Skip to content

Support identity-verified daemon bootstrap on macOS - #260

Merged
lossyrob merged 11 commits into
mainfrom
claude/issue-259-a2d3b8
Nov 15, 2026
Merged

lossyrob merged 11 commits into
mainfrom
claude/issue-259-a2d3b8

Conversation

@lossyrob

@lossyrob lossyrob commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

Closes #259

Current disposition: source clear, native HOLD (diagnostic-only)

Actual pushed head 1fed64e687a5d9bbae6267f6329f2fb217bdff94, tree 49d3b31173318ee0cc254f526ad6f3b8ff3258b8, parent 8eb83f774f349e97e1ed2b0384c9d6efd2cd7edc. Parent8eb is an ordinary merge of af42 and current main 9facc00444852c324c315d371dbdef6e5a5a1b66, tree 73e61c71ee09438d82e16124588fb43445597f66. Incoming6a01562e/215da490, d71 integration and all repair history are preserved. No rebase/force push.

The new two-file delta (+92/-20) adds precise macOS finalization diagnostics, NOT a behavior fix. It records thread-correlated owned-state/refusal phase, raw group-KILL return and immediately captured errno, signal seal versus generation, skipped-reap versus reap-stage failure, and exact consuming-wait outcome/status. Signal failure STILL skips reap: splitting the previous short-circuit expression preserves its order and failures. No ESRCH guess/exception, retry, containment bypass, cleanup waiver, deadline extension, parked-slot action or broader post-CONT handling.

App accepts this exact candidate's unconditional macOS library stderr as bounded TEMPORARY diagnostics, not permanent consumer-output approval; af marks it LOW/nonblocking, not a contract deviation. No logging change is requested now. AFTER the native cause is known and BEFORE final product/source acceptance, explicitly decide retention/gating/removal while preserving raw evidence; routine stderr must not silently become a public consumer contract.

ThreadId is PROCESSLOCAL, not globally unique across nested fixture processes: correlate with enclosing role/process context. Synchronous diagnostics can affect timing/interleaving; source conservation is not zero-overhead or hard-deadline proof for kernel/diagnostic sinks.

All three retained exact1fed SOURCE dispositions are clear: bc SOURCE CONSERVATION PASS/zero required findings, af CONFORMS/no deviation, App REPORTING SOURCE CLEAR/no schema change. bc also conserves ALL21 accepted RC2 import payloads, strict Telex-owned APIs and six unchanged identity/proof dependencies. Owner fully read the11364-byte bc report/SHA88a881d9fda6f6a7c62cf6978ed7c88cf6f03787e8f0c90aa4e891e876b14209 and later publication addendum2011/SHAda9f45184357aa01fa438b03e8b179b3122941dcffd0977b59d870a7ba0af367, independently rehashed21+2 bound files; af/App dispositions are Local-attributed. These are bounded SOURCE findings, not native/product/final output-policy acceptance.

Native adapter source is exact af42; no public error/enum/receipt shape or private proof seam changes. Current-main integration brings the already accepted RC2 changes/version0.3.0-rc.2, not a new bridge trust vote or release action. The existing disposition is updated in place; earlier public captures/packets remain frozen.

Actual1fed required Mac daemon job132090138558/CI43484755946/a1 FAILED. SAMEbc's final frozen classification14025bytes/SHAaa4a74ba23775fcc6cba88320887532f6d08e151796eae898ccd013a3fd62488 binds Campaign's supplied existing full joblog78863bytes/913lines/SHAec9bbeb9c2cfeef4d256f9e6f22d406a44a38e018e9f6f3c9e4d7fa91ad11a38. Owner FULL-read report/FollowupHashes1042/SHA523e05638e655604d12ebaa65880b0e6ffc943f89da91bfec661ae2178b4e240 and rehashedALL4, no duplicate trace. This job emits exact1fed/tree49d3/SDK26.5, not old-head carry. In enclosing FIRST deadline-custody role/ThreadId2, NORMAL control logs Stopped726-727, Exitedtrue728-729, groupKILL=-1/immediateerrnoSome(1)/postattemptownershipSealed/generation_known=true730, phase=signal/reap=not-attempted731, Bquarantine732 and normalcontrolpanic735. Actual SIGNAL-stage failure/rawerrno1 at branches1083-1104/1806-1813, NOT ESRCH or consuming-wait failure. WHY errno1/kernel/host/group-membership remains unestablished; no groupabsence/containment proof or cleanup waiver. Guarded semantic disposition must precede any behavior fix. SOURCE PASS remains separate from native HOLD.

Separate replacement ThreadId12 logs the same signal-stage result/skipreap801-806. Distinct successful refusal-path kill/reap770-784 and descendant ThreadId5 kill/reap756-760 do NOT prove normal-control cleanup; enclosing role/process context matters. Library37pass/3fail/5ignore. Forced custody/reuse/ECHILD/EINTR after the normal control and downstream process/socket/genuineMacwriter/A3 are not reached. No af42/f535 cause is carried.

Initial1fed watch is retired/52d cleared by sole2a; Campaign is separately assigning SAME-attempt terminal/artifact diagnostic continuation within ORIGINAL20:10Z, not yet confirmed live in this receipt. Overall1fed CI/archive/official-writer/A3 outcomes remain unknown. The full provided job log is NOT independently attested raw admission.log; actual artifact/run/job/byte-length/SHA still needed. Campaign retains the final fresh current-main/CI/merge gate. The af42 watch stays RETIRED on main drift, with unfinished af42 outcomes unknown. The independently admitted merged-RC2 shared scheduler is untouched. No owner duplicate fetch/reproduction/retry/new observer.

Compatibility and authority contract

Enable identity-verified InstalledCurrent/ExactExecutable macOS bootstrap, including the reachable missing-admission manifest produced when the OFFICIAL v0.2.1 updater installs a 0.3.x candidate. Refusing that eligible shape or rewriting its immutable manifest is not compatibility.

macOS uses the accepted WEAKER (dev, ino) identity of the lowest-addressed file-backed PROC_PIDREGIONPATHINFO mapping: regular vnode, offset0, readable/executable and nonwritable. Unexpected first mappings refuse rather than being skipped. Reopening a pathname is not mapped identity; renaming/replacing/unlinking a path does not change its mapped vnode. This is NOT Linux exec-link equivalence, a content digest/signature or intra-user isolation. Same-user malicious code is excluded; nonzero-offset executable slices/incompatible mapping layouts remain unsupported packaging. The O_NOFOLLOW witness prevents inode reuse while held, not path replacement; canonical-path spawn detects replacement rather than preventing it.

UID, canonical process path, PID/start pin, admission-backed capability evidence and exact HelloAck package/build/version/schema/protocol/capability fences remain. Child-side mapped-vnode proof precedes readiness. Unsupported/indeterminate identity fails closed; other Unix platforms remain unsupported. No helper, privilege or entitlement is added.

