2.4 KiB
Row253 implementation plan
Goal: migrate the two host scripts to the private tester socket without widening their unit or root execution boundary. Spec: DESIGN.md, approved by duckworth-0100 with C1-C5; Sage2173 approves the two design decisions. Execution: Gander implements inline; Duckworth reviews the final manifest.
Constraints: four product/test files only; no edits to the brief's legacy paths; no tests before Sage's slot, including patterns. Static inspection precedes the full slot request. All tests use the existing scratch/fake seams.
- Task1: write factory/acceptance socket fixture and refusal tests first. Add config/marker and isolated probe fakes; assert target-uid argv, strict parsing, all C1 path variants and C4 descendant cleanup. Preserve baseline functional/P checks while changing only the intended socket outcomes.
- Task2: implement the marker parsing, bounded target-uid query, strict bash parser, component/no-follow location checks and parent preflight in installer. Preserve hardening and update its header, including sudo dependency.
- Task3: migrate saved args and P8/P10. Implement metadata snapshots and the bounded isolated system-Python peer observation; keep P10 unnamed with explicit observed/failed evidence and same-uid residual. Update header and remove the obsolete accepted-mos-1101 mark. Keep F11 default DATA_ROOT behavior unchanged.
- Task4: static audit every remaining retired-port occurrence; prepare unique M1-M33 including C1 middle-symlink/sibling-prefix, C2 character, C3 parser and C4 process-group checks. Request the slot once the candidate static check passes.
- Task5: on Sage's grant, execute RED cases against isolated prior behavior, GREEN candidate checks, CLI/WebUI whole, all shell files and the checks. Record counts/times, clean restoration, leak sweep and exact frozen manifest. Return the slot, then queue review only in a coordinated write gap.
Review focus: prefix collisions; intermediate symlinks; forged output lines; nonzero query with valid-looking JSON; ignored signals/leftover descendants; kernel versus filesystem inode; wrong Console peer or metadata replacement.
Ruling: tests are written first, but their execution waits for the user-required slot after static preparation. No pre-slot test or passing claim; RED/GREEN execution is an explicit part of the granted verification, not claimed in advance.