Skip to content

audit: full-project Adaptive run on Windows, macOS and Linux — 12 confirmed defects - #6

Draft
Dimkox wants to merge 2 commits into
masterfrom
audit/adaptive-runtime-addendum-20260910
Draft

Dimkox wants to merge 2 commits into
masterfrom
audit/adaptive-runtime-addendum-20260910

Conversation

@Dimkox

@Dimkox Dimkox commented Sep 10, 2026

Copy link
Copy Markdown
Owner

Verdict: RED — evidence, not repairs

Full Russian report with reproduction, preconditions and acceptance criteria

Completed three-OS run 34521506460

Audit the entire 81-file GitMir baseline 0c9db4635407cb30faab7a3f1887634db097324e using the unmodified external Adaptive preflight engine at be752872f3e5a9d6fe179872d9c8bdaec4338238, plus actual HTTP/relay/installer probes on disposable fixtures. Executed driver: 69fa4934f0f41d46acc23bb5d75261e956aa53d9; the following commit adds only the report.

Actual results

Native runner Original Google Cloud tests Runtime probes Full audit job
Ubuntu x64 20/20 pass 11 pass / 10 fail failure
Windows Server 2025 x64 20/20 pass 11 pass / 9 fail / 1 Linux-only skip failure
macOS ARM64 20/20 pass 11 pass / 9 fail / 1 Linux-only skip failure

TypeScript, JS/TS/shell/PowerShell syntax, npm dependency audit and tracked-source integrity checks passed on all three runners. Installer conflict probes are additional to the runtime totals.

Confirmed product defects

Twelve distinct behaviors, not a sum of duplicated wrapper/report failures:

  • Windows installer recursively deletes a pre-existing user-owned skill directory; POSIX installer instead reports success while leaving the intended skill unavailable.
  • Linux project path is interpolated into shell execution.
  • Host-prefix allowlist bypass; untrusted preview HTML served on dashboard origin; mapped-IPv6 loopback proxy bypass.
  • Twelve successful same-title task submissions overwrite one filename.
  • Incoming relay model follows a pre-existing symlink outside the project; local mode sends queue content despite its displayed privacy promise; completed-task replay recreates todo after fresh module state.
  • JSON null is treated as an internal error; a regular file is accepted as a project directory.

Each finding has a bounded, inert reproduction and explicit preconditions in the report. There was no real cloud access, hosted relay traffic, victim-browser attack or SDK/registry repair.

Framework/environment findings — separate from product defects

On POSIX the raw PR verifier rejects ambiguous local main/master refs from the runner checkout. On Windows it aborts on os.killpg; the factory installer also fails closed because no-follow directory primitives are unavailable. Full-tree secret scan flags a reviewed artificial Google Cloud privacy-test fixture, not a confirmed live credential. Architecture/governance are not configured, and no independent Trust CI or model-review attestation is claimed.

Change boundaries

Exactly five paths relative to the baseline: one additive audit-branch-only job in the existing workflow, two probe files, a plan and the report. No product source changes, factory changes, merge, deployment or release. Existing Google Cloud tests remain intact. This PR is a draft and does not claim the application is fixed.

The concurrently advanced audit/adaptive-full-project-20260910 branch was preserved; this addendum did not force-push over it. Evidence artifacts currently expire on 2026-09-24; report, source SHAs, run links and artifact digests are recorded in Git history.

…nd native probes

Keep application sources unchanged. Preserve existing Google Cloud tests and add an audit-branch-only three-OS job with disposable HTTP, relay and installer fixtures. Run the unmodified external factory verifier plus full tracked-file scans. This is preflight evidence, not independent Trust CI or merge approval.
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.

1 participant