled as current-head proof.

  • Reviewed repairs: signed 07c1db79 lets online workers own IPC closure after server draining; a state === "none" fallback preserves disconnect for custom entries that never initialize cluster. Two event-driven regressions require final IPC during an active TCP connection and clean shutdown of a no-cluster entry. Calls before online retain existing semantics; the broader early-bootstrap issue is separate. Generic worker-owned teardown was already proposed by @robobun in upstream 43762 and 30548; this scoped port does not duplicate those PRs or adopt their broader changes. Signed 9c4e2b1f raises only the aggregate macOS step from 20 to 30 minutes after the measured interruption at file 444/461; all individual test deadlines remain unchanged. Fresh independent P0–P3 review completed clean after resolving the custom-entry finding; formatter and diff checks pass. The native macOS committed-test comparison is complete: on the admitted pre-fix binary, the active-connection test fails with connectedWhileDraining: false and finalMessage: false; the custom-entry control passes. On the repaired binary, both tests pass, exit 0. Bounded text artifact SHA-256 f215bca26a0f23c3b5085c7b3513674881d9cbe71ac26553646de759610f9426, report SHA-256 751337596cdd0d84a31ae6f57484c79c8ccb087742ed91007d9255cd79493451. Both binaries use the same pinned engine; source/tree and executable digests were verified. This is focused runtime proof, separate from the full native and shipping gates. No local native execution.

  • Prior a75d9fc1 candidate gates: native qualification: Linux succeeded. macOS attempt 1 had one 30-second VM leak-test deadline failure; its one explicit failed-job-only replay passed that same test unchanged in 28.58 seconds, but exposed a separate cluster worker exit 1 and later hit the aggregate 20-minute test-step limit at result 444/461. Retries are stopped. No results.json was produced by the interrupted replay, so it is incomplete proof. The original repaired class-guard fixture, both run-AOT suites (72/36 passes), all four engine IR cases and direct networking passed in the completed attempt-1 reports. Six-target shipping/consumer dry run succeeded: all six builds and smoke lanes (364 runtime cases per target, zero failures), all eight compiled-runtime steps on each Linux/macOS target, both 44-row Windows compatibility lanes, ARM64's 50 owner-exit stress trials, manifest, OpenClaw and Birdclaw. Publication was skipped. Rust checks, Lint, source-lints, Packages CI, bun-types and autofix.ci are green. Synthetic PR merge 55284cc25ad430f9bbc46a4923ea120dd4dea637 has exact candidate tree cc3029841eb0506d15c55fe583c8e6a457573b4d and expected base/head parents.

  • Preserved macOS deadline evidence: attempt-1 text report, 60,828 bytes, SHA-256 bc6c76368640540c03f0ceaf86d4f7c2a70c60e77e368a124b07f23ba40d69a0; the VM test timed out at 31,542 ms, without an RSS assertion result. Test workload (50,000 × 50KB), 3000MB limit and 30-second deadline are unchanged. Existing same-workload candidate/stock evidence completed in 16.28–24.57 seconds under the same limits; prior native head 4079136f passed this test. This supports one bounded replay, not a proven diagnosis or a timeout increase. The single replay's separate cluster and aggregate-budget failures are recorded next; no further blind replay is planned.

  • macOS replay disposition: job 132876713213 completed the unchanged VM leak test (1 pass, 0 fail, 28.58 seconds). test/js/node/cluster.test.ts:277–305 instead printed correct permission bits 66 but worker exit 1 with no signal; the parent exits 0 regardless, and the test currently discards captured child stderr. That is a worker-shutdown failure, not a chmod failure or a VM diagnosis. At 20 minutes the aggregate step was terminated during result 444/461. Attempt-2 report: 37,403 bytes, SHA-256 ce1e59f069a8c1ae90b6cbdeda8104210cebc007070f9cc0cfd8058772578428; job text-log SHA-256 38250c562ccf50884eac41df9d06464d3107ee5a666525ab8aae25a0399003d0. The completed secretless macOS probe reproduced 3/20 candidate worker exits with unhandled write EPIPE versus 0/20 on stable 1.4.4; every failing trial had correct permission bits. Text report artifact SHA-256 98d8178ccd2d28de3c87796424bc3d92d063fe0acdb6b08c01805165a55ba59e, report SHA-256 86889a04c921be8d2cb27d3961f9427aef349dfd45eab4a8ad5e384b550c6779; diagnostic-only, no shipping credit.

  • Same-run consumers passed: OpenClaw at frozen 7c26e8cccd19b74c3443fb3f0ed6fbf9d7d410a3 completed all 22 phases, including Bun-only smoke, SQLite, hooks/handoff, complete standalone build/runtime/version/help and final identity checks. The 65,298-byte text log binds expected source/Bun/WebKit inputs (SHA-256 dd51bc4c211cdf9bec8df0b75ed3d2b6ef6874ca683e5ade6e463042ccd09eec); its large smoke-output artifact remains remote. Birdclaw report, 22,383 bytes, SHA-256 a8720f4bb6f385c42a4e0caf91bcd4251f03064c13509e222ff58f8a6ebafe72, independently verifies all 11 phases and source preservation at 59c9ee27f39238654efc93ff7862b27ad474c044. Both owning jobs select the same immutable same-run release artifact and enforce admission; Birdclaw's read-back manifest SHA-256 is 883033f21037c42dd860353080f379a4e25460eb15c776fa6be21af228cd904a, ZIP SHA-256 3871a65b81818c7b3e996887345f3e0e6b5f512ccc26763de1315cf9e8555371, executable SHA-256 13f2651c977153711fd8b7162e9aa6138d436b4d2201a31aff162ff28cdbcbf0.

  • Earlier engine-qualified runtime candidate: dd694cf4c8226ffb24821b6184b1c15ac36c76b7 on the integration branch. It contains signed merge f46720ac of upstream 2596ac5d391d47804c93a42bdbe615804aa62387 and the retained fork series. Frozen fork base: d9be087429c21e563bb5ac29f8a11805e64b56a3; earlier import: 3f3a9ff87cdccf3d32d7ec6b648c664eddaae840.

