WORK_LOG.md

Work log

2026-10-28 — Explicit numeric render-node policy

• Pond recipe wiring is independently accepted: RED 40710876277, GREEN 40713790188, source 9b0d20e, no P1/P2 review finding. Continue the next independent CLI prerequisite, not persistent remoted-owner research.
• Specify opt-in numeric render groups in the adapter and a systemd generator symlink to the same Rust binary. No auth/account mutation, host device chmod/chown, service start/enable or physical card/TTY grant is permitted.
• Regression ec36837 observed the intended install/profile RED in all three GitHub jobs of 40715702736; existing ordering tests still pass. Runtime fixture cases are present but were not reached after that earlier failure.
• Implement descriptor-pinned character-node inspection and ephemeral, additive service group drop-ins. A missing mapped render node is a no-op; malformed selected nodes are rejected before service policy is written.
• All compilation, formatting and executable validation remain remote. Source implementation is not yet accepted device or desktop behavior.
• Implementation run 40717635975 reports one formatting-only import-order difference in main.rs; apply the runner's exact edit and require a fresh full run. Do not infer whole-gate acceptance from successful sibling jobs.
• GREEN: full run 40718742320, source 30f7651, passes stable, MSRV and static/runtime gates: 49 Rust tests per required build and 20 chroot cases.
• Final review: fresh read-only review found no P1/P2. Ruling: clarify the group-wide permission boundary rather than imply per-device isolation; same-GID resources may also be accessible, so an operator must explicitly accept that group policy. No extra card/input/TTY groups are discovered.
• Final: minor (deferred): independent owner-only/transport-only opt-in test.
• Final: minor (deferred): single-output-directory and empty early/late tests.
• The accepted mechanism can now be selected by the Pond profile. Actual service credentials/GPU use, full Pond shell, persistent Niri owner and platform proxy/optional VT acceptance remain open. No OCI build is resumed.

2026-10-28 — Desktop dependency gate remote verification

• Run 40677322538 passed the minimum-Rust and static-runtime jobs, including the new platform-service process tests. The stable job stopped at two runner-reported rustfmt differences; apply exactly those source changes and require a fresh full run. No build, formatter or test ran on this host.
• Retrieved the LightOS lifecycle, seed-injection and remote-authentication memories again. Keep platform-owned accounts and authentication unchanged; a remote graphical session does not require a physical VT.
• Full run 40678793875 passed. Independent review nevertheless found prefix and template drop-in names bypassing the concrete-service restriction, plus stale owned drop-ins surviving a profile change. Add public-CLI regressions for both before changing the implementation; observe their RED in Actions.
• Run 40680435930, MSRV job 122789498149, observed all three new failures: an accepted broad unit name reached the staging-root check, stale files remained after A-to-B, and a modified stale file was accepted. The ordinary repeated-install case still passed. Apply the two runner rustfmt changes too.
• Reject prefix/template names and reconcile stale owned files before replacing the profile. A bounded staging reader pins an O_PATH-verified regular inode before data access, preserving the no-symlink/no-cross-mount boundary. Require fresh remote GREEN and independent closure; do not resume image packaging.
• Fix run 40681691642 passed MSRV and static-runtime, including all four platform-service tests. Its stable job reported three formatting differences; apply exactly the runner diff and run the entire remote matrix again.
• Exact 16083e02db95aeb74683e697a26458a5b9866e4f, run 40682679574, passed rustfmt/clippy/tests/docs/build, Rust 1.89 and static-runtime acceptance. Independent re-review closed both P2 findings and found no new P1/P2 in this extension. Accept the ordering sub-gate only: device policy and native Pond/Niri session support still keep the overall desktop gate open.

2026-10-28 — GitHub-only execution policy

• The user now requires all downloads and builds on GitHub. From this point, dependency/rootfs downloads, Rust compilation and runtime acceptance run only in GitHub Actions. Local work is limited to source editing, static review, SSH source pushes and workflow/log inspection; no artifact download is needed.
• Documented remote workflow dispatch without adding GitHub-specific behavior to the reusable CLI mechanisms. Earlier local evidence remains historical; it does not authorize further local builds or downloads.
• Confirmed the completed CLI gate's documentation head daaf572 also passes all three GitHub CI jobs in run 40663154259, after the source/runtime gate at cbf327f passed run 40662397249.

