dsh-thinking-level-override

LIMITATIONS

Known limitations

- Governs conversation requests dispatched through the agent loop's `agent/request` waterfall, including subagents. Direct `ctx.llm.stream()` callers with their own configs — session titles, compaction — are unaffected. - Cannot add reasoning capability an adapter does not declare: a model whose catalog entry carries no reasoning metadata sheds every effort under `clamp`/`drop` and fails under `fail`. Wire-level preset repair (offered levels and their spellings) belongs to the adapter's own configuration — for `@deepseek-ai/dsh-llm-pi-ai`, the `modelOverrides` and `reasoningEfforts` fields in its settings section. - The first matching rule wins; overlapping rules do not merge. - Under `onUnsupported: fail`, forced/mapped/defaulted efforts apply without capability validation, so a misconfigured force still fails loud at the LLM seam. - The settings section is edited through the settings document (the `thinking-level-override:` key in `settings.yaml`) and is discoverable via `settings.describe`. The in-page editor additionally needs the harness exposure-allowlist patch documented under [Web settings page](#web-settings-page); without it the section remains fully live-editable through `settings.yaml`.