You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Launcher test suite: trusted build identity for executables it runs, and run the suite in CI #255
The launcher test suite only runs programs it can identify as testing builds, and it refuses non-testing programs before they start (PR #253). Two hardening steps were deliberately left out of 0.6.5 and are tracked here.
1. Trusted build identity for executables the suite runs
Today a program counts as a testing build when it carries the testing-only named resource, the testing hook strings inside its sections, and a matching PE checksum. That rejects release builds, modified copies and the runnable canaries. It is not a provenance proof: the PE checksum is unkeyed, so a binary crafted on purpose could carry the same resource, strings and a recomputed checksum.
Proposal: have the testing build write an identity the suite can verify independently of the file itself, such as a build manifest with the full SHA-256 of each executable produced in that build (launcher, restorer, uninstaller, canaries), checked before every launch, or a keyed marker. Mixed or changed images should fail. Add a negative control with a runnable canary that carries the resource, the strings and a valid recomputed checksum, and show that it is refused on every guarded launch path.
2. Run the launcher suite in CI
CI builds the testing launcher and runs the core tests, but it does not run run_launcher_tests.py. The suite's integration checks can regress while CI stays green: the runnable-canary refusals on each entry point, the inherited-override stripping, the helper-permission negative control, and the edition/junction temp-folder checks.
Proposal: run the default launcher-suite cases on the stand-in payload in CI, including both launcher entry points (--launcher and --no-direct-launcher) and the helper and temp integration checks, with failures failing the job.
Two more items from the #253 review rounds, tracked here rather than in 0.6.5.
3. Independently trusted canary hashes
The suite now starts the two canary programs only after checking them against the SHA-256 list the testing build writes. That list is read from the same folder as the canaries, so a replaced canary together with a matching replaced list entry would still pass, as would a wrong canary that an incremental build hashed after linking. This is the same question as item 1: the hashes need a source the suite can trust independently of the launcher folder (for example the build manifest or keyed marker proposed above).
4. Exercise the helper-guard negative control in real-runtime builds
The negative control for the RR helper permission guard runs only when the installed runtime is the stand-in. With a real embedded runtime the control is skipped, so that build never exercises the guard. The control itself is safe. The gap is coverage: run it against a separate fixture with an independently verified non-loadable runtime, even when the launcher embeds a real one.
Scope note
Further findings of the kind "a link or file planted in the contributor's own work folder or machine" also belong to this issue, not to a blocking review of the suite.
Activity
phantoslayer21 commentedon Nov 16, 2026
Two more items from the #253 review rounds, tracked here rather than in 0.6.5.
3. Independently trusted canary hashes
The suite now starts the two canary programs only after checking them against the SHA-256 list the testing build writes. That list is read from the same folder as the canaries, so a replaced canary together with a matching replaced list entry would still pass, as would a wrong canary that an incremental build hashed after linking. This is the same question as item 1: the hashes need a source the suite can trust independently of the launcher folder (for example the build manifest or keyed marker proposed above).
4. Exercise the helper-guard negative control in real-runtime builds
The negative control for the RR helper permission guard runs only when the installed runtime is the stand-in. With a real embedded runtime the control is skipped, so that build never exercises the guard. The control itself is safe. The gap is coverage: run it against a separate fixture with an independently verified non-loadable runtime, even when the launcher embeds a real one.
Scope note
Further findings of the kind "a link or file planted in the contributor's own work folder or machine" also belong to this issue, not to a blocking review of the suite.
Launcher suite: green on a fresh clone, inside its own isolation (#253)