Skip to content

Batch 18: ONNX Runtime from NuGet, old-kernel leases, withdraw and display fixes, protocol bumps - #1469

Open
aft-alfonso[bot] wants to merge 26 commits into
mainfrom
ci-speed/post066-batch-18
Open

aft-alfonso[bot] wants to merge 26 commits into
mainfrom
ci-speed/post066-batch-18

Conversation

@aft-alfonso

@aft-alfonso aft-alfonso Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Batch 18 on top of batch 17 (main ea5431c1).

  • ONNX Runtime installs from Microsoft's NuGet package. The GitHub repository it came from returns 404, so new installs couldn't fetch it. Hashes are checked; Linux and Windows files are byte-identical to the accepted ones, and macOS adds Microsoft's signed copy. Install hints use the approved new links.
  • Storage leases work on Linux kernels older than 5.8 by reading the mount id from /proc/self/fdinfo (Refs RHEL 8 AFT is crashed in opencode #1468).
  • tool.withdraw answers a task that died in a reboot as stopped, after checking its process (and, on Linux, that its scope is empty).
  • Display records announced before an AFT restart stay readable through display.read for the promised window.
  • The desktop app paired from another computer can list and read shells.
  • subc-protocol 0.37.4: AFT reads the extended federated route stamps, and admission is unchanged.
  • tool-provider 0.15.12: read tools are marked read-only, and fork_id calls are limited to them and never file an approval card.
  • Test fixes: Git automatic maintenance is off in the publication fixture (overlapping repacks on Apple Git 2.54), and the wire-version pins and confine golden are updated.

Search quality: engine_unwired. The only fence file is semantic_index.rs, for the ONNX install and the hint move.


Summary by cubic

Batch 18 fixes ONNX Runtime installation, storage leases on older Linux kernels, withdrawal liveness after a reboot, display readability across restarts, and macOS worker sandboxing, and bumps the subc and tool-provider protocol versions.

Bug Fixes

  • ONNX Runtime now installs from Microsoft's NuGet package; the previous GitHub releases source returns 404 and broke new installs.
  • Storage leases work on kernels older than 5.8 by reading the mount id from /proc/self/fdinfo, fixing AFT startup on RHEL 8 (Fixes RHEL 8 AFT is crashed in opencode #1468).
  • tool.withdraw answers a task that died in a reboot as stopped, after confirming its process is gone (and on Linux, that its scope is empty).
  • Display records announced before an AFT restart remain readable via display.read for the promised window, backed by a durable database cache.
  • A desktop app paired from another computer can list and read shells.
  • macOS jailed workers can run the build-slot rustc launcher and follow sibling-repository links, on both the strict and relaxed (default) worker profiles.
  • A cgroup.kill write failing with ENODEV counts as stopped only when the cgroup directory is gone; reconnect polling now uses production polling.

Protocol and Tool Changes

  • subc-protocol 0.37.4 reads the extended federated route stamps; admission is unchanged.
  • tool-provider 0.15.12 marks read tools read_only; fork_id calls are limited to them and never file an approval card.
  • Catalog goldens, notice-template vectors, and wire-version pins are updated for the new protocol versions.

Written for commit 4bc84ec. Summary will update on new commits.

View guided diff Turn on auto-fix

ualtinok and others added 22 commits November 18, 2026 17:25
Reuse startup recovery's PID/start-time probe and Linux scope adoption without changing recovery's existing boolean decisions. Missing-registry withdrawals freeze only after a terminal record or proven process death; live and uncheckable identities remain refused.

Keep withdraw_incomplete and its current detail unchanged: pinned tool-provider 0.15.10 does not classify transient as retryable. Retry-code changes await an upstream provider contract update.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Saved runtime or task-scope memory policy proves the task acquired a Linux scope. After its recorded PIDs are gone, reuse scope lookup and inspect cgroup.events before freezing: absent or empty scopes are stopped; populated, uncheckable, or unreachable scopes remain refused. Existing recovery decisions and refusal wire shape remain unchanged.

