fengzhiyushui/dsh-think-intensity
Backfills reasoning-effort declarations into custom and third-party DSH models so the native Reasoning Effort menu appears, then applies configurable default strategies when no manual choice is made.
Official catalog models (DeepSeek Reasoner, OpenAI o1/o3) already ship with reasoningEfforts declarations and show the Reasoning Effort submenu natively. Custom models do not. This plugin backfills that declaration per configured model so the native DSH UI reveals the submenu — the plugin does not add a separate UI, it activates the one that was already hidden. When a user picks a level manually the choice passes through untouched. When no manual choice exists, the plugin injects the configured default/subagent/compaction level. If a level is rejected by the upstream gateway, the plugin records it in a session-scoped blacklist and retries with the next lower level instead of looping.
Install
dsh plugin --profile desktop add github:fengzhiyushui/dsh-think-intensity#v0.1.0The GitHub form from the README, pinned to tag v0.1.0 (dsh-think-intensity 0.1.0 on npm). The problem it addresses: DSH treats custom/third-party models (StepFun, Zhipu, proxy gateways) as plain text models and hides the reasoning effort selector because those models have no reasoningEfforts declaration in the DSH custom model settings UI. Sending the parameter anyway causes upstream gateways to reject with UNSUPPORTED_REASONING_EFFORT.
Compatibility
MIT. Three configurable levels: defaultLevel (applied when the user has not chosen manually), subagentLevel (for subagent turns), and purposeLevels (for compaction turns). Explicit manual selection in the UI takes priority 100% — the plugin only acts when no manual choice was made. A self-healing fallback automatically blacklists rejected levels inside the session to prevent retry loops. Zero invasive hacks; clean uninstallation leaves no leftover overrides.
Details
- Repo: fengzhiyushui/dsh-think-intensity
- Category: Models & Providers
- Stars: 0
- Version: v0.1.0 (tag + npm)
Alternatives
shengbinxu/dsh-llm-latency · evan7ang/dsh-agent-hints · yusufameri/dsh-t3-model-picker