[Bug / patch available] Switching models resets the last reasoning effort / 切换模型后推理强度被重置

[kermars39-web]

Summary / 问题

在 DeepSeek Harness 0.2.1-rc.2 中，为模型选择推理强度后，切换到另一模型再切回来，会恢复模型目录的默认值，而不是该模型上次选择的值。希望原生 composer 模型菜单与 /model 都能按 provider/model 记住上次成功选择。

Selecting a reasoning effort for model A, switching to B and back to A resets A to its catalog default instead of its last selection. This is a user-preference recall request, not a request to carry A's effort blindly to B.

Reproduction

1. Pick a model with several advertised reasoning levels.
2. Choose a non-default effort, e.g. Max rather than High.
3. Switch to a different model, then switch back using the composer or /model.
4. Observe the default effort being selected again instead of the previous choice.

Expected: restore that provider/model's last successful supported choice; use the provider default for an unseen model or when the remembered capability is no longer available. Explicit effort changes must continue to work, and failed selections must not overwrite confirmed preferences.

Patch available / 已有修复分支

• Branch: https://github.com/kermars39-web/deepseek-harness/tree/fix/remember-model-effort
• Compare: master...kermars39-web:deepseek-harness:fix/remember-model-effort
• Commits: kermars39-web@83b1a150 and kermars39-web@18585e01

The patch keeps browser-local preferences on the model-directory resolver, shares restoration between both native entries, validates against current advertised capabilities, and only remembers successful choices. It includes regression tests and the English/Chinese README pair. Maintainers are welcome to review or cherry-pick it, or suggest a preferred implementation.

Tradeoff: recall uses localStorage and is limited to the same browser/origin; it is not cross-browser synchronization. No Session format, credential or permission changes.

Verification

• Model-selection suite: 84 tests passed.
• GUI regression: 643 files passed; 10,772 tests passed; 1 skipped (pnpm run test:gui --maxWorkers=2).
• Full build: 9 gates passed; pre-push typecheck: 5 gates passed; scoped type-aware lint and named translation pairing passed.
• Documentation aggregate: 42 passed; one local untracked change-scope report triggered the repository-reference scanner. Moving that report outside the repo made that check pass separately; not claiming a single all-green documentation aggregate.
• Not yet validated in a real browser/desktop flow or installed into the daily client. test:web was not run under the local browser-launch restrictions. Automated tests do not establish full UI acceptance.

Why a Discussion instead of a PR

I attempted to open the existing fork branch against deepseek-ai/deepseek-harness:master, but GitHub returned kermars39-web does not have the correct permissions to execute CreatePullRequest. No PR was created. Issues are disabled, so I am posting here through the enabled Discussions channel rather than changing identity or permissions. Please advise on the preferred contribution route.

Related but distinct: #9345 asks for pi-ai per-model default configuration; this report concerns restoring a user's last native-picker choice. The separately owned Codex speed-persistence fix has already been submitted to the subscription plugin upstream: V1ki/dsh-plugin-subscriptions#162 .