Add deterministic file-backed withdrawal cases for absent, empty, populated, unknown and failed scope observations, including runtime-only scope acquisition.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Keep the registry within its pinned hand-written line limit by moving the missing-registry identity and scope helpers into a sibling impl module. Preserve withdrawal and startup recovery behavior; leave only the module connection and shared recovery call in registry.rs. No size pins are raised.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
The size gate requires exact downward pins after a split. Record 19058 hand-written registry lines, below the previous 19070 pin; no pin is raised.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Persist frozen pull bytes before publishing a wake, with protected count/byte bounds and bounded expiry sweeps. Keep the optional delivery cache outside numbered schema migrations so schema 29 and its reader floor remain compatible with rollback builds. Existing route admission and discovered-reader checks precede both memory and durable reads.

Clarify the existing empty-after-restart outbox test as a memory-only invariant; the new service test separately proves durable recovery and publish ordering.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Blocking database work waits for the process-wide connection mutex instead of refusing ordinary contention; pull reads also wait for the brief outbox lock. Poisoned locks still refuse. Reply-side optional enqueue stays nonblocking.

Wake completion already uses ReadCache backoff rather than the push outbox's eight-attempt budget. Document the separation and extend the capacity regression to twelve deferrals followed by successful publication after one protected row expires, asserting the push attempt count remains zero.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Borrow startup inputs and derive the aft.db path in display::start, restoring the central subc module's pinned hand-written line budget without raising its limit.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Callosum relays a desktop app paired to this computer as reserved:callosum
with a fed-app:<fingerprint> harness. It gets the same view-only shells
operations as the person's phone. Other federated harnesses stay refused.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Read only the pinned platform library and ZIP metadata by default. Keep the existing cache paths, aliases, and legacy macOS pin. Replace whole-archive extraction and copy fixture tests with coverage for the ranged download, atomic failure cleanup, and cache aliases and permissions, since archives no longer supply filesystem paths.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
User-facing text only. Hold this commit for Ufuk's approval; the installer and internal recipe commits do not depend on it.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Batch 17 made the local-recovery fixture take an explicit mount, so recovery
never probes the host's cgroup tree. The withdraw test that came in alongside
it still used the removed single-path form.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
…odule

The approved install-hint wording added a line to semantic_index.rs, which
is past its size pin. The two wording helpers now live in
semantic_index/ort_hint.rs and the pin drops to the file's new size.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Pin subc-protocol to 0.37.4, which requires subc-config 0.1.1. Exercise wire binds for app and computer peers with and without the new optional metadata. Both must retain project-free admission without changing policy.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
…vider 0.15.12)

Bump cortexkit-role-tool-provider to 0.15.12 and its conformance kit to
0.15.13 (0.15.11 pins tool-provider 0.15.10). llm-runner stays at 0.8.53,
which accepts tool-provider ^0.15.7.

Every catalog now marks read, grep, glob, search, outline, zoom, callgraph,
ast_search and conflicts with `read_only: true`, under every name a preset or
plugin serves them. The contract sends the mark without any request opt-in,
so every catalog digest changes; the goldens are regenerated.

The mark promises that AFT honours `fork_id` (tool.call section of the
contract), so a fork call is now admitted before anything runs: a malformed
marker is refused `invalid_request {field: "fork_id"}` and a fork call to an
unmarked tool `fork_call_refused {tool, reason: "not_read_only"}`. A marked
tool runs under AskRoute::Fork on routes that could ask the person, so an
access question becomes `fork_call_refused {tool, reason:
"approval_required"}`, sent as the role's ERROR, and no card is raised. A
fork call sends no code-search nudge to the head. Calls without `fork_id`
are unchanged.

The codemode refusal moves from AppContext::answer_ask into
access_policy::unasked_refusal unchanged, and provider_shell_selector moves
to subc/bash.rs, to keep the pinned context.rs and subc/mod.rs within their
size pins; both pins are lowered to their new counts.