  • Immutable engine published and verified: Canonical run 43567714470, attempt 1, passed all 27 jobs, including all 26 same-attempt prerequisites and publication. The exact-run native authorizer was independently audited and reviewed, then invoked once; its receipt and protected approval were independently read back. Release autobuild-3357f9d63e6dd87a53f2aade723482130a55c0cd was published at 2026-11-14 15:51:32 UTC and is immutable. All 26 public asset names, sizes and SHA-256 digests were verified against the manifest, SHA256SUMS and publication receipt; the hosted publisher downloaded and verified every public asset. Source 3357f9d63e6dd87a53f2aade723482130a55c0cd, source tree d9c63edaafc778bd3c4ae9d610559fac11338d03, 21 native archive lanes and 22 archives include matching AOT generators. Canonical manifest: 866,175 bytes, SHA-256 28f89c4b7172907a83987a512619c732bcd5cdbec4b4ebbe3265bb6b8da95303. Publication receipt: ZIP 1,538 bytes, SHA-256 8114e6c27dad14769afff0e02cccc6db2e69babb6b76df18b7cbbad3e823c671.

  • Previous canonical attempt: Canonical run 43545655436, attempt 1, completed at 2026-11-14 05:59:39 UTC on exact approved source 3357f9d63e6dd87a53f2aade723482130a55c0cd. Preflight, all 21 archive builds, Linux qualification and Windows ARM64 qualification passed (24/26 prerequisites). Darwin qualification built the exact candidate successfully, then recorded 54 passing result rows and one failure: the unchanged vm.SourceTextModule memory test exceeded its explicit 30-second test deadline. The report does not establish a memory-threshold failure or a root cause. Assembly and publication were skipped; no authorization receipt or deployment approval was issued. Raw Darwin report: 285,533 bytes, SHA-256 a6b06485643834ffb40026271998c50eb75de5e3325f43630fc19c7ddebc97e1. That attempt supplied no publication credit; the later successful canonical run above independently completed every prerequisite.

  • Focused deadline diagnosis: Probe 43565761256 verified the exact failed executable's digest, revision, engine, platform and architecture; both candidate and stable test harnesses loaded. It stopped before replay because the probe expected a CLI canary label in Bun.version, which reports 1.4.4. Setup report: 4,064 bytes, SHA-256 b3b82e9d9cb1fba7dfbe37263301e6255641b2a10912f64fb6ca34525c899390. Only that one expected string was corrected and independently reviewed for probe 43566151732, attempt 1, controller 2ccb1b020cfb53737a30446610eb9edd5480f93d, dispatched once at 2026-11-14 06:24:05 UTC. The raw target test passed once in 20.44 seconds with one expectation, but the surrounding replay could not produce a complete results report: the diagnostic's sanitized environment omitted GITHUB_RUN_ID and GITHUB_REPOSITORY, which the existing runner requires. Its raw log also preserves a watch cleanup error and a watch assertion failure; these are not waived. The probe stopped before isolated candidate/stable trials. Replay report: 37,393 bytes, SHA-256 031ba9a866e4cd35b477e8c822b61ed862c43041d8100024e1d132a24250115a. The complete reporting-owner metadata contract was audited. Probe 43567034837, attempt 1, controller ab303bb3e6322462d507784638869f5ae5939d59, was dispatched once at 2026-11-14 06:34:53 UTC with an explicit nonsecret job-metadata allowlist, a bounded preflight exercising the frozen owner's reporting getters, and the canonical owner's cleanup of the task-created process group. The original replay and two candidate/stable pairs retain all coverage and failure checks. The repaired probe completed successfully at 2026-11-14 06:39:51 UTC. Its complete original-order replay passed all 55 result rows for the frozen 53-file selection; the target passed in 19.70 seconds. The two fresh-process candidate trials passed in 16.28 and 22.75 seconds; current stable Bun 1.4.4 passed in 17.88 and 24.57 seconds. Every direct trial recorded one passing test, zero failures and one expectation. Complete diagnostic report: 47,514 bytes, SHA-256 2b1afbfa230a9b56249bab98a556e371cd943302d3da226f0690799231cb4cd4. The exact failed binary, unchanged 50,000-iteration workload, 3,000 MB memory assertion and explicit 30-second test deadline were preserved. This is a whole-runtime comparison, not engine-only attribution. The original deadline failure's cause remains unproven. Earlier incomplete reports remain incomplete; the passing diagnostic supplies recovery evidence and no qualification or publication credit.

  • Prior publication run stopped at 2026-11-13 10:42 UTC: WebKit publication run 43378184244 completed 24 prerequisites successfully: preflight, all 21 archive builds, Linux qualification and Windows ARM64 qualification. Native Darwin qualification exceeded the 115-minute job limit while building Bun, before tests or a completed candidate identity receipt; assembly and publication were skipped. Target source remains 3357f9d63e6dd87a53f2aade723482130a55c0cd. At that point the tag and release were absent. The previous successful dry run is prior-source evidence, not admission of these new artifacts. The single native diagnostic 43532280691 completed all 18 steps successfully using the exact retained macOS archive, unchanged DD candidate, build command, compiled-input inventory and overlay bytes. Its restored build receipt and identity-bound canonical gate record 53 files / 55 results and four AOT passes. The native build took 87m32s, with Rust bun_sema at 71m59.9s on the critical path, linking at 5m12s and dSYM at 3m11.8s. The original timeout's cause remains unproven. Diagnostic report: 884,831 bytes, SHA-256 50648284bc368c1ed6b2dcd504c44ab8bf68cdda4555e528f8fb4d9cbef464e6. Nested raw test reports were not exported; this diagnostic supplies retry justification and no publication credit. No timeout increase, test waiver or partial-run publication is accepted. The later successful canonical run above independently completed all 26 same-attempt prerequisites, its fresh authorization receipt, protected approval and immutable public readback.

  • CI-only follow-up: signed 5c68de14a3 recognizes the historical rewrite merge's subject and first body record while preserving actual-parent validation, ancestry and trailer precedence. It prevents 19,881 unchanged imported paths from being treated as fork changes. The existing real-Git CLI regression failed alone in red run 43388774817 (257 pass/1 fail), then all 258 source tests plus upstream-sync and asset-admission checks passed in green run 43389279954. No runtime source or WebKit pin changes. The later native run confirmed that real-history selection names upstream 2596ac5d and includes the restored compiled SDK alias regression.

  • Bun now commits the exact complete published WebKit manifest, byte for byte, in signed 68ff388d45, with one release note. git diff --check and the pinned Prettier 3.9.9 check passed; fresh local autoreview reported scoped-clean, no P0–P2 findings, with complete coverage. Historical diagnostic identities remain bound to their original engine and executable. No Bun release or permanent consumer pin has advanced; WebKit approval does not authorize Bun publication, deployment or consumer repins.

