Verification evidence — 2026-10-06¶
Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.
Model working environment delivery (2026-10-06)¶
The full sequential Debug matrix passed all 12 projects: 5,475 passed / 0 failed / 0 skipped (including 1,193 GUI tests). Recorded from the feat/model-working-environment branch after commits 1f9971f1 (Task 7 closure) and 51a264db (review fixes), plus the matrix-closure commit that carries this record. The delivery was verified offline only; no live provider, desktop automation, native keyboard/resize, or token/cost measurement was run. The Python guards passed 178 tests with one skip.
All twelve declared workflow scenarios now have a registered offline executor with fresh evidence under .superpowers (task-7-evidence): Vietnamese text, ten-turn constraint sequence, restart continuation (renewed approval identity, exactly one execution), selected handoff (marker-present target vs marker-absent control), independent read batch, UTF-8 read paging, sparse grep paging, terminal middle-output paging (explicit approval, UTF-16 cursors, single execution), dependent three-file edits with hash-bound receipts and a fixture verification run, edit recovery, failed-test repair (fail → read_output diagnostics → edit → identical rerun), and explicit denial without bypass. Each scenario asserts client-readable receipts, ordered identities, and the exact workspace file set; they prove wiring and effects, not model quality.
The full matrix surfaced and closed several latent defects that focused filters had missed: a culture-sensitive duration/size rendering in the CLI TUI and GUI (now invariant), a stale prompt-composition golden (Task-2 wording), two test fixtures that predated the work-state-on-create and v16-schema contracts, an image-only prompt that failed run creation on an empty work-state goal, and a Task-5 regression where the prepared-body guard failed every provider retry closed as invalid_request (fixed with a genuine RED/GREEN regression; the two native Anthropic loop tests now pass). Independent formal Task 7 and whole-branch reviews returned no load-bearing findings; their four non-blocking items were fixed or recorded. One transient agent_store_unavailable 503 flaked once under parallel load and passed isolated and in the recorded green run.
build.bat (Release) passed all 12/12 stages on 2026-10-06 and refreshed the self-contained flat distribution with the four executables side by side. Managed-runtime smoke, native GUI acceptance, and live-provider quality remain not run.
Update (2026-10-06, later): the selected saved-work handoff feature was removed by owner request (composer button, /handoff, run-detail work_state_* fields, and its evaluator scenario). The evaluator now declares eleven scenarios; the remaining executors are unchanged, and the .NET matrix and Python guards were re-run green after the removal.