The owned raw posix_spawn (not spawnp) starts suspended in a new group with explicit descriptor actions/CLOEXEC-default/reset signal mask/defaults and the existing stripped environment allowlist. It runs non-serving --json version --root. Both pre-authentication AND fresh pre-CONT require owned Stopped(0 or SIGSTOP), never Running; selected witness/mapped-vnode/start checks remain, then only the anchored direct-child PID resumes. Darwin requests WSTOPPED, not WCONTINUED; trapped/continued/unsupported-stop/PID mismatch/unknown refuse. Linux options/results and Windows decisions remain unchanged.

One original10-second budget bounds verification/collection/ATTEMPTED cleanup, not guaranteed reap or cancellation of blocking kernel calls. A synchronized slot, SAME opened F_DUPFD_CLOEXEC witness and native PID storage are armed BEFORE successful spawn. During the original active attempt, still-known owned stopped permits one cleanup group KILL for ordinary wrong-vnode/expiry refusal. Signal seal is NOT generation loss: changed generation/observation error forbids later numeric signal/reap. Stopped remains pending, never terminal/success; only terminal WNOWAIT evidence permits one consuming terminal wait. A terminal leader alone is not containment or successful cleanup.

B quarantine: ANY parked child-bearing record permanently refuses native legacy admission across roots until CLIENT restart, with NO later PID observation/signal/reap/release, background cleanup, clock extension or selector-lock retention. Even apparent exit/matching start/stale ownership flags cannot release it. Strict admitted manifests/cached same-key completion remain outside this gate. Only normal ACTIVE original-child confirmed terminal reap clears a child-bearing slot. Matching-start/WNOWAIT/matching-start alternative A remains rejected because of its final consuming-wait race. The real competing-reap/different-real-child oracle simulates numeric reuse; it is not an OS allocator reuse trace or native passing receipt.

Exclusive admission spans probing and final root/selector/raw-manifest/target validation; validation errors propagate before equality. Completion requires exact metadata, writes no manifest, releases exclusive admission and revalidates under fresh shared admission. Present-invalid keys, non-0.3.x targets and failed probes still refuse. A NEW admission-aware immutable-manifest reselection is remediation for ineligible selections, not genuine-v021 compatibility proof or a cure for unresolved CLIENT custody. Public enum/display strings are unchanged; unconfirmed cleanup retains incompatible-manifest precedence and primary context.

Proof route remains mandatory

The native job builds exact head and records source/tree/OS/architecture/SDK. It runs primitive, suspended-image/state/start/failure/custody coverage, complete installed-current process/synthetic compatibility and actual composed-socket length/bind under a unique short private base.

Any future diagnostic packet using these lines must cite raw admission.log by ACTUAL artifact/run/job and independently retain its byte length/SHA256. report.json or a job label alone does not attest the raw lines. This is the unchanged capture route, not a new receipt schema or native-run request; no such current-head raw artifact is credited here.

The genuine proof pins/verifies the matching OFFICIAL v0.2.1 archive from commit212b76a4, invokes THAT updater to install/select exact candidate, and verifies raw admission-key absence, exact package/build/schema/protocol/capability metadata, installed bytes/hash/identity and immutable before/after root/selector/target/raw-manifest bindings. Windows retains real directory identity via BACKUP_SEMANTICS.

The EXISTING public run_watcher_probe must use that SAME genuine-writer root: actual SendOnly send, stable logical store/fresh runtime, registered responsibility/store/runtime/capability/health assertions, sender-ready, no receive/deaf state and zero inbound pending/actionable work. Schema2 reporting retains exact source/tree/host/run/artifact, actual argv/environment KEY NAMES ONLY/exits/timing/byte counts, command stdout/stderr and hash references, all before/after bindings, release/archive/sidecar/extracted/candidate build/digests, stdout reference/hash, real message/durable/occupied/replay/recovery and BOTH nonempty store/runtime tuples. No credential environment values, new harness/private Application Client seam, fallback, NEW manifest or attach-only substitute. attach.ready alone is not A3.

Evidence and limits

Exact committed1fed Windows bootstrap36pass/0fail/2roleignore; Darwin Postgres tests cross-type-check and production-lib clippy pass with warnings denied, on Windows ONLY. No local native Mac/Intel or SQLite-only Mac genuine-writer/report integration compilation/execution. Historical broader Darwin test-clippy failure is NOT green; no suppression/baseline run/unrelated repair.

Retained bc actual-af42 SOURCE PASS/S1 SOURCE RESOLVED by B, af CONFORMS and App reporting SOURCE CLEAR are HISTORICAL source dispositions, not successor approval. All four af42 reporting blobs were unchanged; no reporting fix was requested. The initial reviewed4417-byte memo/SHA68b171f8 PLUS explicit C1-C5, NOT later UNREVIEWED4418-byte/SHAaec63409 copy, remains the plan anchor. Pinned upstream source/conditional kernel-mask explanation is not shipping-kernel configuration or f122 cause proof.

Actual af42 Mac job132065388872/run43477596181/a1 FAILED: primitive2pass/1roleignore, bootstrap37pass/3fail/5roleignore, exit101. Its exact checkout/tree and SDK26.5 are emitted. Code5/status0 is accepted as Ok(Stopped), options0x2d; code1/status0 Exited=true is observed. Its initial NORMAL authenticated-probe control reaches ordinary post-terminal finalization, then unconfirmed cleanup/permanent quarantine. Group-kill return/errno and exact failed cleanup substage are UNKNOWN, not old f535 Stopped rejection or guessed ESRCH. Expected-success replacement also reaches finalization before incompatible manifest; later expiry refusal follows parked custody, not proved OS spawn failure. State-table, descendant-held-pipe deadline and automatic-reaping-policy tests actually PASS only their bounded surfaces. Later output-cap and forced deadline/ECHILD/EINTR/reused-custody roles are NOT reached; complete process/socket/genuine-updater/A3 stages are SKIPPED. Final supplied bc classification11800/SHAf5a3c18965820d47be5c2d6d414f576d0a404e51f698fc2e7719d380d2413ac8 is separately frozen from its source PASS.

Historical f535 S1 SOURCE HOLD/opportunistic-reap conformity withdrawal is not retroactively cleared. f535 CI43470584011/a1 ended21success/1knownMacfailure/1cancelled Watcher Mac diagnostic; archive43470584067/a1 SUCCESS3/3. Its Mac bootstrap35/4fail/5ignore rejected code5/status0 BEFORE downstream proof; concurrent diagnostics do not identify a unique originating test or complete cascade. No S1 reproduction/vnode/socket/API-impossibility cause follows. The ACTUAL Windows official-v0.2.1 proof job132041097584 SUCCEEDED, but only supplied f535 PR-head/base guards are credited; detailed report/hash and Windows-job-specific checkout/tree/host/SDK remain unavailable. Mac metadata is not copied. No result/watch carries forward.

Historical f122 CI43463739914/a1 FAILURE16/3fail/4cancelled, archive43463739998/a1 SUCCESS3/3, and f122 source HOLD remain historical. Windows failed at the directory-incapable helper; Mac primitive2/1ignore preceded bootstrap35/3fail/4ignore, before process/socket/genuine-writer/A3. Its ownership/policy/non-Running primary cause remains unknown; cancelled axes are unavailable, not passes/assertion failures. Original Drop-loss and R2 fixes remain source-present. Incoming215da CI20/1fail/2cancelled and older6a author results are not current clearance.