The existing PR headref claude/fix-aot-require-facade-20261102 was advanced by normal fast-forwards to signed 65be1ba3dd2a7f1c603b497e3c61ebfcd6ae2528, preserving all 182 imported/fork commits after the original PR head bca751591913221e79af2d2aec09088632b7f248. Frozen main base d9be087429c21e563bb5ac29f8a11805e64b56a3 remains an ancestor. No squash, rebase or force-push was used. Keep the PR draft until final native and shipping/consumer evidence is current.

Prior-head integration failures and repairs: Native Linux/Darwin qualification and six-target shipping/consumer dry run are terminal failures for PR head 65be1ba3dd2a7f1c603b497e3c61ebfcd6ae2528. PR-event builds used synthetic merge commit ba40bff90529d2da90d8f4a4686a664ade834042, whose tree exactly matches the signed head (edb139c14b98a8fea0dd90527673746bc8ce5b18). Both native binaries compiled. Linux tests have one failing row: the migrated Node class type-position test reports five by-name accesses instead of zero. Darwin tests have that same row plus the existing AOT inherited-file-descriptor case (ENXIO on descriptor 3). Source review traced the Darwin failure to startup close-on-exec marking without the Linux-only AOT restoration path; signed 49217763e5 extends that existing owner to macOS, with native runtime proof now complete in 43699734361: its Darwin build passed and 460/461 runtime file rows passed; the sole failing row is the unchanged class type-position assertion. test/cli/run/run-aot.test.ts passed all 72 cases/228 expectations, including ordinary/AOT handover of descriptors 3, 4 and 64 and bidirectional IPC. Startup CLOEXEC/spawn passed (168 pass, 24 skip), as did fs-open-cloexec and IPC siblings, four engine IR cases and direct net compatibility. Verified text report: 61,354-byte ZIP, SHA-256 55435c4273a422f76175b4811c11b2921d8142c3db51c226883ad29c8dfd7294; actual Bun revision d1e8ee9e11c1ae75a8c6f5f95f28e7476627f13d, engine 3357f9d63e6dd87a53f2aade723482130a55c0cd. This verifies the descriptor repair, not final integration readiness. Format 43696178364 and Lint 43696180085 passed on its signed descendant controller a35aa4709352dcb3653ec137c9001963856ec01d. Focused same-binary comparison 43696426661, attempt 1, completed successfully using the immutable failed Linux executable: run 43614132198, artifact 14121114774, actual revision ba40bff90529d2da90d8f4a4686a664ade834042. Controller f91b3765dd7856ecdc8edd0da48ca65e1acd2beb verified the artifact's source, digest and executable revision, then compared the current split fixture, the same fixture with a strict dependency config, and the original single-file literal from 05ac1e0fda2eef1f3890882f19223b4e7ee2d440. All three retain the same five answer() member causes: unknown-object, empty types, loose-type; all three ordinary/AOT pairs produce the exact expected output with empty stderr and exit 0. The complete ranked JSON causes and source positions are retained. Dependency strictness is not a fix, and no fixture or assertion has been weakened. Diagnostic text report: ZIP 10,999 bytes, SHA-256 080fa393ed1be5cf5ae752875dc99efcfb8ca04dd6a6ccd0af55334b287f66f8. The earlier 43696169627 stopped in an unrelated legacy trim diagnostic before this comparison; a reviewed one-condition routing repair separated the two. That first setup failure supplies no type proof. The hosted frontend probe 43697766997 at signed 19a17f83725753b43b13f37aaeeb46dc80a54c0a confirms a checked entry, both resolved Node parameter types, 77/71 listed members, and both runtime constructor declarations. Frontend IR probe 43698241247 at signed 82aa70dbe7b7eb275aa4f662e0ba72a3e3d6e114 retains both expected parameter layouts (undefined|null|L4 and undefined|null|L6). There is no observed annotation erasure or declaration-bridge failure. The compact handler's known 5*4<30 parameter-guard profitability rejection remains unchanged. Subsequent guarded echo probes established that the Node layouts are valid, in use and shape-testable; the remaining by-name counts include generic fallback arms inside Body. Personal inspection and independent audit identified a reporter-contract mismatch in f737cc1f8d: old foldSiteLevels favored nearer/non-generic observations, while new makeExplainedAccess takes the worst outcome/minimum cost across source-site copies, including those closed arms. Aggregate byName == 0 was therefore not a semantics-preserving migration of the old fast-path assertion. Final matched proof 43704383751 completed successfully on signed controller b3671d347d2e4506c5135d62e482d96ed9813a23, frozen Bun d1e8ee9e11c1ae75a8c6f5f95f28e7476627f13d, and diagnostic engine 6c1519bc0aca8e50384003d4c6e3e298437d93ad. The typed callback's public --explain=guards-json report names both IncomingMessage and ServerResponse; removing only the parameter annotations yields no callback guards. Successful fast arms emit GetSlot for url and NewClosure for end, while legitimate fallback arms retain the name/filled-slot checks. All four ordinary/AOT pairs preserve exact output, empty stderr and exit 0. Verified text report 14158357387: 94,286-byte ZIP, SHA-256 98335ecbd8300af798d6c40c087e2d8c29c504530ccbc22de849592e5795737b; 12 text files/845,568 uncompressed bytes, within the existing 100-file/64-MiB/seven-day cap. No native executable is downloaded locally. Signed 4079136f5f repairs the test migration using those source-attributed native guard names, adds the annotation-removal negative control, and keeps exact real HTTP/EventEmitter/Readable runtime parity. It explicitly protects declaration-to-native-layout recognition; it does not claim a universal no-fallback per-access cost. Fresh independent review is clean and formatting passes. No optimizer policy or runtime workaround was added. Signed caeca56fb9 already removed completed temporary frontend diagnostics from integration, retaining their commits in history. Native qualification 43705441067 completed successfully on signed integration head 4079136f5f5b790d5bfae1369ea4af19fd18bd02, published engine 3357 and PR-base selection: Linux build/tests, Darwin cross-build and Darwin native tests all passed. Verified reports contain 461 Linux and 465 Darwin runtime rows with zero failed rows; the permanent typed/untyped class fixture passes on both. Both also pass all 72 run-AOT cases/228 expectations, four engine IR cases and direct net compatibility. Text report digests: Linux 14158613481, SHA-256 d0869cf0859fedb278c3ad4d06932bba5c1e44e50fa8da88c5f32f2c995069f6; Darwin 14159183963, SHA-256 3907111688a1dc9c5f3365d965be7dee35f562a394b0317cdc6a6fe9add1fa06. This is prior-head proof for the promoted xwin-only follow-up; new exact-head CI remains required. The diagnostic engine does not qualify shipping. Selection correctly names upstream 2596ac5d391d47804c93a42bdbe615804aa62387; the restored compiled SDK alias passes all five cases and 15 expectations on both platforms, as do the four selected engine IR files. The selector's invalid-parent stderr is an expected negative fixture, not a test failure.