2026-10-28 — Initial implementation

• Fixed the domain boundary before implementation: generic fetch/inspect/clean, an explicitly named LightOS adapter, and separate Pond release/build recipes.
• Implemented verified atomic fetching, read-only archive inspection and explicit offline cleanup policies; exercised them against temporary fixtures.
• Added lighos-init post-setup integration without taking ownership of platform account creation, credentials or SSH. Tested failure retry and idempotence.
• Downloaded and verified the Pond artifact, then inspected it without execution; the payload remains excluded from version control and public distribution.
• Read-only review identified root-ancestor symlink resolution and mount-check TOCTOU risks. Added three regression tests and observed all three fail on the original implementation before changing the filesystem implementation.
• Replaced ambient reopening and generic recursive deletion with pinned directory descriptors and openat2 no-symlink/no-cross-mount traversal. Removed cap-std rather than adding another filesystem abstraction alongside it.
• Re-ran targeted tests: the mount-boundary probe and all nine cleanup integration tests pass. The mount test only queries existing procfs; it creates no mounts.
• Documented Linux requirements, the isolated-staging trust boundary and why distribution-neutral tools are not universal operating-system adapters.
• crapper, dryer and mutator are not installed in this environment; no result from these tools is claimed. Rust formatting, clippy and behavior tests provide the available automated checks.
• First full local verification passes: fmt check, clippy with warnings denied, 26 Rust tests, rustdoc and release builds of all four binaries.
• Git had no author identity. Configured only these two repositories with the authenticated GitHub account's public name and noreply address; global Git configuration is unchanged. Aligned their initial branch names with main.
• Follow-up review closed the two original findings and identified a special-file marker risk: opening before checking the file type can invoke device behavior. A Unix socket fixture reproduces the wrong I/O-first error without any device access. Changed validation to pin with O_PATH, require a regular inode, then read that exact inode through its held procfs descriptor instead of its name.
• Final local verification passes with 27 Rust tests, fmt check, strict clippy, rustdoc and all release binaries. Read-only re-review closes all three safety findings and reports no new blocker in the fixes.
• GitHub API access works, but ordinary Git HTTPS and a bounded HTTP/1.1 retry fail or time out. Publication is pending a working authenticated transport; local commit success must not be reported as remote push success.

Source publishing and CI results are recorded only after their commands succeed. OCI construction and a real LightOS/Pond boot remain separate acceptance gates.

2026-10-28 — Source publication over SSH

• The user explicitly requests SSH transport. Existing SSH configuration already maps GitHub to ssh.github.com:443; authentication succeeds as czyt using existing keys and known hosts. No SSH configuration or key changes are needed.
• Changed this repository's origin to git@github.com:ca-x/oci-kit.git and pushed initial source commit 2ef0ba1 to main; Git confirms the new remote branch and configures origin/main tracking.
• A missing ssh-agent did not mean that on-disk identities were unavailable; inspect the existing SSH configuration before diagnosing authentication.
• The earlier HTTPS failures do not block this SSH push. Remote CI and image construction still require their own evidence; publication alone proves neither.

2026-10-28 — CLI completion gate

• The user requires full CLI completion before starting the image project, continuous execution afterwards, and independent agents for review. No image workflow has been dispatched. The follow-up plan is docs/superpowers/plans/2026-10-28-cli-completion.md.
• Existing helper/fixture tests and source CI do not close process-interface, minimum-Rust or static-binary acceptance gaps. These are now explicit gates.
• A read-only adapter audit was requested, but the reviewer could not start due to model capacity. Its slot was closed. Review remains pending, not approved.
• Recorded the CLI-first ordering and continuous SSH/review workflow in Nowledge preference memory 64fa168c-9cf9-4f88-a874-00a8c5233986.
• Task 1: observed three real inspector failures: absent requested text silently succeeded, and corrupt gzip/zstd checksums beyond the tar end marker escaped validation. Added regressions, observed all three fail, then drained the decoder and required every requested text path. All seven inspector tests now pass; the truncated-member test already passed and required no speculative fix.
• Added eight real generic CLI subprocess tests covering help/version, verified download/provenance/cache corruption, interrupted/digest/size/HTTP failures, source rejection, three archive formats/text errors, and marked cleanup with dry-run/apply/repetition/link safety/all-rule preflight. The initial cleanup test expected the wrong diagnostic wording; corrected the test, not production.
• All 39 workspace tests and strict clippy pass after this slice. The adapter's process contract, MSRV/static acceptance and whole-change review remain open.

2026-10-28 — Initialization runtime acceptance

• Generic CLI phase a15b20c is pushed over SSH; its source CI run 40659011161 succeeded. Pond packaging remains untouched.
• Added four adapter subprocess tests for help/version, deterministic dry-run, profile rejection, root refusal and repeatable offline installation of the actual executable/profile/unit/enable link. Observed explicit failure emitting a success plan, then changed application to emit one JSON result only after success, with changed; failure stdout is empty.
• Built real musl binaries locally using the installed Zig compiler and Rust's self-contained linker. Direct zig cc initially failed because cc-rs supplied a Rust target triple; a disposable workspace wrapper translates that argument. This is a local toolchain workaround, not a production dependency; CI uses musl-gcc. No GNU compiler is passed off as a musl compiler.
• Added a static, non-GPG pacman-key fixture and a process harness that uses rootless bubblewrap on this host, or disposable chroot on CI. It runs the real initializer as namespace root, mutating only disposable roots with fixture credentials. It does not change host system configuration/services, execute Pond rootfs programs, or perform network operations. Chroot is a trusted-code test harness, not a sandbox for hostile root programs.
• The first eight runtime checks caught child stdout contaminating JSON on success and failure. Observed failures, redirected child stdout to stderr, then re-ran the real static CLI successfully.
• Independent read-only contract audit confirms that account/seed/SSH lifecycle remains platform-owned, and that keyring is the specified adapter action, not complete desktop initialization. The audit found two important issues.
• Confirmed in the installed pacman-key source that populate can ignore import failure while --add propagates it. A failing-import/successful-populate fixture reproduced a false success state. Now explicitly add each preflighted keyring before populate; failures do not commit state.
• Observed unsafe state directory/link acceptance and blocking FIFO state reads. Now pin and validate root-controlled state ancestors, reject symlink/special lock/state files before I/O, bound state reads, use nonblocking no-follow opens, and sync the state directory after atomic replacement. Hostile concurrent root processes are outside the cooperative-lock contract.
• All 14 isolated real-CLI runtime cases pass, including failure retry, changed profiles, unchanged-run idempotence, real process lock contention, corrupted state, failed import, unsafe modes, symlinks and FIFO rejection. Added a ninth generic process case for links/devices/FIFOs without extraction. Its first tar fixture omitted UID/GID; fixed that invalid fixture, not production parsing.
• Full local workspace verification now covers 44 Rust tests. A fresh whole-code review and CI MSRV/static gates are still pending; this is not yet permission to begin image packaging.