The separate synthetic0516 Windows credential timeout is retained ONCE at #167 existing record; tree/cause unknown, no changed-line defect, retry, credential repair or five-second budget waiver assigned here. Distinct cap-seed occurrence remains the 182 receipt, co-tracked as additional208 Live-setup debt, not208's named scoped-scan recurrence/167/new260repair. No production480ms/attempt-policy relaxation or duplicate residual post.

Native failure oracles, genuine official Mac writer, SAME-root actual public sender and App A3, exact-head CI/current-main and Campaign's final fresh gate/ordinary merge/follow-through remain mandatory. Intel/other OS versions/full backend parity are not claimed. Broad application-client conformance/fixture harnesses remain Mac-gated; #172 stays open. No shared installation, operator daemon/permission mutation, release/tag/publish work, new reviewer/observer or scheduler replacement.

Prepared with PAW.

macOS refused every bootstrapped daemon connection because it had no way
to prove which executable file a running peer is executing. Read that
identity from the kernel instead of refusing: the (dev, ino) of the vnode
behind the peer's mapped main image, via
proc_pidinfo(PROC_PIDREGIONPATHINFO). It is unchanged by renaming,
replacing or unlinking the image path, and needs no privilege for a
same-user peer.

- platform_fs: process_main_image_file_identity, fail-closed when the
  lowest file-backed mapping is not a read-execute offset-zero image
  header, the process is another user's or has exited, or the bounded
  walk finds nothing.
- daemon: macOS peer identity read bracketed by process start time and
  held to the same instance through the capability-record pin; a held
  descriptor keeps the selected inode allocated through HelloAck.
- daemon_bootstrap: the spawned daemon reads its own mapped image
  identity on macOS rather than re-statting its path.
- Other Unix platforms still refuse before selection or spawn. The
  v0.2.1 missing-admission probe stays Windows/Linux only.
- Enable the InstalledCurrent/ExactExecutable process tests on macOS and
  add a macOS CI job; document the guarantee and its limits.

Closes #259

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@lossyrob

Copy link
Copy Markdown
Owner Author

Please confirm the external agent ownership for this #259 PR: your owner label, whether this is the operator-routed #259 handoff, the base commit you started from, and the exact source commit used for the reported native results. The current public branch is claude/issue-259-a2d3b8 at 6a01562e4b7fce0483a0e5e32b6e83b994893b00, whose parent is 3bb75ce397a58f790429bee5b348d1ea0c9106a1. The shared GitHub login alone does not identify the coding agent, and #257 ownership does not automatically transfer to this separate task.

Please also link any existing source-bound native records supporting the PR's reported results, including the actual architecture and OS. No new run is requested.

The retained Local implementation/design reviewers are reviewing this exact head. The proposed mapped-vnode guarantee and unported missing-admission probe need explicit contract dispositions. The existing ApplicationClient coordinator is handling #259 acceptance item 3: actual sender readiness through the public InstalledCurrent seam. Bootstrap tests alone do not establish that acceptance; full #172 SQLite/Postgres lifecycle and health parity remains separate.

This requests ownership and existing evidence only. It does not assign another writer or request a source change, new test run, installation, release or merge.

@lossyrob lossyrob left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Changes required for #259 at 6a01562e4b7fce0483a0e5e32b6e83b994893b00: provide a bounded compatibility repair for the reachable official v0.2.1 missing-admission installation state on macOS. An intentional migration exception or deferral is not accepted. This concerns the new platform support; it does not claim older macOS bootstrap already worked.

The official v0.2.1 writer lacks application_bootstrap_admission_version and shipped both Mac architectures. This candidate's released-writer proof installs the exact candidate with that updater and confirms the field is absent. On macOS, InstalledCurrent then refuses before attach/readiness. The two ignored missing-admission compatibility tests do not establish support for this reachable path.

First confirm the external owner and source bindings requested in the custody comment. After that binding is verified, provide a bounded repair proposal for the existing implementation, design and ApplicationClient reviewers. The plan must preserve witnessed image identity, admission/version/build/protocol/capability fences, selector/manifest/target consistency and fail-closed unsupported or indeterminate results. No particular mechanism, including suspended spawn, is preaccepted.

An admission-aware updater installing or reselecting the exact candidate into a NEW immutable manifest may be documented as user-facing remediation. It is not a compatibility waiver or proof. Do not edit an old immutable manifest or substitute ExactExecutable, a raw connection or a private Watcher path for the public InstalledCurrent seam.

Required #259 acceptance item 3 remains pending before merge/closure: accepted same-candidate bootstrap/image evidence and an exact-source public InstalledCurrent real-sender readiness receipt are separate requirements. Full #172 SQLite/Postgres, restart/lifecycle-health and containment acceptance remains downstream; #172 remains open.

This records the required outcome and review hold. It authorizes no unbound writer, source patch, new native run, installation/root/selector mutation or release. The remaining implementation, normative-documentation, native/CI and current-main integration findings still need their own dispositions.

@lossyrob lossyrob left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Source delivery HOLD at 6a01562e4b7fce0483a0e5e32b6e83b994893b00 (tree 83457885b6f9f4c63729a863e09c590aa0edd1ee, actual parent/merge-base 3bb75ce397a58f790429bee5b348d1ea0c9106a1).

The retained independent implementation review covered all six changed paths (+620/-16) and the selection/witness, PID/start-time, child admission, HelloAck and negative-test dependencies. It established no additional high-confidence identity/lifecycle implementation defect within the explicitly accepted mapped-vnode guarantee. That property is weaker than Linux's exec-image link, but the design review accepts it within the existing same-user/intra-user exclusions. No additional operator identity-contract vote is needed. This is not an unqualified source or product pass.

F2: reconcile the normative contract and required evidence. The new descriptive macOS section does not update the active Linux/Windows-only pre-Hello and platform-evidence clauses, or ADR 0053's platform identity and compatibility text. Update docs/design/daemon.md:2150-2164,2227-2240,2242-2303 and docs/design/DECISIONS.md:2470-2485,2560-2571 coherently. State the actual mapped-vnode guarantee, replacement detection rather than descriptor execution, same-user limit, compatibility boundary and required macOS evidence. Do not claim Linux exec-link equivalence or unexecuted Intel/backend coverage.

F3: the reachable official-v0.2.1 installation path requires a bounded compatibility repair. The missing-admission macOS refusal and two ignored compatibility-success tests remain unresolved. The published required-outcome review records the reachability evidence and scope decision: no migration exception or silent deferral. Actual external owner/contact binding must precede its bounded proposal and exact implementation/design/ApplicationClient plan assessment; no mechanism or repair execution is preapproved.

Acceptance item 3 remains required and PENDING before #259 closure. attach.ready in process tests is not the public sender-health receipt. The existing public InstalledCurrent consumer probe must establish the actual send-only sender's readiness with no receive readiness, attended-but-deaf state, or inbound pending/actionable work, on the bound candidate and isolated frozen absolute root. The ApplicationClient owner reviews that receipt.