Prior-head shipping failure and approved repair: Linux x64/ARM64 and macOS x64/ARM64 built successfully. Both Windows targets failed all four provisioning attempts before compilation because xwin 0.9.0 does not recognize the current SDK catalog's forward-slash filenames. SDK 10.0.26100 and CRT 14.44.17.14 remain present; current catalog data reproduces the selector mismatch. The approved, reviewed and signed a75d9fc1e8 adopts xwin 0.10.0 with publisher-verified archive checksums and its released catalog-path fix, retaining both SDK/CRT pins. Formatting and whitespace checks pass; fresh independent review and the promotion audit are clean. The promoted-head matrix must prove provisioning, builds and consumer acceptance before landing. The same three pin changes are already merged upstream in 48986, 346b430dbe, authored by Dylan Conway with diagnosis/fix credit to @aphrody-dev; no duplicate upstream PR is needed. In the failed prior-head run, signing, package smoke, manifest assembly and both consumer jobs were skipped. There is no final shipping or consumer qualification credit. Release input tag remains empty and publication disabled. No duplicate manual matrix was dispatched. Superseded prior-head native 43613554388, shipping 43613554473, native 43613094500 and shipping 43613095676 remain prior-head evidence.

CI repair evidence: The first promoted pin head exposed an unused AOT HTTP helper in Lint 43613094519 and one Rust import reorder in Format 43613094514. Signed a38a0f049d deletes only the unreferenced, unexported waitsItsTurn helper; the canonical HTTP module and matching AOT header hash stay unchanged. Fresh autoreview is scoped-clean at P0–P2, plus independent source/history review. The pinned Prettier check and diff check passed. Signed 318cc248a0 applies the exact hosted formatter import reorder, preserving the existing cfg attachment. Neither changes HTTP or watcher behavior. Upstream still calls its own helper, so the dead-code cleanup belongs only to this fork. These failed reports remain recorded; only the new runs can establish the repaired checks.

Built-in type contracts: Lint 43613554352 passed lint, build/CI script types and the engine generator, then exposed five AOT built-in type errors. Signed 65be1ba3dd reuses the existing shared stream callback/error contract in AOT TTY and preserves the nullable HTTP-version comparison using ?? 0 in both texts, with the matching canonical SHA refreshed. Three files, six insertions and six deletions; no casts, narrowed request contract, changed TTY runtime or changed HTTP decision. Pinned formatting and diff checks passed. Fresh autoreview is scoped-clean at P0–P2, with independent HTTP owner review. Before promotion, exact signed-head preflight 43613915168 passed the complete Lint workflow: JavaScript lint, build/CI script typechecks, fork-engine schema generator, AOT-aware built-in module types and remaining source typechecks.

Remaining acceptance: both native lanes and the six-target, untagged publish=false shipping/consumer workflow must qualify the final manifest-bearing head and exact resulting artifacts. Prior-source or fixture-only proof below does not substitute. Preserve enforced floors, size/ICU ceilings, expected-failure assertions, review/landing gates and historical failed reports. Resolve or explicitly report remaining defects; do not turn a passing replay into a repaired original run.

Retained changes

  • CommonJS/ESM namespace identity, hook registration/alias/computed-call discovery, resolver/VM ownership, compiled-test main identity and Jest mock factories; compiler/binding ports and all four positive/negative engine IR regressions.
  • ChildProcess initialization after tracing publication, failed-spawn EOF/close, IPC/extra-fd identity; lazy memoized stdio sink acquisition preserves the original constructor exception and supports retry.
  • External bun run retains interpreter, builtin text and locales while honoring explicit removals and intentional JIT removal; Darwin shared-memory initialization remains covered.
  • TLS-only peer-certificate/trust boundaries preserve plain HTTP; HTTP socket/close ownership and public HTTPS validation; typed-array allocation/GC, startup diagnostics, data-URL MIME and SQLite publisher-export repairs.
  • Runtime-transpiler-owned cache writes, explicit cache activation and TypeScript stripping; first-write/second-acceptance regressions. TLS retention changes only the evidenced fixture GC boundary, not production TLS.
  • Canonical LLVM 23.1.1, correct Windows bun.exe target, shipping trim tables, artifact-bound consumers, Birdclaw checkout naming and restored extensionless external-SDK regression.
  • Diagnostic branches, rejected source cuts and withdrawn mock annotations are not promoted merely because their measurements improved.

Current paired engine/Bun evidence

WebKit dry run 43236817413 passed on de72d14dc8d937924f15514c5f0d86ae87631052 with exact Bun DD694 / tree b60f7b8898514913a8125bef65707996fcd88bff: 21 archive builds plus preflight, Linux, native Darwin/Windows and checksum assembly—26 successful jobs, publication skipped. Linux has 85 sync files/87 passing rows and 44 comparison files/46 passing candidate rows, nine cache cases, four AOT checks, 1,901 engine stress and 1,639 module rows. Darwin has 53 files/55 passing rows, four watcher suites and four AOT checks. Windows ARM64 passes startup/DFG/FFI/accounting in all five modes. The unchanged candidate-pass-baseline-observed-v1 policy retains historical HTTP/Intl baseline gaps: baseline_all_passed and all_passed remain false.

WebKit PR53 merged as bcb792d4; its earlier run failed Darwin compilation before runtime tests. DD694's watcher repair passed diagnostic replay. PR54 merged as de72d14dc8 with verified parents/tree/signatures. Docs-only PR55 merged as signed 3357f9d63e6dd87a53f2aade723482130a55c0cd, correcting release prose without changing machine inputs.

