7.9 KiB
Agentic boundary hardening — 2026-10-12
The implementation review identified seven concrete gaps. All seven have code changes and regression coverage. Native package and sandbox acceptance passed on an installed-system overlay; a fresh installer run remains pending. This is not a new ISO release.
| Gap | Implemented change | Evidence |
|---|---|---|
| Raw OpenCode provider credentials entered logs before catalog parsing | Sensitive output suppresses both streaming sinks, logged bodies and attached command failure bodies; curl ignores user configuration | Unknown synthetic provider credential tested on success and failure; real HTTP fixture suite |
| Investigation evidence could carry registered secrets and service journals/argv | Shared evidence boundary redacts, removes private-path lines and bounds UTF-8 output; service evidence uses only enumerated state properties | Registered service token absent from the second scripted model request |
| Approval digest omitted execution context | Versioned envelope binds id, kind, creation/expiry, root, uid, privilege, source, application, full candidate and material effects; lifetime limited to 900 seconds | Context substitutions and expiry extensions refused after resealing, including recomputed digests against retained approvals |
| Developer jail exposed home and could unmask credentials | Synthetic home, clean environment, private PID/IPC/UTS namespaces, no capabilities, only a selected writable workspace; mask overlaps refused; workspace bound by an open descriptor | Real Bubblewrap denied synthetic private files, environment, host processes and offline network; a post-review workspace symlink swap still bound the original workspace |
| Target writes checked paths and later reopened them | Root and parent descriptors, no-follow component traversal, atomic temporary-file replacement and descriptor chmod; root symlinks refused, hardlinks cannot receive writes | Root replacement, hardlink and concurrent atomic directory/symlink exchange tests |
| Installation verification only proved presence | Portage exact atom/version with explicit source/binary mode; Flatpak ref/commit, Snap revision/version/channel, Nix active-profile store path, AppImage bytes/version checks | Wrong Portage identity/version and wrong Flatpak commit refused; mocked install success/failure transactions remain truthful |
| PID-based per-action locking raced and allowed plan interleaving | Persistent inode plus kernel flock; transaction-wide reentrant ownership through revalidation, execution, verification and outcome recording | Competing acquisition, nested guard lifetime, other-thread exclusion, killed-process recovery, and a worker lease that survives broker death |
Additional enforced limits
The process runner drains stdout/stderr concurrently with nonblocking stdin, limits capture per stream (16 MiB by default), bounds log emission, and applies wall-clock deadlines. Streaming builds retain a bounded tail. Oversized query responses fail. Provider limits/deadlines also apply to sandboxed calls, and the systemd manager receives RuntimeMaxSec so the actual service worker is bounded. Model units have MemoryMax=256M and TasksMax=64. Commands issued under a mutation hold inherit a duplicate lock descriptor; broker death cannot release exclusion while an inheriting worker still runs. A real-process crash regression proves that ownership passes to the worker and ends when it exits. Long build commands retain a 24-hour default; common catalogue, candidate and verification reads use 120 seconds. Service evidence uses 30 seconds and 64 KiB.
Actual negative testing exposed a further gap: the original systemd user unit could read a synthetic file beneath home. The unprivileged model path now uses explicit Bubblewrap filesystem/PID isolation inside the transient user unit. Systemd retains process/syscall/address-family restrictions. Its user-manager filesystem properties were removed because they implied a namespace that prevented Bubblewrap creating its jail. Root model calls retain the hardened DynamicUser system-unit path. Missing tools or namespace failures stop the request; there is no unconfined fallback. Effective uid is read from the kernel, and developer home comes from the account database, rather than PATH or HOME.
Artifact semantics
Flatpak does not support an install --commit option. The reviewed flow downloads without deployment, uses OSTree's native REF@COMMIT pull with the configured remote's signature policy, then deploys with Flatpak --no-pull and checks the active commit. Phoebus now installs dev-util/ostree with its Flatpak lane. Snap installs with --revision. Nix evaluates from an immutable path flake and verifies its active store path. These pins bind the selected application; backend-managed runtime/dependency closures are not claimed to be completely content-pinned. Portage exact-version selection is not an immutable repository snapshot. Portage, Flatpak and Nix were exercised in a disposable installed-system VM; native Snap revision acceptance remains pending.
Primary command references: Flatpak, OSTree pull, Snap revisions, Nix profile install, Portage emerge.
Verification and remaining acceptance
Real host denial tests are intentionally ignored by the portable suite and run explicitly with:
cd installer-rs
cargo test -p phoebus-executor --test isolation_acceptance -- --ignored --skip model_system_unit
# The system-manager probe is run separately as root only in the disposable VM.
Both user probes passed on the development host. The system-manager probe also passed as root in the disposable Gentoo VM. They use only synthetic fixture files, never account credentials. Their fixtures are deleted at test completion. Portable gates pass: 981 Rust tests, 22 VM-harness safety tests, strict Clippy, fmt, repository checks and staged secret scan. Dynamic and static release builds passed. No physical disk was touched.
The native run serial-boot-20261012T105530Z-9c09094e passed all 14 driver cases:
detached generic boot/login, running systemd, developer/user/root isolation,
deterministic history, descriptor-backed Btrfs snapshot, exact Portage install
and removal, persistent lock permissions, Flatpak commit install and removal,
Nix active store-path install, and clean poweroff. It used a separate qcow2
backing overlay; the source-build base image and local ISO are unchanged.
Two preceding attempts exposed fixture errors (password mapping and the KVM
virtualization identifier); neither is claimed as a product acceptance pass.
Initially doctor found 6 GiB free. Compiler-cache cleanup restored sufficient
space and host doctor passed with a tight-space warning. A fresh installation
was not repeated. Native pins/isolation/snapshot behavior were exercised on the
installed Gentoo overlay. The subsequent lock-lease addition has its own real
crash-process regression. Final static binary smoke
serial-boot-20261012T112454Z-47d3cfae passed 8/8 with clean poweroff,
repeating sandbox denials and executing history and a Btrfs snapshot in an
explicitly network-free namespace. Its SHA-256 was
1ea91f6f0ac82d88e9a8c400f510819aefaf73251778b248ef2443fdf18e2b62.
Remaining acceptance: a fresh encrypted installation with descriptor I/O; native Snap revision installation and adversarial metadata drift across the lanes; interrupted native package-worker recovery and broader network-free repair/undo acceptance beyond the verified history/snapshot commands. Older staged approvals must be reviewed again because the envelope digest has changed. Developer tools/auth state outside the selected workspace are intentionally hidden; narrowly scoped explicit grants are future work, rather than silently exposing the whole home directory.