Other unresolved gates are the exact-head failed macOS image-identity job, external owner binding, source-bound native evidence and current-main integration. The failed job is 131633015347 in run43348340690/attempt1; its cause was not established by this source review. The head remains ten commits behind the observed main base a9a4e780fa4d39c161a27da1900b871bc67717ad; head-only results are not merged-current-main proof. Reported arm64/macOS26.7.1 SQLite results are not reviewer executions, Intel/Postgres coverage or full #172 acceptance.

The accidental #172 closing reference is resolved: parsed closing references now contain only #259, and #172 remains open. Its full lifecycle/backend/containment gates and PR #257's separate findings remain unchanged. No source patch, native run, CI retry, installation, release or merge is authorized by this review.

@lossyrob lossyrob left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

CI addendum to the source-delivery hold at 6a01562e4b7fce0483a0e5e32b6e83b994893b00. This classifies the existing failed-job logs; no rerun or native reproduction was performed.

CI-MAC-1: the isolated fixture socket path exceeds the macOS budget. Job131633015347 checked out the exact PR head on arm64 macOS26.7.2/25G332. The primitive command passed 2 tests, with 1 ignored. The installed_current_process command then reported 10 passed, 15 failed and 3 ignored.

One failure is concrete and occurs before peer-refusal assessment: tests/installed_current_process.rs:2250:59 fails to bind its foreign endpoint with InvalidInput: path must be shorter than SUN_LEN (log849-850). The new CI job's base under RUNNER_TEMP is already 41 bytes. Composing that base with the fixture label/PID/millisecond/counter namespace, /run, and the daemon socket's actual 12-character hash gives source-derived lower bounds of 105 bytes for ic-foreign and 109 for ic-spawn-reuse, exceeding Darwin's 104-byte sun_path field. These are lower bounds, not claimed observed full endpoint names.

After actual Mac owner binding, the bounded correction/proof must restore a valid composed isolated-fixture path budget without removing uniqueness, private roots or negative coverage. Relevant sources are .github/workflows/ci.yml:466-474, tests/support/telex_isolation.rs:110-113,174-178,245-249,466-472, and src/daemon.rs:313-321,1905-1911. The other failures include cold-start attach timeouts; the available daemon evidence does not establish that path length explains all 15. No kernel-version, vnode-identity or PID-bypass cause is established.

CI-WIN-1: a separate existing credential-writer assertion failed. Job131633015366 actually checked out synthetic merge 8779a23cc871fe46113cad2618709002d91dc60c, not the event head. Its workspace test invocation failed credential_command_shell_exit_preserves_delayed_inherited_writer at tests/credential_command.rs:848:9: the inherited writer was not collected within the five-second budget, with elapsed time 5008.5ms. The suite reported 10 passed, 1 failed and 2 ignored.

The scheduling/writer/collector cause remains unresolved. Missing markers mean unpublished at the post-cancel read, not code never reached; the snapshots are not an atomic deadline state. The analyzer reported no native Telex dump to analyze. This receipt is routed to the existing #167 residual owner, without asserting a new PR260 regression or an issue256 crash cause. Relevant source blobs match the known head/base sources; whole synthetic-tree custody is not claimed.

The sole observer reports terminal CI FAILURE (21 successful jobs, 2 failed) and archive SUCCESS (3/3). Archive success clears none of these failures, F2 normative documentation, F3 required compatibility repair, mandatory public sender-readiness proof, external owner binding or current-main integration. No source patch, retry, new native run, installation, release or merge is authorized by this addendum.

The macOS bootstrap identity job put TELEX_TEST_BASE_DIR under
$RUNNER_TEMP. The isolated daemon socket nested beneath it exceeded
SUN_LEN, so daemons could not bind and every process test that attaches
failed. Use a short /tmp base, as the suite already expects locally.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@lossyrob

Copy link
Copy Markdown
Owner Author

Operator-directed ownership handoff: PR #260 / issue #259 is now being taken over from the Mac agent by local Copilot session 52d8b24c-708d-4e75-863a-2e8fe1cae0d8 on Windows, coordinated by Campaign and the existing Local Daemon coordinator.

The incoming branch is claude/issue-259-a2d3b8 at 215da49. Both incoming commits are retained. The local session is now the sole implementation/push owner of this PR; the former Mac writer should not make further pushes. Existing Mac evidence remains attributed to its actual source and host.

The local owner will assess the pushed socket-path fix, resolve remaining review findings, obtain fresh independent implementation/design and required ApplicationClient sender-readiness evidence, integrate current main ordinarily, and carry the same PR through Campaign's merge gate and post-merge follow-through. The published holds remain until their outcomes are actually satisfied; this notice is not a review approval or merge clearance.

Parsed closing references must remain issue #259 only. Full #172 lifecycle/backend/containment acceptance and PR #257's separate findings remain unchanged.

@lossyrob

Copy link
Copy Markdown
Owner Author

F3 repair proposal for PR260, based on215da490 and to be integrated with maina9a4e780

The official v0.2.1 updater can install this candidate under an immutable manifest
without application_bootstrap_admission_version. I propose implementing the same
bounded, in-memory compatibility completion on macOS, with a native suspended
launch that authenticates the mapped image before allowing the metadata command
to execute. No existing manifest or selector is rewritten.

Mechanism:

  1. Retain the current raw-manifest eligibility, 0.3.x lineage, strict metadata,
    exclusive selector admission, byte/file-identity binding and final revalidation.
  2. Open the selected executable O_NOFOLLOW, compare its dev/inode to the candidate,
    and retain that witness until the probe has been terminated/reaped.
  3. Use Darwin posix_spawn with POSIX_SPAWN_START_SUSPENDED, a new process group,
    close-on-exec-default descriptor policy, null stdin and explicitly piped output.
    The environment stays cleared except the existing temporary/system allowlist
    and launcher guard. The argv remains --json version --root <trusted_root>.
  4. While the child is still suspended before userspace execution, require direct
    unreaped-child ownership, capture start time, read the kernel's mapped-main-
    image dev/inode, and confirm the same start time and witnessed selection.
    A path replacement, exited child, inaccessible/indeterminate image or changed
    process instance fails closed without resuming the command.
  5. Resume only that authenticated child. Reuse the bounded nonblocking pipe and
    metadata validation contract: ten seconds including cleanup, output caps,
    no credentials or daemon serving. Keep the Unix unreaped leader as group
    signal authority; seal group authority before reaping and never signal a
    numeric PGID after ownership loss. Any launch/identity/collection/cleanup error
    produces the existing typed incompatibility or identity failure.
  6. Require exact admission1/package/build/schema/protocol/capability metadata,
    then revalidate selector/manifest bytes and identity/target path and identity
    as today. The daemon still independently admits the child and publishes
    runtime identity; pre-Hello and HelloAck remain load-bearing.

Implementation boundary: macOS-only spawn/identity adapter, sharing the existing
bounded collector and Unix ownership logic where Darwin exposes the same waitid
WNOWAIT semantics; no Windows/Linux semantic changes. The Darwin start-suspended
flag is an OS API, not an entitlement/helper, kernel-version allowlist or new
release requirement.