Report receipts, independently verified; native archives/execution remain hosted:

  • Earlier Linux 13913489036, ARM64 13914966404, Windows ARM64 13912809601 preserve their recorded source/gate identities. Merged-source Linux 13940087836: SHA-256 f35b74881e24ced929d4e808cdf9f4d5c390c5b12ca552ea52709332b4fb6e13.
  • Final dry-run reports: Linux 13966372044 / 6dc9b14a11831ff0f05023c19b00bb896a2035428fc4d0e5cda90f39915ba2cb; Darwin 13967693937 / bdcd0f5ecafbf9513e5116b5438c2eefac6abcfe3c4e5b67cabc7734d75d538b; Windows 13965365283 / ac6b01c4a34fbace77a8fec9cf998fc4c1b312b4c587726e469c475e22a87e78.
  • Assembly 13968358754: 10,577,683,272 bytes; ZIP SHA-256 18b9617a8aee134d1a4be938716fcba4847dd22b024a9b661d5c7017c6ea3815. Earlier source/system-ICU evidence remains historical; canonical Linux closes that shipping-ICU gap, not final Bun shipping admission.

Restored alias regression 3666c0447f preserves the compiled executable → disk external ESM → extensionless SDK subpath → runtime resolver boundary from PR208. Committed-test replay 43187688823 passes 1/0/0 on cd861bb9f5/651eeaca6e27; missing-hook control fails, ordinary output is 42, standalone output observes the postcompile disk edit as 43. Report 13929177885, SHA-256 9556a61608fae0f8dd02407869e757d0757b7fe296556e0b98f8084dce1e68ea.

Actual Execa 10.2.0 replay 43210256810 passes on Node 26.12.1 and cd861bb9f5/651eeaca6e27: failed-process handle, writable own streams, original spawn error and no child start. Report 13939649345, SHA-256 d48720cab6bce210d49c2d2b263d2036b8e1cd28b2195bf52babe9870decf4f3. PR216 acceptance is complete; closure waits for integration landing. Upstream44933 is the existing owner proposal; exact upstream-head proof remains separate.

Rewrite-era regressions and explicit limitations

Evidence here uses the recorded earlier executable/engine, principally 5cd0cffa4a/651eeaca6e2792378928033b1ae7ac66d6f91eab; it is not final-head shipping proof.

  • Native42658362697 builds/packages and passes Rust policy, written-position/AOT CLI and net (154 pass/8 platform skips). Grouped setup times out before selected execution. Exact-selection resume42666743338 uses the same executable/selection after frozen dependency setup: 43 files/45 rows pass, no missing rows. This closes that selection, not the earlier full fork delta.
  • Close repair399dcd747c schedules drains only for initialized listeners. Red42657924975: two failures, four deferred errors each. Green42661605049: 171 close/abort passes, then dependency setup stops. Resume42662475558 adds 614 distinct passes (listener generations, path/glob, HTTP/HTTPS and CJS/stdio), total 785 without repeated-case double-counting.
  • Receiver oracle42661839475: 176 attributed cases/runtime, all original visible-surface walks retained; candidate deferred errors 30→14 (12 borrowed-close plus four disposal errors removed). Fourteen known listen errors, 125 synchronous differences and two pending Node disposal Promises remain recorded limitations. Migration105eefc85f changes only justified observations, preserves Windows TTY bytes; strict no-update replay42663088458 passes 35/0/0. No wholesale snapshot waiver.
  • Source42664730416 and format42664735888 pass at a327122d35; correspondence retains all 16 ordinary/AOT pairs.
  • Stdio42670431146: three failures before/three passes after on fork-native aa024→5cd, original sentinel/no uncaught reporting. Recovery42796705564: stdout/stderr red before, all three streams green after 5cd→cd861; stdin passes both. Native42793326534 passes the stdio file; its existing ICU-minimum failure remains recorded. Full-stream AOT42883219981 proves restored stdout/stderr writes and sentinel identity in-place and standalone, with image/ELF checks.
  • Light-stream follow-up: the unforced probe never calls replaced fs.WriteStream when only light properties are read. The full-stream proof demands _writableState.objectMode; it does not repair/waive light-stream monkeypatch parity or alter compiler retention. Original false receipt retained.
  • Crash42788680638 and review replay42791245947 prove read-only-procfs red/green, writable control and four fatal-init/UV cases with unchanged arguments. Best-effort, process-local helper retains Android shell/filter behavior, no global kernel policy. Android execution and optional TTY recovery coverage remain gaps. Upstream allocates inside its constructor; fork recovery does not establish that same allocation defect upstream.

Matcher: stock removes110/98, initial fork109/93. Origin42607192709 identifies minimatch private names. Extractiona9371d61cb separates unchanged matching from traversal; paired42655481259 recovers arguments/date/encoded-text/string-builder families (109/97), outputs unchanged, no added family. Only source/signature is imported, not diagnostic engine/controllers. 75ab4916e9 explicitly budgets one required private-name feature and requires the four recovered families removed on Linux; old unadjusted110/98 is not claimed. Final42665211403 passes eight retention/runtime cases, both glob programs, unchanged20-table/two-error assertions. Existing AOT NODE_SKIP_FLAG_CHECK avoids absent embedded source metadata; no declared flag/assertion is skipped.

Access costs: Entry diagnostic42672262387 attempt2 uses native 11e3dd7f and logging-only engine 5e52309a; five parameter uses×four<30 body values rejects the guard. Typed layouts remain, output matches, costs are five by-name/zero tested/one free. Stock42790302528 with 72e01da7/5a61071f, identical literals/types, matches those costs, dependency12/0/2 and four rejoining guards. The unchanged zero-by-name assertion remains unsatisfied on that fork pair; no optimizer/schema policy change is made. This rules out a fork-specific regression for those inputs, not all compiler costs.

Upstream48354 is withdrawn: stock candidate42795224747 on ecebf1d016/5a61071f passes Rust/table/parity and six JSX/TSX cases but CLI is13/16, three older CJS failures introduced within that PR. Stock2596 already supplies the package hint for in-place concurrent imports. The fork's different ordinary parser justifies its local extension policy; the disputed 40940 runtime policy is not imported upstream. No further parser port or test waiver.

Historical source, shipping and TLS evidence

These older results are retained for attribution; later passes do not rewrite failed runs or qualify later sources.

