dsh-reasoning-options

GitHub: Scorp1o117/dsh-reasoning-options · npm: dsh-reasoning-options · 中文

A small DeepSeek Harness plugin that automatically adds a reasoning-effort picker to every pi-ai (third-party gateway) model, and auto-injects required routing headers (such as x-opencode-session for OpenCode Go).

Why

1. Reasoning Effort Selection: dsh's built-in DeepSeek models show a reasoning-effort selector in the Web UI because the DeepSeek adapter declares reasoning capability for them. Models configured through llm-pi-ai (OpenCode Go, GOAT, Volcengine Ark, ...) don't — pi-ai only offers effort levels for models that explicitly declare reasoningEfforts, and hand-declared gateway models never do.
2. OpenCode Go Compatibility: OpenCode Go (opencode.ai/zen/go/v1) strictly requires an x-opencode-session HTTP header to route requests and manage prompt cache. Without it, requests fail with 400: {"type":"MissingSessionID", ...}.

This plugin closes both gaps: it scans the llm-pi-ai namespace, adds the full level set (off / minimal / low / medium / high / xhigh / max) and default reasoning: high to models without declarations, and injects x-opencode-session into OpenCode Go provider headers if not already set. Writes go through DSH's native settings pipeline into the Profile patch.

The plugin only gives users a convenient way to pick and ensures gateway requirements are satisfied. Which level a model actually supports is the user's call; the plugin does not judge model fitness.

Desktop install

Use the Desktop-installed dsh command (Application → Manage dsh Command), or the app’s Plugins page. Then install into the Desktop profile:

dsh plugin --profile desktop add dsh-reasoning-options@0.3.1

Restart the Desktop app to load the client bundle. Desktop keeps its profile under $DSH_HOME/profiles/desktop.

Compatibility (v0.3.1)

Verified with DSH 0.1.7-rc.2 (Web) and 0.2.0-rc.2 (Desktop runtime) in isolated profiles. The Desktop app uses its own desktop profile. Other DSH prereleases remain unverified.

Install

dsh plugin --profile web add dsh-reasoning-options

Or mount manually in a profile patch:

- insert:
    - id: reasoning-efforts
      name: 'dsh-reasoning-options'
      config:
        enabled: true

How it works

1. Read the current user layer of the llm-pi-ai namespace (settings.describe() re-reads the profile patch from disk every time).
2. For each model without reasoningEfforts, generate a mutation writing the full seven-level declaration at the exact path.
3. For provider routes without a reasoning default, add reasoning: high.
4. For OpenCode Go providers (baseURL containing opencode.ai), auto-inject x-opencode-session: dsh-session into headers if missing.
5. Writes are schema-validated by pi-ai, persisted, and hot-committed; dsh's native UI/request path takes over.
6. Idempotent and serialized: models and headers that are already declared are untouched, and a scan that finds nothing writes nothing. Only one pass runs at a time, so repeated triggers never contend for the settings file lock.
7. When it runs: on settings/document-updated (a model changed through the Web UI), on app-boot/config-reload, and as a fallback every pollIntervalMs. A trigger only flags the namespace; the write itself runs from this plugin's own timer context. dsh emits the settings event from inside the hot-reload transaction that is applying the change, and a write issued inside that transaction is refused outright (HMR transactions cannot be nested).

Config

Field	Default	Meaning
enabled	true	Set false to stop auto-patching
autoSessionHeader	true	Auto-inject x-opencode-session for opencode.ai gateways
sessionHeaderValue	'dsh-session'	Value for injected x-opencode-session header
pollIntervalMs	30000	Fallback rescan period in ms; 0 disables polling

Want different defaults or wire values? After the patch they live in the llm-pi-ai entry of the active Profile patch — edit them freely; the plugin never overwrites existing declarations.

Notes

• The plugin reads and mutates the llm-pi-ai namespace but does not own it (pi-ai registers it exclusively). All writes use the public settings.update API — equivalent to editing via the Web UI.
• Arrays are replaced wholesale, so every write restates the full models list of each provider it touches. Each write carries the revision it was planned from, so a namespace that moved in the meantime is re-read and re-planned instead of clobbering the concurrent editor.
• The plugin writes the config of the llm-pi-ai row in the profile patch, which is rewritten as a whole (YAML comments inside that row are lost).
• Hand-editing cordis.patch.yml no longer needs a dsh restart: the next poll (within 30s by default) adds the missing declarations and hot-applies the new models along the way. Lower pollIntervalMs to shorten the delay.
• Wire spellings are OpenAI-compatible (low/medium/high/...). Most OpenAI-compatible gateways accept them; if one expects its own spelling, adjust the values in the Profile patch.