Validation: enable both currently ignored missing-admission macOS process tests;
add a suspended-child replacement-refusal regression proving no user-code marker
appears before identity acceptance, successful exact metadata collection, and
bounded output/deadline/cleanup cases. Retain bytes/mtime/file-identity immutability
and cross-process serialization checks. Run the existing primitive and complete
installed_current_process targets on native hosted macOS at the integrated head,
including actual architecture/OS/source. App separately assesses and owns the
public InstalledCurrent real SendOnly sender-health receipt using the existing
run_watcher_probe fixture; bootstrap attach.ready is not that acceptance.

Alternative considered: an admission-aware updater can create a NEW immutable
manifest as remediation, but does not by itself establish the required reachable
compatibility path. Accepting a version string or reopening the path after an
unsuspended quick-exit probe would forfeit the witnessed image guarantee.

Please assess this concrete plan before identity/admission implementation.
Native success, APIs on the hosted SDK, and the final exact-head review remain
unproved; this proposal is not an approval or evidence receipt.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@lossyrob

Copy link
Copy Markdown
Owner Author

F3 revised repair plan for PR260

Source: d71b896,
tree40a6072a9788eeb71a2f38ae11979a15f60b694a,
ordinary merge parents215da4909c789c8a07e27baa02634fb3a3b19ec9 and
a9a4e78. No F3 code is changed yet.
Independent uncommitted documentation/fixture edits will join the final candidate.
This supersedes proposal6858210199 and incorporates retained af R1-R6,
App APP-R1-R5/A3 and bc P1-P4. Their direction assessments are not native API,
implementation or evidence approval.

Outcome and authority

Support the reachable official-v0.2.1 missing-admission macOS installation using
bounded in-memory completion, never rewriting an existing immutable manifest.
Keep existing eligibility (raw key ABSENT and 0.3.x target), selected package,
build, schema, protocol, capabilities and admission1. Keep root/selector/manifest/
target consistency, child admission, pre-Hello image credentials and HelloAck.
No new helper, entitlement, privilege, version allowlist, release or trust change.

Owned native adapter

The existing collector creates std::process::Child through Command::spawn.
Introduce a macOS-only owned raw-posix_spawn child/pipe adapter, with ownership
beginning immediately at successful native spawn. Never fabricate Child from a
PID, detach it, or transfer it to a std/Tokio reaper. Deliberately extend the
existing ProbeGroupOwnership and nonblocking/observational Unix helpers to Darwin,
retaining the Windows/Linux implementations and error behavior.

Initialize owned spawn attributes/file-actions and close-on-exec pipes BEFORE
spawn. Every pre-spawn setup failure drops those owned descriptors/attributes.
Use canonical-path posix_spawn, NOT posix_spawnp, with START_SUSPENDED,
SETPGROUP(0), CLOEXEC_DEFAULT, an empty signal mask and reset signal defaults
including HUP and CHLD. Only /dev/null stdin and explicit stdout/stderr pipe writes
survive. Environment is cleared and contains only the existing
SYSTEMROOT/WINDIR/TEMP/TMP/TMPDIR allowlist and launcher guard; no DYLD, credentials,
store/session or inherited variables. Argv stays --json version --root .
Unsupported attributes/spawn errors, including EINVAL, fail without fallback.

Open selected executable O_NOFOLLOW, compare its dev/inode to the frozen token,
and retain that exact witness through termination/reap. Before spawn, reject
SIGCHLD ignore/automatic-reap policy. A single absolute10s budget starts BEFORE
spawn, covering verification, resume, output drain and termination/reap. Retain
the existing cleanup reserve and check remaining time at each boundary; never
reset the deadline. This is a bounded polling/collection policy, not a claim that
synchronous kernel calls can be forcibly cancelled at10s.

While the direct child is suspended:

  1. Establish direct unreaped-child ownership with waitid/WNOWAIT.
  2. Capture PID/start time and read (dev,ino) of its LOWEST-ADDRESSED FILE-BACKED
    mapping under PR260's accepted vnode rule: regular vnode, offset0, RX/nonW.
  3. Recapture start time and require it unchanged; require mapped identity ==
    held witness == frozen selection. Unsupported/inaccessible/indeterminate image
    or missing start time refuses identity, never skips to a later mapping.
  4. Immediately before SIGCONT, recheck SIGCHLD/wait ownership and the same start
    time. Resume only this authenticated PID, never the group.

This is the weaker mapped-vnode guarantee, NOT kernel exec-image identity or Linux
exec-link equivalence. Pathspawn detects replacement, not prevents it. Suspension
removes the child's own userspace from the pre-verification forgery set; other
same-user code remains excluded. A main-entry marker's absence proves only that
marker was not reached, not independently that no loader/initializer executed.

After successful spawn, EVERY setup/identity/resume/pipe/metadata/deadline failure
enters owned cleanup. Verification failure NEVER sends CONT: while a fresh direct
unreaped leader anchor authorizes signaling, kill the owned group, seal authority,
then reap the exact leader within the original deadline. Normal success also
seals group authority before reap. ECHILD, auto-reap or indeterminate ownership
seals without a numeric group signal; no Drop or late retry signals a recycled
PGID. An interrupted observation is not positive ownership evidence. No claimed
cleanup success without exact termination/reap. Preserve bounded descendant-pipe
collection and caps. The orphan-stopped-group HUP/CONT reasoning is a design
expectation requiring native evidence, not an already-executed result.

Validation and typed failures

Witness/mapped/start mismatch OR inability to establish image/instance identity:
ExecutableIdentityMismatch. Unsupported spawn/attrs/ownership/resume, malformed/
nonzero/mismatching metadata, pipe/cap/deadline or unconfirmed cleanup:
IncompatibleManifest. Only valid-but-changed final bindings: SelectionUnstable.
Preserve actual root/authority/selector/target validation errors BEFORE equality
(InvalidTrustedRoot, UnsafeInstallAuthority, MissingCurrent, MissingExecutable,
InvalidManifest, and applicable compatibility errors). Never silently remap
Windows/Linux. When cleanup also fails, surface IncompatibleManifest and record
both the primary typed cause and unconfirmed-cleanup diagnostic without paths or
credentials; otherwise preserve the primary failure.

Hold exclusive selector admission across probe AND final revalidation. Bind the
frozen canonical root and identity, current selector tag/file identity, raw
manifest bytes/file identity, canonical target/file identity and source/build.
Reacquire shared admission and revalidate again as today. The daemon still makes
its independent child admission and runtime capability proof.

Native tests must establish the actual stopped-child region walk and Darwin
waitid/WNOWAIT semantics, not extrapolate from a running-after-dyld primitive.
Cover suspended success, replaced/wrong image with no resume/no main-entry marker
and confirmed kill/reap; unavailable/failed setup, attrs, image query, start capture/
change, resume, SIGCHLD ignore, ownership loss/EINTR and unexpected first mapping;
output caps, original deadline/descendant-held pipes and irreversible ownership
sealing. Enable both synthetic missing-field tests, retaining byte/mtime/identity
immutability and cross-process serialization.