Recorded source/surface Evidence and outcome
b254 Linux Native42396910724:341/342 files,4/4IR; only size fails. Source, lint/typecheck, format pass.
827 Rust/cache Four Rust jobs42393773828 pass Clippy/all-targets, Miri,lol-html,Mordant. Cache 03c red42397008173 is0/2; 827 green42397216902 and Darwin42397771969 are2pass/28assertions/0skip; classic fails before success marker, AOT loses first cache write.
TypeScript removal 827 counterfactual42396900891 passes8/9: intended failure strips after explicit Interpreter removal; b254 fixes this boundary.
ChildProcess Linux42293426539:29/29 records,117 cases/6skip/1todo; Darwin42294315113:24/24 records.
Retained runtime Red42295458249:external ESM exit70 with interpreter removed; repaired4ff covers nested CJS,SAB sharing,embedded entry and Linux policy.
Prior native 4ff42300565294:Darwin349/349+4IR;Linux340/341,size only. 037 Darwin42397898542:350files+4IR pass.
Prior shipping c4ca42311827376:six builds/two Windows compat pass; Linux/Winx64 smoke pass; Darwin clone budgets and WinARM TLS fixture block consumers. All eight AOT/trim/external steps pass on four supported compilers. 03742397790517:six builds/five smokes pass,Macx64 takeover timeout blocks admission.
Clone controls MacARM42391125997:647cases/8skip/nine suites; Macx6442391134988:348clone+SAB/Buffer pass before old retention failure.

037 takeover timeout: attempt1 has Macx64 client timeout5690.33ms (120pass/1fail),64MiB payload,five-second deadline,no ordering assertion failure; server sibling passes. Mac report13585048864 retains raw evidence. Windows offset report13584668990 records135.38ms against100ms,650pass/1fail,not wrong offset/hang; its one failed-job retry passes. Mac's one attempt3 retry fails again6704.43ms; manifest/consumers still skip. No further blind retry or deadline/payload relaxation.

Isolated42401637739 and receipt13584859186 pass original443.25ms,instrumented client/server and Node26,64MiB before attachment,one plain callback,handshake/two TLS callbacks. Full-file42402894358 reproduces untouched failure6657.54ms while instrumented121/Node119 pass. Quiet42403845102 passes both121-case Bun files and119Node; smoke-context42404521914 also passes preceding JIT/SQLite/AOT/clone/retention and later TLS workload. Buffered snapshots finish with exact bytes/callbacks. Observer/order effects leave phase unexplained; these passes do not establish a TLS production repair or repair original shipping.

Retention fixture: Heap/stopped-stack13583245642 finds an unrooted lexical environment reaching the dropped TLS cycle and an exact post-GC stack reference; retained pair is a rooted control. This is not proof the scanner consumed that word or a general JSC defect. Controlled42394776966 is old-red/new-green on identical binary. Only the first30 existing collections move into existing Immediates; startup/count/second loop/assertions unchanged. Actual fixture Macx6442395706599 and WinARM42395715841:10pass/0fail/one inherited platform skip each (Windows-only pipe/platform-excluded upgradeTLS). Current-source shipping proof remains separate.

Historical size investigations and rejected experiments

No ceiling has been raised. Earlier b254 is68,720,006 plain/26,305,926 trimmed against68,440,000/25,880,000. TypeScript repair is+64/−16,320 versus827;363,173 bytes of dispatcher reachability is not marginal savings.

  • Matched42308717420:3f/4d68,126,484/25,949,972 versus03c/S68,711,814/26,400,134. Only69,972trimmed overage is established upstream; source+engine are confounded.
  • Layout42398848764 and report13584890787 reconcile b25468,720,006:67,546,343 unchanged named native,30,883 changed native,3,774headers/gaps,479,550unmatched native,659,456new storage. Source function39bytes; image262,144code/376,832data,including252,735shared stubs. Live consumers and deduplicated engine data do not justify deletion.
  • First42402871582 fails compile on obsolete namespace APIs, no measurements. ABI-adapted42404040694 removes36obsolete lines via canonicala7c2bb61 plus recorded download-host routing only. All four programs print exact hello. Same-S upstream68,163,156/25,982,548 versus b25468,720,006/26,305,926:556,850/323,378 paired gap. ABI-adapted hello is not unchanged-upstream/module/source qualification or per-fix savings.
  • Reports42406427307 reproduce all four totals; AOT code/data equal262,144/376,832. Range42461120720 with passing lint/format proves659,456 new storage each; own-native overlap67,503,700→68,060,550 plain and25,323,092→25,646,470 trim. Entire gap is native ranges,including549headers/gaps. Compare only each image's own VAs/maps,union aliases/ICF/overlap; anonymous names and reachability are not marginal savings. Joint SourceProvider cut rejected because prelinked/bytecode consumers need it.

Accepted narrow syntax guard: fa54b248b66ed7eb48217ad759120502be489e97 assigns the external-source syntax probe to Interpreter; compiled loads return earlier and its generated throw stub avoids false syntax verdicts. Native42405779322 passes341/342+4IR,including explicit-Interpreter-removed aliased/computed hooks; size alone fails. Source, lint, format pass. Artifact42406613788 reproduces68,720,070/26,289,606: +64plain/−16,320trim. Maintainer explicitly accepted that measured tradeoff (cont), despite the pre-run no-increase rule; it does not waive either ceiling. Plain intentionally retains Interpreter. Remaining excess280,070/409,606 is historical fa54, not a final-rewrite measurement.

Rejected native-source helper: 99267c65 passes341/342+4IR and affected behavior but 42430664842 adds4096 to both (68,724,166/26,293,702),fails measurable trim/no-further-plain-increase. Signed 798cf1f0 reverts to exactfa54 tree. No promotion,waiver or new build of unchanged source.

Withdrawn mock annotations: aa639be8/090fafc3→be7f28fd, 42466260961, fixes MalformedRecords and passes342/345+4IR,normal mocks/Test-removed plugin. Sizes68,715,974/26,277,318 improve4096/12288 but still fail ceilings. Affected behavior failed acceptance; signed f052f580 removes seven Test/no-inline annotations,keeps selector/plugin coverage,restores fa54 runtime tree.
Original-fa54 replay42467905191 and report13611873200 reproduce four assertions (314pass/4fail/1skip): not attributed to experiment. Upstream9e76acc6 records stale AOT mock expectations. a68e966b corrects only those values/title; retains plain/AOT parity,requires,identity,keys,destructuring,ESM live bindings. Corrected42468364393/report13612253906:317pass/1fail/1skip,builtins fixed,old CJS remock namespace parity still exposed. Correction source/format and rollback source/format pass. These are historical results,not qualification of the later rewrite or a reason to reintroduce guards.

Transport and prior same-artifact consumers

API/native42466260961 downloads/verifies the unchanged LLVM23.1.1 x64 archive2,014,958,380bytes and pinned WebKit LTO-AOT809,104,824bytes; build consumes checksum-keyed cache. Asset/source regressions, lint/typecheck, format and reviews pass. Canonical extraction/checksums remain; API failure preserves normal recovery,integrity failure is fatal,warm cache skips prefetch. Step-level runner context fixes a rejected-before-start declaration. BuildKit43148076486 later verifies pinned zstd1.5.7/ICU79.1/WindowsASAN downloads after92-byte504 failures. Transport success is not shipping proof.