The notice-template samples are published, so aft-notice-templates-v1 stays
byte for byte as it was. The regenerated catalogs, now carrying the marks
and new digests, go in a new aft-notice-templates-v1.1 with its own README
and SHA256SUMS, and AFT's sample test now checks v1.1.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Git 2.54 may launch overlapping detached repacks when the fixture quickly commits a copy of the entire checkout and then a train edit. Keep maintenance disabled only in this disposable repository. Check the Git trace of a fixture commit so the regression does not depend on winning the repack race.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
…talog golden gains read_only marks

The 0.37.4 bump changes the wire crate version AFT reports in HELLO
provenance, and tool-provider 0.15.12 marks the read tools read_only in
the confined catalog. Regenerated on Linux with
AFT_REGENERATE_CONFINE_VECTORS=1; only catalog-reply.json changed (four
read_only marks and its digest).

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
…ndex.rs ONNX install and hint move only)

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
@socket-security

socket-security Bot commented Nov 18, 2026 •

Copy link
Copy Markdown

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff Package Supply Chain
Security
Vulnerability Quality Maintenance License
Updatedcargo/​cortexkit-role-tool-provider-conformance@​0.15.11 ⏵ 0.15.1310010093100100
Updatedcargo/​subc-protocol@​0.37.2 ⏵ 0.37.499100100100100
Updatedcargo/​cortexkit-role-tool-provider@​0.15.10 ⏵ 0.15.12100 +1100100100100

View full report

ualtinok and others added 4 commits November 18, 2026 18:54
…ling links

Batch 17 limited a macOS worker's reads in the worker pool to its own
checkout. Two things the worker system places beside every checkout broke:

- The build-slot launcher (worktrees/.build-slots/prefrontal-build-slot),
  which every worktree's Cargo config names as the rustc wrapper. The exact
  grant for it expected `.root`/`.cache` files beside it, matched nothing in
  the launcher's real layout, and granted nothing, so every Rust build in a
  Mac mason failed with "Operation not permitted". The launcher's directory
  is now read recursively (the wrapper itself is still listed exactly, which
  the strict profile's RUSTC_WRAPPER restore looks for), and only its queue
  and slot locks are writable. The directory goes in a new `read_dirs` list,
  kept out of the shell write-deny list so the lock grants inside it hold.
- The sibling-repository links in the checkout's container folder
  (`../subconscious`). Their targets were readable, but the lookup stopped at
  the link in the denied container. Each link's own metadata is now
  readable; the sandbox still checks the target, so a link into another
  worker's checkout reads nothing.

The launcher code moves to sandbox_spawn/build_slots.rs, and the
sandbox_spawn.rs pin drops from 10797 to 10719. Native macOS worker_jail
test covers prefrontal's real layout, a sibling link and a link into
another worker; 156 Linux sandbox tests and the Windows check pass.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Treat a cgroup.kill write returning ENODEV as an already-stopped scope only when the cgroup directory is gone. Preserve existing kill refusals and surface syscall diagnostics on failure. The running-withdraw path still signals the process group before attempting the scope kill.

Keep the connection-loss control-frame assertions on the closing socket and on reconnect. Use production polling on reconnect so scheduled polling before route installation cannot masquerade as a scope-end action queued at EOF.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
Workers run the relaxed sandbox profile by default, which builds its grants
in build_relaxed_profile, separately from the strict path's
worker_jail_grants. The previous fix added the build-slot launcher's
directory and lock grants only on the strict path, and its native test ran
strict, so real workers still couldn't read `.build-slots/` and every Rust
build failed with "prefrontal-build-slot: Operation not permitted".

Both paths now take the launcher grants from one function,
build_slots::worker_build_slot_grants. The relaxed path adds them to its
data policy as read exceptions (directory, launcher, locks), write
exceptions (locks only), and a write deny on the launcher. The native macOS
test runs the same build-slot and container checks under both profiles.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
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.

RHEL 8 AFT is crashed in opencode

1 participant