Genuine released-writer and App proof

Extend the EXISTING released_v021_writer_installs_exact_candidate proof to native
macOS. Use the architecture-matching official v0.2.1 binary, verify its reported
build212b76a4c586101bdc2a53264e2a4c3e2326671a, and pin the release archive digest:

  • aarch64-apple-darwin:
    07ab2088645363716c1fa826aa877866e7b3195f155ac8463c973cd2c5256f0f
  • x86_64-apple-darwin:
    220c49cf3924024d733d0429f2ceaa01289ac95b5eb984c30647100c50478d34

These are GitHub release asset digest metadata read at planning, not local native
executions. CI must check actual downloaded archive AND published checksum sidecar
against the pin before extraction. No Intel execution claim unless its leg runs.
The released updater installs/selects the exact integrated candidate in an isolated
root. Assert genuinely ABSENT raw admission key, exact package/build/schema/
protocol/capabilities and target path, selected tag and hash/bytes-identical
installed candidate. Record unchanged manifest bytes/mtime/file identity afterward.
The existing Windows official-asset proof remains intact; neither synthetic shape
nor an admission-aware updater/new manifest substitutes for the native old writer.

Use the SAME genuine-v021 root with the EXISTING external run_watcher_probe for
separate mandatory A3. Add explicit registered assertion and raw responsibility/
store/runtime/SendOnly/health tuple reporting INSIDE that fixture if necessary.
Preserve stable logical store/fresh runtime, real public send/durable acceptance,
sender_ready=true, receive_ready=false, attended_but_deaf=false and zero inbound
pending/actionable. No new harness, private seam or attach-only substitute.

The existing exact-source macOS CI job carries these targets and durable receipts:
source/tree, OS/arch/SDK, candidate/build/hash, official asset pin, canonical frozen
root/selector/raw manifest/target identities and actual sender health. Correct and
measure composed Unix socket budgets first; preserve private unique roots and all
negative tests. Sole2a4 through Campaign observes the pushed integrated candidate.
Local retains bc/af delta review and App acceptance receipt. Campaign owns merge.
The cancelled215da Mac job supplies no native proof. All final receipts must be on
the same integrated candidate. Full172 parity/containment acceptance stays separate.

Preserve the accepted mapped-vnode contract, add an owned suspended metadata-probe adapter and extend exact-source native compatibility and public sender evidence.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

@lossyrob lossyrob left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Source delivery HOLD at f122c6d0fdc1378383019017069c7c4b5c19583f (tree 9f82ef57737bfb804882dd9f3e5c87940d138cab, parent d71b896d23106b2766621e7f0fcbe0117e9c4628, integrated main a9a4e780fa4d39c161a27da1900b871bc67717ad).

The retained implementation reviewer covered all nine changed paths and their integration dependencies against the assessed v2 plan. Two required P2 repairs remain:

  1. R1: preserve unresolved probe custody on cleanup failure. At src/daemon_bootstrap.rs:1616-1618 and the related ownership/deadline returns, cleanup-unconfirmed errors discard the raw macOS child/group and the caller's selected-file witness. A failed observation or unsuccessful reap can therefore return without retaining the unresolved child and witness promised by v2. This is a resource-ownership/conformance finding, not an established identity bypass. Preserve an explicit owned failure hold, keep the caller's deadline, and never resume or signal a numeric group after authority is lost. A bounded, non-signaling in-process quarantine is an acceptable design direction, but its exact code, plan/docs wording and failure-path tests still need retained review. The separate panic/orphan-group reasoning does not cover these normal error returns.

  2. R2: make the Windows root-identity observation directory-capable. New calls at tests/installed_current_process.rs:1781,1832 pass directory frozen_root to the helper at240-262, whose CreateFileW flags remain zero. The production directory-capable helper uses FILE_FLAG_BACKUP_SEMANTICS. Preserve the real before/after root-identity assertion and correct the observer. The exact-head official-writer Windows job panics in that helper with INVALID_HANDLE_VALUE; its log does not supply a last-OS-error value or reliable caller argument. The directory-open defect is source-derived, and the ordinarily ignored genuine-writer proof was not covered by the reported normal process-suite passes.

Retained design review finds the implementation and active normative documentation conform to the accepted weaker mapped-vnode contract, without a new trust decision. App finds the existing genuine-writer/same-root public Watcher route substantively correct, but requires self-describing command metadata, explicit before/after source and authority bindings, structured send/reconnect evidence, and the complete asset-to-candidate hash chain in the final report. Those reporting revisions remain required for an acceptable receipt.

The supplied macOS job passed the primitive command, then failed bootstrap-library tests. The suspended tests' primary error occurs in ownership/leader-state checks before mapped-vnode/start verification; the precise wait/signal-policy cause remains unresolved. The process/socket suite, genuine official macOS updater and public sender proof never executed after that failure. No successful native or acceptance3 receipt is established.

This is a nonapproving source-hold review, not CI or merge clearance. Fresh reviewed-source native evidence, the separate same-root public InstalledCurrent actual-sender receipt and the normal Campaign merge gate remain required. The existing #208 residual is outside this repair; #172 remains open.

lossyrob and others added 2 commits November 13, 2026 12:08
Record exact commands, source/tree and before/after bindings, release/candidate digests and structured public sender evidence in the existing proof route. Include native custody tests in the existing bootstrap filter.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Prearm exact-child and same-opened CLOEXEC witness custody before native spawn. Quarantine returned errors without signals, retain indeterminate generations until client restart, and use non-consuming terminal observation before one exact reap. Add real-child failure roles and sanitized native diagnostics; document attempted-cleanup bounds.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@lossyrob

lossyrob commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner Author

Latest observation addendum: exact1fed terminal CI FAILURE; native HOLD unchanged

Campaign/sole2a now reports actual CI43484755946/a1 terminal FAILURE at19:03:54Z, with retained daemon132090138558 failing; archive43484755959/a1 SUCCESS. The existing diagnostic continuation is retired, b414 cleared/readback ABSENT. No PR260 watch is revived. The separately admitted RC2 tag record is independent and is not Mac proof.

Actual raw member admission.log is now available from SAME-run43484755946/a1/job132090138558 macos-bootstrap artifact14066468520:13880bytes/143LFrecords/SHA9637897564201209b63a6b0c69e54f9fb9de6bc266998c8d54e9a94d32f4f003, artifact/run/job linkage sole2a/Campaign-attributed. Owner independently verified length/hash/LF, not a duplicate content/cause trace.

SAMEbc's FINAL additive provenance report5840/SHA6a461304b85c0334204716bf2b2ba3f4c37b438f957ad2c5828cbbf30529fa76 matches ALL143 raw lines uniquely to job704-846 (rawN->jobN+703), comparison-only timestamp-prefix removal, ANSI/content/order preserved and no files rewritten. Raw27/job730 is the nested normal-control KILL-1/errno1; raw28/job731 skips reap. Distinct outer successful paths and later replacement failure remain separate. Owner FULL-read report/manifest906/SHA4de30dc3aa6925412df4114a6314b09c1e2964f75d9c4b3da99d423fe43c2fbf and rehashedALL4. This closes raw custody/correspondence ONLY; no new fetch/repro/retry/cause classification.