Immutable WebKitS / publication42249525393 passed27jobs,verified26assets/21native+generators; manifestSHA a3984642bc71ab89451ed5000e5efed46282020f216c0bc02f9dc9e45ee66c50. This remains Bun's committed old pin until3357 admission.

Frozen consumer owners: OpenClaw7c26e8cc canonical pin, Birdclaw59c9ee27 pin, website52b86196 installer. Website derives the selected OpenClaw version's pin through installer sync; discovery is not deployed-runtime/operator-override proof.

fa54 six-target42429807742 passes six builds/native TLS smokes,two Windows compat,manifest and all54JUnit suites,with inherited8Unix/10Windows skips pertarget. Four compiled-runtime reports pass AOT/trim/external-run; emptytag,publish=false,nothing published. This is new-source proof,not an explanation of old Mac timeout. Birdclaw report13599050162 passes11steps/source preservation on59c9:install,check,coverage,build,demo,runtime,packaging,identity. ZIP SHA 66cca2a25db9e1a07ab0b52d25ba86f43267ee6bc3bb2e77eb88e56b52cb5d74; admitted archive 4ab7beeb7ae551fed3a214bdfc66448c13bb26aaa3150d5801fb5aeccd5726cd; ELF 91a3f5036e5ff0f2a957e0f40338738f8901c21607a5c86c2db7f36104a1112e,141,421,864bytes; manifest 1a7e187e643b25778b959a82b309ad0098b3fb2b99c10a6dac2ae5e0473d49a2.

Two OpenClaw attempts lose runner communication,not product verdicts; second stops standalone-build before unchanged30-minute deadline. Instrumented42460225353 also shuts down midbuild:cgroup15,733,264,384bytes,host813,628KiB available,swap5,468KiB free,disk72.6GB; sampled OOM counters zero. This supports pressure,not proven OOM/Bun crash/standalone success;no final report or third blind retry. 628e3c21 moves only OpenClaw+matching diagnostic to existing32vcpu runner; Birdclaw unchanged,following capacity owner. Resource-corrected42463585662 passes samefa54 artifact/frozen consumer:all22phases,31hooks,2SQLite,zero fail/skip/todo,exact source/engine/ELF/manifest,sidecars,outside-checkout CLI. Standalone14:07.91,maxRSS20,681,776KiB;actualhost~124GiB,cgroup50,463,993,856,no swap/OOM sample. No fanout,permission or deadline relaxation.

Consumer artifact13611492095 is488,326,135bytes; only ZIP directory/selected entries16,522,022bytes were retrieved,CRC/content hashes verified—not the whole archive SHA. ReportSHA b303a0c6771a7effbe62955296f3589501394b753c32d6f37eaa8e4c8d84982c; standalone manifest ba15723aa36e5e19f46984ad841b5c1370ec4b68d56d87141fc3363bdb463478. It closes sameartifact consumer behavior,not the historically failed shipping run or a later source. Earlierfe7842294321154 passed its older OpenClaw31hook/SQLite/browser/descendant contract; Birdclaw2014pass/one package-root fixture failure was repaired by4ff checkout naming. No prior result admits new artifacts.

Contributor credit and queue disposition

Keep replacements credited; after this integration actually lands, link landed commits and close superseded stacks. Do not import their retired harnesses or restart settled scope.

  • PR216:Peter Steinberger's 03db8ba2 setters are retained; c09/ad1/a990 reset/ownership repairs and actual Execa acceptance above complete the replacement. Existing upstream44933 owns that generic proposal; no duplication.
  • PR193:Jason (Json), fuller-stack-dev, authored original work; steipete split it. Credited910 moves finalartifact clone/SAB/Buffer,31frozenhooks and descendant proof to canonical consumer owner,replacing competing harness/obsolete pin/tag-only/all-record-v15 claims.
  • PR208:GNU verifier 2f741a3d retains Peter's authorship. External ESM/CommonJS/SAB/runtime fixes credit Peter Steinberger and Jarred Sumner; upstream owns extensionless resolution. Restored regression above closes alias boundary. Old full-application benchmark campaign is outside replacement.
  • Earlier sync PR218 and diagnostic PR219 remain preserved until credited replacement landing. No cleanup by inference.

Upstream handoffs and independent proof gaps

  • Resolver 47958, Jest namespace 47966, constructed Module.require 47747.
  • Compiled-test main 47969 and maintainer disposition; retained external runtime 48041 and maintainer disposition:both closed unmerged,superseded by incoming rewrite;keep closed.
  • Explicit cache 48129:signed 9622e83a1f592c09954b5e84151392fcbaaa2ca3 on 7ca0042d3ba23ed60fd9cd8af1a1ce6ec3f8d074,initializer+classic regression only,clear independent/scoped review. Buildkite138349 needs upstream approval. Resolver/Jest jobs have same external approval boundary; OpenClaw admin cannot unblock them,no upstream-native pass claimed.
  • WebKit evaluator 998:closed unmerged after rewrite. TLS argument-order handoff, existing reconnect 38980, earlier namespace, Darwin hard-link, synchronous-hooks preserve original credit.
  • Upstream47942 merged2026-11-07 21:21:27UTC as2596; actual engine pin is5a61071f79134bf41844e7ea31a8449147d85dd4. Compiled-main owner supersedes closed47969's runtime repair with original parity test retained; exact new-upstream hosted proof remains separate. Closed48041's retained-run/ESM/SAB/nested-CJS/JIT-removal issue is not established fixed by source:reproduce on rewritten main;new PR only if it persists,as requested.
  • Stdio48352 and crash-wrapper48339 were checked at 2026-11-15 09:25 UTC: both remain open and ready, with upstream review required, no new submitted review or inline feedback, and no actionable CodeRabbit comments. Live signed-series descendants are a95d5da807dbbbd0994603e253f15a7158308b12 for stdio (Buildkite141706, pending) and 23344f0335ccf25e1ea194c543ce3fd9fc1784d9 for the crash wrapper (Buildkite141683, pending). Their new tips add Vincent-authored regression tests only, preserving the previously reviewed production commits. Addressed review threads remain resolved; optional TTY is deferred and Android unexecuted. These external upstream proof gaps do not authorize changes to active fork policy.
  • Withdrawn48354 and its13/16 candidate proof remain above; no disputed runtime parser expansion.

