john-walks-slow/dsh-set-model

dsh-set-model is a purpose-built DeepSeek Harness plugin (a cordis plugin targeting DSH 0.1.2-rc.1) that gives the root Agent two tools — set_model({ provider?, model?, reasoningEffort?, reason }) and list_models({ provider? }) — so it can autonomously switch provider, model and reasoning effort by task difficulty, with reason required because each switch invalidates the session KV cache. The plugin writes model/selection events into the session (the same channel used by the GUI model selector, so the Web UI state stays in sync), and appends one cache-safe line to the Runtime Context snapshot each step, such as [Current active model: deepseek/deepseek-chat · reasoning: low] or, in Plan mode, [Current active model: deepseek/deepseek-reasoner · reasoning: high] [Plan Mode Active], letting the model self-check without an extra tool call.

It is intended for users who want the Agent to keep cheap high-throughput settings for routine work yet escalate to a high-reasoning model such as deepseek-reasoner when facing hard concurrent or multi-step reasoning tasks, with no manual config toggling in between. Plan mode handles the escalation automatically: on /plan on the plugin snapshots the current daily model and switches to the configured planModel; on exit_plan_mode approval or /plan off it restores the original model when autoRestorePlanModel is true (default), and any failure only emits a warn, never interrupting the step. Installation is dsh plugin --profile web add dsh-set-model, or add github:john-walks-slow/dsh-set-model for source, where the first run is intercepted by pnpm ≥10 and the prompted package must be added to ~/.dsh/profiles/web/pnpm-workspace.yaml under allowBuilds before rerunning.

Dependencies are pinned to @deepseek-ai/cordis 4.0.2, @deepseek-ai/dsh-agent, dsh-llm, dsh-session, dsh-tools 0.1.2-rc.1, zod ^3.24.2, and Node ≥ 22.5; the plugin performs no external network calls and no direct filesystem writes, routing all model resolution through DSH core ctx.llm.resolveCallConfig. Safety guardrails reject switches when the session token pressure exceeds the target model contextWindow (prompting the user to compact, e.g. via clear_mind), honour the allowedProviders whitelist, validate every candidate combination through DSH's adapter registry, and restrict the tools to the root Agent only — set enableAgentTools: false to disable them globally. License is MIT; local dev is npm install && npm run build && npm test, and releases run npm run release followed by the WebAuthn fingerprint script, with npm view dsh-set-model version as the post-publish check.