The measured native SIGNAL failure remains KILL-1/errno1, then reap NOT attempted. WHY/kernel/host/group membership remains unestablished: no ESRCH, wait-error, groupabsence/containment proof, cleanup waiver or behavior fix. All three exact1fed SOURCE dispositions remain separate. App accepts only TEMPORARY stderr; explicit post-cause retention/gating/removal and actual official Mac writer/SAME-root public A3/product/final merge gates remain.

This supersedes only earlier captured unconfirmed-watch/unknown-overall-CI/missing-raw-file statements. The original39-file handoff and JOBLOG-only classification/captures below remain honest frozen history, not retroactively relabeled raw evidence. No source/logging change, new push, observer or scheduler action.

Preserved source-publication disposition captured18:57:54Z (availability/status superseded above)

PR260 current disposition: ACTUALLY PUSHED1fed SOURCE clear; actual required Mac daemon job FAILED/native HOLD. No cause/behavior fix or final acceptance.

Head1fed64e687a5d9bbae6267f6329f2fb217bdff94 / tree49d3b31173318ee0cc254f526ad6f3b8ff3258b8 / parent8eb83f774f349e97e1ed2b0384c9d6efd2cd7edc. Parent8eb is the ordinary af42+current-main9facc merge (tree73e61c71ee09438d82e16124588fb43445597f66); incoming/ordinary history is preserved. Normal af42->1fed push and fresh exact public commit/PR metadata are independently confirmed, OPEN/non-draft/base9facc.

NEW ACTUAL1fed required daemon job132090138558/CI43484755946/a1 FAILED. SAMEbc FINAL frozen classification14025/SHAaa4a74ba23775fcc6cba88320887532f6d08e151796eae898ccd013a3fd62488 binds Campaign-supplied direct log78863bytes/913lines/SHAec9bbeb9c2cfeef4d256f9e6f22d406a44a38e018e9f6f3c9e4d7fa91ad11a38 and EMITTED exact1fed/tree49d3/SDK26.5. Owner FULL-read report/manifest1042/SHA523e05638e655604d12ebaa65880b0e6ffc943f89da91bfec661ae2178b4e240 and rehashedALL4, not a duplicate trace. Enclosing FIRST deadline-custody role stderr/ThreadId2 NORMAL control: Stopped726-727 ->Exitedtrue728-729 ->groupKILL=-1/immediateerrnoSome(1)/ownershipSealed/generation_known=true730 ->phase=signal,reap=not-attempted731 ->Bquarantine732 ->normalcontrolpanic735. Actual SIGNAL-stage failure/rawerrno1 at1083-1104/1806-1813, NOT ESRCH/consuming-wait failure. WHY errno1/kernel/host/group membership remains unestablished, no groupabsence/containment proof. Guarded semantic disposition BEFORE behaviorfix/cleanupwaiver; no old-head cause carry.

Separate replacement ThreadId12 similarly fails signal801-806. Distinct successful refusal-path kill/reap770-784 and descendant ThreadId5 kill/reap756-760 do NOT prove normal-control cleanup. Library37pass/3fail/5ignore; forcedcustody/reuse/ECHILD/EINTR after normal control and process/socket/genuineMacwriter/A3 not reached. Full JOBLOG ONLY, NOT independently attested raw admission.log; actual artifact/run/job/byte-length/SHA still needed.

The two-file +92/-20 delta adds precise macOS finalization phase/thread/raw-KILL-return-and-errno/seal/exact-reap diagnostics. It preserves ownership/generation/signal/deadline/terminal-wait/B quarantine and original failure order: failed signal STILL skips reap. NO behavior fix is chosen; no guessed ESRCH, containment shortcut, cleanup waiver, parked-slot release or broader C4. Exact committed Windows36pass/2roleignore, Darwin Postgres cross-tests/production-lib clippy pass ONLY. No native Mac/Intel or SQLite-gated genuine-writer/A3 execution here. Main integration includes accepted RC2 version/Windows bridge changes; own diagnostic delta changes no report/receipt shape or af42 native adapter.

All three retained exact1fed SOURCE dispositions are now clear: bc SOURCE CONSERVATION PASS/0required, af CONFORMS/no deviation, App REPORTING SOURCE CLEAR/no schema change. bc conserves all21 accepted RC2 import payloads/strict Telex-owned APIs and six unchanged identity/proof dependency blobs. Owner FULL-read bc11364/SHA88a881d9fda6f6a7c62cf6978ed7c88cf6f03787e8f0c90aa4e891e876b14209 and separate publication status2011/SHAda9f45184357aa01fa438b03e8b179b3122941dcffd0977b59d870a7ba0af367 and rehashed21+2 bound files; af/App are Local-attributed. Original UNPUSHED-at-capture packets stay frozen; the SAME actual push is separately bound.

App accepts exact1fed's unconditional macOS library stderr as bounded TEMPORARY diagnostics, NOT permanent consumer-output approval. af marks it LOW/nonblocking, not a contract deviation. AFTER the native cause is known and BEFORE final product/source acceptance I must explicitly decide retention/gating/removal, preserving raw evidence without silently making routine stderr public consumer contract. No implementation/logging change is requested now.

bc qualification: ThreadId is PROCESSLOCAL, not globally unique across nested fixture processes; future native classification needs enclosing role/process context. Synchronous diagnostics may affect timing/interleaving and are not zero-overhead/hard sink deadline proof.

Any future diagnostic packet relying on these lines must cite raw admission.log by ACTUAL artifact/run/job and independently retain byte length/SHA256. report.json or a job label alone does not attest raw lines. This is the unchanged capture route, not a new receipt schema/run/admission request.

The already supplied af42 native classification remains historical: accepted Stopped/terminal observations reached ordinary post-terminal finalization, with its failed cleanup substage/errno unknown. That is not the cause of the NEW1fed failure merely by ancestry. Precise actual1fed evidence plus retained semantic disposition must precede any behavior fix. All af42 retained source passes and older runtime/proof results remain HISTORICAL.

Initial1fed watch retired/52d cleared by sole2a. Campaign separately assigning same-attempt terminal/artifact diagnostic continuation within ORIGINAL20:10Z, not yet confirmed live here. Overall1fed CI/archive/official-writer/A3 outcomes unknown; Campaign retains final fresh current-main/CI/merge/follow-through custody. The af42 watch stays retired; independently admitted merged-RC2 shared scheduler untouched, no Mac dependency imposed. No owner native repro/retry/provider query/new observer/reviewer. Native failure oracles, genuine official Mac old updater, SAME-root actual public sender and App A3, exact-head CI and merge gates remain open; attach.ready is not A3.

The separate synthetic0516 Windows credential timeout remains ONCE at #167 (comment) ; no cause/budget waiver/credential repair/retry scope or owner revival. Closing set remains259 ONLY;172OPEN. Existing body and THIS SAME comment updated, no duplicate post; frozen af42 and prepush1fed proposal files/captures are preserved with a new publication receipt.

Historical af42 disposition (unchanged pre-1fed snapshot; not a current watch)

Current guard addendum: Campaign/sole2a retired the af42 diagnostic record on actual main/base drift to 9facc00 (RC2 merge). NO active af42 watch remains; unfinished overall CI/archive/official-writer outcomes remain unknown. The held failed Mac job/classification and source dispositions stay bound to af42 only. Integrating current main and any actual pushed successor require ordinary exact-source review and separate Campaign/same2a admission, with no af42 pass/result carry. The shared scheduler is active ONLY for separately admitted merged-RC2; it is not cleared or replaced for PR260. No Mac dependency is imposed on RC2.