Promotion and authority boundary

  1. Completed: approved3357 publication, immutable public verification and signed exact complete manifest pin; imported history retained through normal fast-forwards.
  2. Completed: native regression repairs and the approved xwin 0.10.0 catalog repair are reviewed, signed and fast-forwarded to the existing PR head without force/rebase.
  3. Six untagged publish=false shipping targets and both consumers are complete. Resolve the observed native macOS cluster teardown and aggregate-budget failures while preserving assertions, coverage, review gates and the original failed evidence.
  4. Land with merge commit,retain imported/signed history,then retire credited replacement PRs.
  5. **Separately authorized only:**Bun publication,permanent consumer repins and live deployment. OpenClaw integration also needs paired CI/Bun-only smoke,native macOS two-binary checks and matching-host staging. Green builds do not authorize restarts/deployment.

vincentkoc and others added 24 commits November 2, 2026 08:17
Keep bare plugin loading separate from filelike resolution and preserve lstat errors. Resolve Darwin paths through the canonical owner with its hard-link guard and ENOTDIR realpath fallback. Treat compiled namespaces like interpreted namespaces when CommonJS forwards their defaults.

Preserve raw native-module mocks and align namespace, cycle and UTF-16 cache regressions with the paired engine. The real AOT forwarding case fails before the C++ adapter and passes after it through bun bd. Full Debug suites report 1201 passes and 12 timeouts under investigation; this commit does not claim full qualification.
Keep the kernel ENOTDIR result when the libc realpath fallback cannot run because process or system descriptors are exhausted. Successful fallback results and other errors retain their behavior, including Darwin trailing and intermediate path resolution.

The unchanged native descriptor-exhaustion fixture reproduced file/below returning EMFILE before this repair. The preceding run passed the Node path oracle but had five POSIX lock-test timeouts after successful loop iterations. Fresh native qualification of this change remains pending; no passing full-suite result is claimed.
### What does this PR do?

Both Prisma guides break at step 3 today. The `latest` dist-tag of
`prisma` now points to Prisma ORM 8 (`8.0.0-rc.24`), while
`@prisma/client` and the driver adapters are still on `7.10.1`. So `bun
add -d prisma` installs a Prisma ORM 8 CLI next to a Prisma ORM 7
client, and `bunx --bun prisma init --db` fails with `No flag registered
for --db` (`--datasource-provider` fails the same way in the SQLite
guide).

This rewrites both guides for Prisma ORM 8, which has no `prisma init`,
`prisma migrate dev`, `prisma generate` or `@prisma/client`:

- **`prisma.mdx` (Use Prisma with Bun):** `bun init`, then `bunx
prisma@latest orm init`, which installs the packages with `bun add` and
writes the schema, the client and `.env`. Then `bunx prisma db init`, a
script that creates a user and counts the users, and three runs that
print 1, 2 and 3. `orm init` targets PostgreSQL or MongoDB, so the
SQLite file of the old guide becomes a PostgreSQL connection string,
with `bunx create-db` as the no-account option.
- **`prisma-postgres.mdx` (Use Prisma Postgres with Bun):** `bun create
prisma@latest` with the `minimal` template and Bun as the package
manager, `bunx create-db` for the database, `bun run --env-file=.env
db:init`, the same test script, then a `Bun.serve` server for `GET
/users` and a deploy of that server to Prisma Compute with the
scaffold's `deploy` script.
- **Both guides** use `sslmode=verify-full`, because with `require` the
`pg` driver prints a security warning on every connection.
- The Prisma Postgres guide notes that `bun run` does not pass `.env` to
`package.json` scripts without `--env-file`, and links the environment
variables page. The old note about dynamically loaded subcommands
needing npm is gone, since `orm init` installs with `bun add`.

### How did you verify your code works?

I ran both guides end to end as written, with Bun 1.4.3, in fresh
directories, against databases created with `bunx create-db`:

- **Use Prisma with Bun:** `bun init`, `orm init` (installs
`prisma@8.0.0-rc.24` and `@prisma/orm-postgres@8.0.0-rc.18`), `db init`
printed the output shown in the guide, and three runs of `bun index.ts`
printed 1, 2 and 3 users with no warnings. `bun check` passes.
- **Use Prisma Postgres with Bun:** `bun create prisma@latest`
(create-prisma 0.14.3), `contract:emit`, `db:init` with
`--env-file=.env`, three script runs, `curl localhost:3000/users`
returned the users from `Bun.serve`, and `bun run deploy` put the server
live on Prisma Compute, where `/users` returns `[]` from the freshly
provisioned database. `tsc --noEmit` passes once `@types/bun` is added
as the guide says.
- Before the change, I reproduced the `No flag registered for --db`
failure with the current guide.
- `prettier --check` passes on both files, and I opened every link in
the changed lines.

Disclosure: I work at Prisma.
@autofix-troubleshooter

Copy link
Copy Markdown

Hi! I'm the autofix logoautofix.ci troubleshooter bot.

It looks like you correctly set up a CI job that uses the autofix.ci GitHub Action, but the autofix.ci GitHub App has not been installed for this repository. This means that autofix.ci unfortunately does not have the permissions to fix this pull request. If you are the repository owner, please install the app and then restart the CI workflow! 😃

Match the catalog-compatible pins from oven-sh#48986 while retaining the Windows SDK and CRT versions. Upstream credits @aphrody-dev for the diagnosis and fix.
Let initialized workers close their IPC channel after their servers drain. Preserve the primary-close fallback for entries that never load cluster and therefore never install the disconnect receiver.

The generic worker-owned teardown is already proposed in oven-sh#43762 and oven-sh#30548 by @robobun. Keep the broader early-bootstrap changes separate.

Add distinct active-connection and uninitialized-entry regressions. Hosted runtime validation follows this signed source commit; no local native execution.
Increase only the aggregate native test step from 20 to 30 minutes after a measured run was killed at file 444 of 461. Preserve all per-file and individual test deadlines. Reviewed together with the cluster drain repair; hosted validation follows.
@vincentkoc
vincentkoc marked this pull request as ready for review November 15, 2026 18:17
@vincentkoc
vincentkoc merged commit 9debcc4 into main Nov 15, 2026
40 checks passed
@vincentkoc
vincentkoc deleted the claude/fix-aot-require-facade-20261102 branch November 15, 2026 18:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

7 participants