Repository navigation
PLATFORM-TEMIZLIK: macOS/Windows platform verification from 148 failures to 2 (alpha.41 candidate) - #92
Merged
Conversation
… PLATFORM-TEMIZLIK
Owner 2026-10-15 ("Beta öncesi temizlik"): the nightly macOS/Windows
platform verification has been red since at least 2026-10-07 and runs
the whole suite on one runner (8–18 min per job, 20-min step bound).
scripts/platform-shard.mjs keeps every check of `npm run verify`: each
shard runs verify-context, build, the native test variants and its
quarter of Vitest; shard 1 also runs lint:checks, test:native,
test:host and smoke. Unlike verify, a failed test step no longer hides
the later checks; a failed build still stops the shard. The frozen
verify-job guard admits exactly these four edits.
The roadmap gets PLATFORM-TEMIZLIK (pre-beta) and the post-beta
WINDOWS-NATIVE gaps item. Astra 2578 P2 notes: ARCHITECTURE names SDK
2.3.1; dependencies.json names the 4.6.7 core zod.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…product gaps named PLATFORM-TEMIZLIK worker A. The nightly windows-latest verify (run 38793038837 at b63d92c) failed 118 tests in 37 files. Linux behaviour is unchanged: every touched file runs the same tests on Linux as before (394 = 392 + 2 new simulation tests, 1 existing skip). Fixture plumbing (the tests assumed a POSIX host): - tests/contracts/support/host-platform.ts: hostLayoutPlatform() (the mapping production passes to resolveProductLayout) and homeEnv() (HOME + USERPROFILE; the product reads USERPROFILE on native Windows); host-platform-fixture.test.ts simulates both for win32 on any host. - USERPROFILE beside HOME: query-error-disclosure, config-surface, terminal-delivery-window, init-policy-command, policy-upgrade-command, terminal-first-test-fixes, terminal-info-windows-workline (its "timed out waiting for budget choice" was the unresolved home), the instruction window CLI test. - Literal '/tmp' -> os.tmpdir(): source-build-bwrap, config-surface home. - PATH delimiter/dirname in the config discovery bound test; '/'-normalised source paths in worker-tree-gate-paths; the reader's canonical root (8.3 temp names) in the project instruction path assertions. - worker-tree-gate: no FIFO on win32, so the expected findings drop only the special-file row there (Linux expectation identical). - audit-checkpoint: the doctor/verify fail-open test registered no `terminal` config section and passed only after an earlier composed test (it failed alone on Linux too); it now registers the provider config itself. - binding-capability: hostBindingStrength() follows the product capture on every host (Windows reaches weak/location, as the doctor reports); installationBindingNotRunReason() keeps win32 not-run explicitly, so its consumers are unchanged. - project instructions: native Windows cannot verify a private trust directory; the fixtures open the reader the way the product does there (no directory) and assert that a directory is refused ('unsafe'). Platform skips (Linux never skips; bracketed reason in the title): - [requires Linux local runtime socket]: the two controlled-socket service tests (also failed on macOS: LOCAL_RUNTIME_UNSUPPORTED). - [requires POSIX ...] (win32 only): bubblewrap/Landlock view tests, chmod-0 unreadable entry, controlled POSIX Docker daemon (its stale win32 pre-check expected SUPERVISOR_OPTIONS_INVALID, which the supervisor no longer raises). - [unsupported on Windows: ...] product gaps: OS-user identity (backup, identity preview, admission actor), managed private ledger files, installation policy template (POSIX layout + OS-user identity), layout resource editor, project instruction trust cache, CLI JSON file input (no O_NOFOLLOW). Not changed: the Windows-only audit-checkpoint claim (product reads the OS identity before refusing POLICY_UNAVAILABLE) and the macOS backup BACKUP_IO (worker B). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
An untyped storage failure was mapped to BACKUP_IO and the original error was dropped, so the macOS platform-verify backup failures named no reason. The original error now travels only as the DeckentError cause; the code, the localized message and the audit event (code only) are unchanged, and no surface renders the cause. Vitest prints it as "Caused by", so the next hosted macOS run shows the real reason. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
… Linux with their reason run-create-approval-plan opens a live OS session (LocalOsSessionAuthority reads /proc) and worker-image-cancel runs the Linux process runner; on macOS both refuse by design (SESSION_REQUIRED, PROCESS_RUNNER_UNSUPPORTED). They now skip off Linux with a [requires Linux ...] tag instead of failing; Linux runs them unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
… CoreFoundation encoding variable
The helper is spawned with env {} (helper.ts); on macOS CoreFoundation inside
the child sets __CF_USER_TEXT_ENCODING at startup, so the child's process.env
was never empty there. Only on darwin, and only in CoreFoundation's exact
0x<uid>:0x<enc>:0x<region> form, that one key is set aside; every other key
still fails, and Linux keeps the exact empty-environment equality. The canary
