Windows qualification addendum#

This addendum records the remaining-gates pass against the frozen recovery candidate. It is a qualification record outside that candidate's public export; no runtime, test, dependency, model selector or packaged file changed. The recovery verdict and its process-crash limitations remain accepted. The earlier readiness report is the historical baseline.

Qualified staged tree: 3a7d24cbf8c8d41c14d88fa6a6df326a6fa37ebb. Base commit: 8b02eba3deb02cc24393421b317644b183114a09. The existing staged-file inventory remains authoritative; this document is the only subsequent staged-file delta. No release was committed or published.

Results and remaining decisions#

The Python 3.11 run excludes the 50 already accepted recovery tests and the one privilege-dependent symlink test: 361 passed, 1 deselected in 139.27 seconds. The unchanged candidate's baseline full Windows suite remains 411 passed, 1 skipped; its installed recovery suite remains 50 passed. Those results are historical evidence, not newly rerun tests. Existing fake-provider benchmark unit tests do not constitute a live model benchmark.

CPython came from the official 3.11.9 NuGet package, with the matching signed Python Software Foundation Tcl/Tk component extracted into private storage. No system Python registration, PATH change or installer execution was needed. This establishes an actual 3.11 interpreter test, not fresh-machine acceptance.

The predeclared measurement used 20 seconds idle, 100 settings transactions, one exact gpt-oss:20b synthetic boundary review (423 input / 122 output tokens), and five seconds after completion. The installed runtime was Python 3.13.14. Controller startup was 0.875 seconds. Controller/worker CPU averaged 0.71% of one core idle and 4.63% active; sampled RSS peaks were 56.6 and 124.8 MiB. The separate hidden dock peaked at 57.3 MiB, with 18.1 ms p95 lateness for its 100 ms timer. Mean settings transaction time was 9.07 ms. Both owned processes exited normally. Instrumentation contributes overhead; sampling can miss peaks.

The older observation used only two seconds idle (0.78% of one core, 43.2 MiB), 1.510 seconds startup and 9.76 ms settings writes. Different sampling windows, model loading and one shared development host prevent a regression or speedup claim. Ollama-name process samples omit differently named runner descendants and GPU memory, and are not a complete inference footprint. No live callback latency or human interaction latency was measured.

Owner decision: specify reference hardware, cold/warm workload, CPU/RSS scope (companion, dock, or entire model runtime), startup/status latency and acceptable idle/active resource limits. Set any new thresholds before a scored run. This pass does not invent thresholds from the observed values.

The local review correctly identified the synthetic inclusive-boundary error. Supervision graded it corrected because its downstream-impact claim was not supported by the supplied snippet. No product patch followed. The isolated job remains locally succeeded with a pending managed review and no Codex session; telemetry feedback does not count as callback delivery, event approval or final verification. Existing GPT-OSS preference and the exact installed qwen3:14b-q4_K_M tag/digest remain unchanged.

Human interface checklist#

Record Windows version, display scale, screen resolution, keyboard, observations and screenshots where useful. Automated native Tk checks do not pass these items.

The live callback's reviewing/completion sequence remains authorization-gated. Do not create artificial approval just to obtain a green status screenshot.

Short fresh-Windows procedure#

Prerequisites: a genuinely clean supported Windows x64 machine/VM with approved Python 3.11 including Tcl/Tk, Git, Ollama, and an already available completion model. No model is bundled; stop if a model transfer/download needs permission. Use an isolated Git fixture with an existing HEAD, never the stable checkout. No paid machine provisioning or security setting changes are implied.

1. Copy the exact wheel, sdist, requirements.lock.txt, examples/companion-task.json and this checklist to a clean folder. Verify both SHA-256 values above with Get-FileHash. Retain a PowerShell transcript.
2. Confirm py -3.11 --version; create .venv with py -3.11 -m venv .venv. Install the wheel using .\.venv\Scripts\python.exe -m pip install "./localagentbridge-0.5.0.dev0-py3-none-any.whl[tray,dev]" -c requirements.lock.txt. Run pip check, package --help, and python -m tkinter using that venv. Record sys.version, sys.executable and localagentbridge.__file__.
3. Create only a private config.local.toml: source_root is the disposable Git fixture, data_dir is new private storage, model is an exact installed tag, and thinking = "low" for GPT-OSS. Use python -m localagentbridge --config config.local.toml companion inventory, then select, create, and submit --run RUN_ID --request companion-task.json. Run companion controller --once without a Codex executable. It starts the detached worker; poll companion status --run RUN_ID until terminal local state. Expect local success and pending review, never automatic verification.
4. Run companion dock and complete the checklist. Close the dock after review. If symlink creation is already allowed, extract the exact sdist to source and run the existing tests/test_bridge.py::test_symlink_rejected with the installed venv from outside the source directory: python -I -m pytest source/tests/test_bridge.py::test_symlink_rejected -q -rs -p no:cacheprovider --basetemp symlink-fixture. Adjust source to the actual extracted root. A skipped result remains not tested. If privileges are absent, stop that check and request explicit owner approval separately.
5. Retain hashes, interpreter/package paths, install/test logs, status and human observations. Stop before init-review or any --codex-exe option unless the separate bounded callback request has been approved.

The private evidence bundle contains exact application callback prompts, payload hashes, CLI version/authentication method, anticipated context, retry limitations and the bounded approval request. It contains no exported credential. No production defect requiring a change was found during this pass. Release qualification remains open for the environment, criteria, human and authorization gates above; the accepted recovery verdict is unchanged.