2026-10-28 — MSRV and static acceptance gates

• Initialization phase 2052285 is committed and pushed over SSH.
• Installed Rust 1.89.0 without changing the default toolchain. The locked workspace tests pass on the declared minimum; musl release tests also run on 1.89.0, including all real CLI subprocess interfaces.
• All four 1.89.0 musl binaries pass readelf checks for no INTERP/NEEDED entries and execute help/version. Their actual initializer and a 1.89.0 static fixture pass all 14 rootless isolated runtime tests. This is execution evidence, not merely a successful cross-link.
• Added CI jobs for MSRV tests, full musl release tests, linkage/execution checks and real initializer chroot acceptance. No image/rootfs/artifact upload step is added. Added docs/cli-acceptance.md as the per-command evidence/limit matrix.
• Workflow syntax passes actionlint 1.7.7 from its existing installation; the bare shim initially had no selected version, so no global mise setting was changed. Shell syntax and fixture Rust formatting also pass.
• Remote CI on these new gates and the fresh whole-change review remain pending.

2026-10-28 — CLI completion accepted

• Static/MSRV phase cbf327f0fa20b4e87b491b14aeed373f20cf223b is pushed over SSH. GitHub run 40662397249 is successful; inspected the individual jobs and logs: rust, msrv and static-runtime all succeeded. Stable/native, Rust 1.89 and musl release each passed 44 Rust tests; static checks executed all four binaries; the disposable chroot ran all 14 initializer cases successfully.
• Fresh independent whole-change review of baseline f7b5f12 through cbf327f reports no Critical/Important/Minor finding. It re-ran 44 native tests, strict clippy/formatting, 14 bubblewrap cases and all four static binary checks. Additional temporary probes rejected writable ancestors, differently owned state, Unix sockets and oversized state. No source/index/host setup changed.
• Both important contract-audit findings are independently closed. The evidence matrix now closes every documented CLI function and states the remaining platform/external-command limits. Completion is not inferred from source-only CI, a cross-link, or a reviewer that could not start.
• All four tasks in docs/superpowers/plans/2026-10-28-cli-completion.md are complete. Image packaging is now eligible as the next phase; no image/rootfs publishing or Pond build happened before this acceptance. The accepted scope excludes unimplemented generic account/locale features and physical desktop claims rather than pretending that they work.
• Removed only the empty nested directory left by the earlier wrong-CWD patch; it contained no files. The final review slot is closed and its workspace can be removed after this durable log is saved.

2026-10-28 — Platform contract and reopened desktop gate

• User requests further LightOS underlying-implementation memory retrieval. Retrieved inner runc/outer core observations, three-stage seed/runtime config injection, admin-proxy credential ownership and optional physical VT records. Read only the current setup script, never the actual seed/auth environment.
• Ruling: desktop does not imply a physical VT. vt=false remote instances must still run a graphical desktop. Do not invent a registration API, bypass the admin proxy or overwrite platform credentials. The Niri session still needs real support; the remote program's Hyprland owner is not a substitute.
• Documented those boundaries in docs/lightos-platform-contract.md. First extension: explicitly order configured platform services after successful post-setup initialization, without starting/enabling or replacing them. A skipped systemd Condition is not a successful dependency readiness gate.
• Added real offline CLI installation regressions and invalid-unit cases before production changes. All executable tests remain GitHub-only. This reopens, not silently closes, the original toolkit completion gate.
• Run 40676051140 observed the new valid-profile CLI installation fail in MSRV and static-musl tests, before the feature existed. The formatter also requested two test layout changes; use that runner-produced diff locally without executing rustfmt here. Tightened negative tests to require the platform-service validation diagnostic, not an unrelated absent-root error.
• Implemented optional post_setup_services, validated names/count/duplicates and obvious initialization cycles, and installed only ordering drop-ins. Changed the marker Condition to Assert for fail-closed dependencies. Empty lists are omitted from serialization to preserve original profile identity. GitHub GREEN and independent review are pending; other desktop work stays open.