check now runs on the raw helper output.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…xed sleep terminal-panels-pty (macOS, every run) and config-records-ink (macOS node 24, once on Windows) sent keys a fixed 60/50 ms apart. On a slow runner the child reads two keys at once: Ink keeps "\r\r" as one text event (no key), and a key can reach handlers before the previous key's state is drawn. Reproduced on Linux: the PTY failure with a CPU-starved child (old driver 3 fail, new driver pass) and the Ink failure with a 1 ms pause (old 7 fail, new 7 pass). The PTY driver now sends the next key once the child has drawn and gone quiet (bounded 1 s); '+' steps still look only at what their keys drew. The Ink press waits for a drawn frame (bounded 300 ms) and positive port answers wait via vi.waitFor; every assertion, including the negative ones, is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…n it does not exit On macOS the compiled MCP server answers over the FIFO but is still alive 10 s after its input closes (SIGTERM by the test timer); a Linux replay with process.platform forced to darwin exits in 16 ms, so no JS platform gate holds it. An unref'd 8.5 s timer preloaded into the child now writes getActiveResourcesInfo() and stdin end/destroy state to stderr, and the exit assertion shows that file, so the next hosted macOS run tells a FIFO EOF that never arrived from a product handle. A prompt exit is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…ORM-TEMIZLIK) BACKUP_IO now carries its cause (f81f419). These assertions compared only the code, so a macOS BACKUP_IO would still hide its reason. The code assertions are unchanged; the cause becomes the assertion message. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
… so backup, monitor, model calls, Brain and acceptance do not run there; green platform skips are not support Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…; macOS MCP proof over a pipe; host tooling not run on Windows Hosted platform run 38804553793 (try 1 of 4, owner budget): - macOS backup/restore failed with BACKUP_IO, whose preserved cause is LocalRuntimeSocketError LOCAL_RUNTIME_UNSUPPORTED: backup and restore coordinate with the runtime service over the Linux-only local socket. The 17 tests that failed on macOS now skip off Linux with that reason; the backup tests that passed on macOS keep running there. - macOS MCP identity proof: the server never saw the FIFO EOF (MCP_LIVE_HANDLES: PipeWrap only, stdinEnded false), a kqueue FIFO transport effect. Linux keeps the FIFO proof; macOS runs the same assertions over an ordinary pipe. Both paths pass on Linux. - Windows shard 1 ran test:host for the first time (verify used to stop before it). The .agents/refactor host tooling is Linux/WSL-only, so it is reported as not run on Windows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…orted policy custody refuses POLICY_UNAVAILABLE first
Owner 2026-10-15 ("Sırayı düzelt"). loadConfiguredScopeContext passed
readLocalOsIdentity() as an argument, so the identity was read before
config and policy. On native Windows that refused AUTHENTICATION_REQUIRED
where the existing audit-checkpoint proof requires POLICY_UNAVAILABLE
before any ledger/key/identity access. The local path now reads the
identity after the policy parses. The socket-peer path still verifies
the peer first, so an unauthenticated caller learns nothing about policy
availability. Linux full suite: 8547 passed, 1 failed (the known-env
installation-apply-adoption timeout), 8 skipped.
Known limits: macOS lacks the Linux-only local runtime socket, so the
service transport, backup and restore do not run there.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…s the pin gate; bounded backoff on fetches
Owner 2026-10-15 ("Kapıyı daralt"). In the try-1 dispatch the GitHub
advisories for containers/bubblewrap did not answer; the report counted
that as watch-incomplete and closed the integrity gate. Upstream
repository advisories ("GitHub advisories <repo>") and npm version
metadata ("npm registry") say nothing about the installed tree's
integrity, so they keep the report red but no longer block a verified
pin candidate. Incomplete registry signatures, OSV (which carries npm
malware advisories) and any future source still block (fail closed).
The token stays out of the watch job. Fetches get 3 attempts with 2 s
and 4 s waits instead of an immediate single retry.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…cision, not after a fixed 150 ms Windows node 24 failed "timed out waiting for late refusal of A" in the analysis run 38793038837 and in try 2 (38814829682). Card A expired 150 ms after it was requested, but the test needs that window for render + Ctrl+O + 40 ms + 'y'. On a slow runner the keys arrived after A expired, so A was never decided and no late refusal could come. A now expires only once its decision is in flight; the scenario (an old card's late refusal never touches the next card) is unchanged. Linux with keys slowed by 200 ms: the old version fails with the Windows signature, the new one passes. 12/12 pass three times as written. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…t bound on slow hosted runners Try 3 (38816414038) failed only on macOS node 24 with three multi-process timeouts. Two had already appeared in try 1: PURGE_CHILD_BARRIER_TIMEOUT (5 s) and CONCURRENT_SERVICE_TIMEOUT (3 s); the third was the 60 s allocation timing test. These waits ignored DECKENT_TEST_TIMEOUT_MS, which the platform cells raise to 120 s because of their 3–4 vCPU runners. hostBound() scales a fixed wait by that bound: unchanged at the 30 s default (Linux CI, local), ×4 on macOS and Windows. A hang still fails, only later. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
…(owner 2026-10-16); CHANGELOG alpha.41 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Owner 2026-10-15 "Beta öncesi temizlik"; trial budget 4 hosted runs; 2026-10-16 "Şimdi indir, 2 test açık".
scripts/platform-shard.mjs); ~12 min instead of ~18. Shard 1 runs lint/native/host/smoke. A failed test step no longer hides the later checks.[requires Linux …]. Windows product gaps skip only on win32 with[unsupported on Windows: …]. Fixtures are cross-platform: USERPROFILE, host layout platform, no literal /tmp. Timing fixes cover PTY key drivers, the approval R3 race, andhostBound()scaling of fixed waits. No test skips on Linux.Hosted tries: 38804553793 → 38814829682 → 38816414038 → 38818112622 (14/16 shards green). Two macOS timing tests stay open and are not skipped. Linux full suite on 5db6cc8: 8547 passed, 1 known-env failure.
🤖 Generated with Claude Code
https://claude.ai/code/session_017uL4HZTS1oaVHovPUKoRXP