[vladimirrott]

Summary

I had every maintained page reviewed against the source, verified each finding in the code myself, and fixed 30 documentation files. The ones that change what a reader would do:

• docs/vm-daemon-setup.md told you to write the vsock token into your own home. The packaged daemon runs as sysknife with HOME /var/lib/sysknife, so it never read that file, and the page never switched the daemon to vsock at all. It now writes /var/lib/sysknife/.config/sysknife/token (owned by sysknife, mode 0600) and adds the SYSKNIFE_LISTEN_URI=vsock://:9734 drop-in. The same page said make install handles rpm-ostree on its own; it refuses until you pass the /etc overrides, and still stops at Fedora Atomic: make install has never completed, and it fails halfway #301.
• docs/mcp.md: sysknife_execute halts on any status but succeeded, needs_reboot and rolled_back included, not only on failed or unknown.
• SECURITY.md: the intent screen matches credential values, not bare words such as "password", and the minimum role comes from the action's risk tier, not a hand-written match.
• docs/automatic-rollback.md called signed status transitions future work. They exist, and audit verify checks them on SQLite. It also gains the one case where the daemon skips a rollback on purpose.
• docs/security/audit-advisories.md listed 20 accepted advisories. Two remain: RustSec withdrew the gtk-rs set on 2026-08-14 and the other eight crates left the lockfile.
• docs/introduction.md, docs/quickstart.md: the daemon runs as the sysknife system user with scoped sudo and polkit grants, not as root. The sample plan now matches what print_plan prints.
• Contributor docs: clippy everywhere now carries --all-features --all-targets --locked as CI runs it; npm run tauri (no such script), the semantic PR-title check (no such workflow) and the e2e PR label (the smoke job is dispatch-only) are gone; the story harnesses' default set is described as it is in the scripts.

Smaller fixes cover docs/architecture.md (vsock and the CLI/MCP path, the sysknife-daemon-test crate), docs/ubuntu-action-reference.md (SnapInstall hold path, UfwStatus numbered), docs/distro-support.md (helper and grant counts, the gate on hard-fenced reads, the Blocked tier), docs/typed-actions.md, docs/cli.md (Ctrl+C exits the REPL), the release docs, both user-story documents, ROADMAP (two shipped items removed), and dated notes on ADR 0002 and 0004.

Two non-doc changes ride along:

• Cargo.lock: event-listener 5.4.1 → 5.4.2 for RUSTSEC-2026-0221. Unlike every other accepted warning it reached the daemon and the CLI, through sqlx-core. cargo audit reports it as unsound rather than a vulnerability, so CI stayed green. cargo audit now shows only glib and proc-macro-error, both reachable from the shell alone.
• A doc comment in dispatcher.rs named credential_argv_position, renamed earlier to credential_argv_spec.

Related Issue

None filed; this is a maintenance sweep ahead of 0.26.0.

Validation

• Tests added or updated: none needed; no behaviour changes.
• Documentation updated: this PR.
• Security impact considered: the event-listener bump removes the only accepted advisory on the trust boundary.
• Trust boundary preserved (daemon remains the only privileged executor).
• CI passes: pending on this head.

Locally, scripts/ci-local.sh --no-postgres passed fmt, clippy, doc, the frontend, cargo audit, markdownlint, both link checks, yamllint, shellcheck and the release tests. The claim screen caught a "26-story" figure I had written, which no evidence file backs, and I replaced it with the script's own default. The one nextest failure was #604, the load-dependent planning-timeout flake, which is unrelated.

Notes for Reviewers

I left dated documents as history: the red-team report, the 2026-04-25 gap analysis tables and ADR 0001. The Ubuntu story audit gained a note saying its remediation has since landed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UpyrhbDTW1gBwzGAC44GYU