PR260 current disposition: exact af42 source clear; native finalization HOLD.

Latest supplied af42 disposition

Actual af42 required Mac daemon job132065388872/run43477596181/a1 FAILED. Campaign supplied its held log to the SAME retained bc for exact-source/reached-versus-skipped classification. No cause or f535 carry is inferred, and no retry/native reproduction/provider fetch is requested here. The admission watch stopped; Campaign separately assigned bounded diagnostic continuation, not yet credited live in this receipt.

Retained bc's actual-af42 verdict is SOURCE PASS, S1 SOURCE RESOLVED by B, with zero new required findings. Retained af CONFORMS with no required changes; App reporting is SOURCE CLEAR with all four reporting blobs unchanged and no receipt-shape revision. All three source dispositions are clear for exact af42. These do not accept native behavior or close native failure-oracle, genuine official Mac writer, SAME-root public A3, final CI or Campaign merge gates. The reuse oracle uses a real competing reap/another real child but simulated numeric reuse, not an OS allocator reuse trace or native pass. C4 remains as assessed; optional broader post-CONT stop handling is not required scope.

The source-publication record below precedes these new observations; its absence of successor run IDs and pending source assessments are superseded by this addendum. Historical f535 S1 HOLD is not af42's source verdict. No native success or merge readiness is claimed.

Final supplied SAME-bc classification of the held af42 log (11800 bytes, SHA256 f5a3c18965820d47be5c2d6d414f576d0a404e51f698fc2e7719d380d2413ac8): code5/status0 is now accepted as Ok(Stopped), options0x2d, and code1/status0 as Exited=true. This is not f535's unexpected-state rejection. The first custody subprocess fails its initial NORMAL authenticated-probe control at ordinary post-terminal finalization, before forced deadline/ECHILD/EINTR/reused roles. The expected-success replacement case also reaches finalization before incompatible-manifest refusal; later expiry refusal follows parked custody and does not prove OS spawn failure. The state-table, descendant-held-pipe deadline and automatic-reaping-policy tests actually pass. The later output-cap and forced custody roles are not reached; full process/socket/genuine-writer/A3 remain skipped. Group-kill return/errno and exact failed cleanup substage are not established. Precise owned-kill/seal/reap diagnostics and semantic disposition must precede any behavior fix: no guessed ESRCH, containment bypass, bad-cleanup waiver or parked-slot release. The original generation/signal-seal/deadline fences remain.

Source-publication record (before latest observations)

🐾 PAW Implementer: Re-review Requested

ACTUAL PUSHED successor af42d45 / treef6358e19903d6732e1e67743662318aefb79f6c5 / full parent175cdd90e85dd6609a5622f7bc960f39926a5974 / basea9a4e780fa4d39c161a27da1900b871bc67717ad. Two ordinary commits follow f535: 175cdd9 permanently seals ALL parked custody; af42 adds the separately assessed stopped-owned guards. Incoming commits/ordinary ancestry are preserved. Actual retained implementation/native/App/CI/merge acceptance remains pending.

B is SOURCE-PRESENT: ANY parked child-bearing record permanently refuses native legacy admission across roots until CLIENT restart. NO later PID observation/signal/reap/release, background cleanup or clock extension. Strict/cache paths remain unaffected. The real competing-reap/different-real-child oracle checks refusal, identical witness/slot retention, zero observation and no stolen terminal status. Its reused numeric PID is deterministic test simulation, not a claim of actual OS reuse. The f535 S1 SOURCE HOLD and af withdrawal are not retroactively cleared; the rejected matching-start workaround is not implemented.

The distinct code5/status0 correction implements the reviewed initial4417-byte plan PLUS explicit C1-C5, NOT acceptance of the later4418-byte copy. Darwin requests WSTOPPED, not WCONTINUED. Both pre-authentication and fresh pre-CONT require owned Stopped0/SIGSTOP, never Running; mapped-vnode/start fences remain. Still-known original custody permits one ordinary refusal/expiry group KILL and signal seal. Genuine changed generation/observation error permits NO later numeric signal/reap. Stopped is pending, not metadata/exit/reap; only terminal WNOWAIT authorizes consuming waitpid. Unsupported/trapped/continued/mismatched observations refuse. Linux unchanged. Shipping-kernel/mask/f122 cause and actual stopped-image API remain unproved.

Exact af42 Windows bootstrap36pass/2ignore; Darwin Postgres tests cross-check and production-lib clippy pass. One duplicate-quiet invocation failed argument parsing with ZERO assertions; corrected invocation passed and both outputs are retained. The source-table test is pure and emits no synthetic kernel observations. Actual waitid logs preserve raw code/status/options. Native lifecycle/failure oracles and SQLite-gated Mac genuine-writer/A3 are NOT executed locally.

Supplied f535 CI43470584011/a1 is terminal FAILURE21success/1knownMacfailure/1cancelled Watcher Mac diagnostic; archive43470584067/a1 SUCCESS3/3. Mac job132041097486 failed bootstrap35/4/5ignore after matching code5/status0; no unique concurrent-test origin/cascade is established. Custody assertions after the first deadline role were not reached; downstream process/socket/genuine-updater/A3 were skipped. This is distinct from S1. Actual Windows official-v021 proof job132041097584 SUCCEEDED at f535 (https://github.com/lossyrob/telex/actions/runs/43470584011/job/132041097584), bound only to supplied f535 PR-head/base guards; detailed report/hash/Windows-specific checkout/tree/host/SDK are absent. No Mac metadata or old-head result transfers.

App's four reporting requirements were SOURCE-cleared at f535 and are unchanged here, not A3 acceptance. The v2 proposal and historical f122 HOLD stay historical. This updates the SAME disposition, preserving earlier frozen bytes. Campaign has the actual pushed tuple for separate sole2a4 admission; no successor run/job/artifact/native/genuineMacwriter/same-root public A3 receipt is credited. No second observer or RC2 scheduler change.

Retained actual source review, native failure/image/process/socket proof, genuine official Mac updater, SAME-root actual public sender and Campaign fresh CI/merge gates remain. attach.ready alone is not A3. The cap-seed residual stays the 182 occurrence with additional208 setup co-tracking, not named208 recurrence or this repair; no duplicate post/new owner/480ms waiver. Closing set remains259 only; issue172 stays open.

lossyrob and others added 5 commits November 13, 2026 12:39
Remove numeric-PID observation and opportunistic reaping after custody is parked. Preserve the witness until client restart and add a real competing-reap/reused-child refusal oracle.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Explicitly request Darwin WSTOPPED, distinguish pending stopped children from terminal wait receipts, and preserve still-owned cleanup versus permanent generation-loss sealing. Add focused parser/authority cases without emitting synthetic kernel diagnostics.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@lossyrob
lossyrob merged commit 67cf8ba into main Nov 15, 2026
26 checks passed
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.

[local-daemon] Support identity-verified macOS daemon bootstrap for ApplicationClient consumers

1 participant