dsh-thinking-efforts

Automatically add off / low / high / max reasoning efforts to every model saved under a custom llm-pi-ai provider in DeepSeek Harness (DSH), so the Composer model selector shows the thinking-intensity options for custom providers just like official ones.

Behavior

• Listens to the settings/document-updated event for the llm-pi-ai namespace (DSH 0.1.7-rc.x; the older settings/updated event no longer exists in current DSH builds).
• Reads the current configuration through settings.describe() and scans the namespace's user layer (user.providers, the profile patch layer where your custom providers live). Bundle-inherited providers are never touched.
• For every provider model or modelOverrides entry without an explicit reasoningEfforts, writes:

reasoningEfforts:
  off: null
  low: low      # ← added in v0.2.0
  high: high
  max: max

• Existing explicit reasoningEfforts (including 3-tier { off, high, max } or false) is never overwritten unless force: true.
• Also sweeps shortly after startup and retries while the llm-pi-ai settings namespace is still activating, plus on every app-boot/config-reload / llm/adapters-updated as fallback triggers.
• Writes go through settings.mutate('llm-pi-ai', ops). providers is a volatile field of llm-pi-ai, so the commit is hot — the entry does not restart and the Composer picks the new levels up immediately.
• The write lands in the active profile's cordis.patch.yml, so it survives restarts.

Configuration

Field	Default	Meaning
enabled	true	Set false to disable auto application.
force	false	Set true to overwrite explicit user declarations (entries already equal to the target four tiers are skipped, so writes always converge).

Install

# from a release tarball (see the Releases page below)
dsh plugin --profile desktop add <path>\dsh-thinking-efforts-0.3.0.tgz

# or straight from the repository
dsh plugin --profile desktop add git+https://github.com/DreamRift/dsh-thinking-efforts.git

Prebuilt tarballs are attached to each GitHub Release: https://github.com/DreamRift/dsh-thinking-efforts/releases

Restart DSH and hard-refresh the browser after installing. The desktop profile wires the plugin's cordis.patch.yml in as a bundle overlay automatically (via dsh.bundle.patch).

Test

npm test

Changelog

v0.3.0 (2026-08-20)

• Fix: plugin had no effect on current DSH. Adapted to DSH 0.1.7-rc.x (official desktop nightly):
  • Trigger event corrected from the removed settings/updated to settings/document-updated(ns, revision).
  • Current values are now read from settings.describe() user layer; the old ctx.settings.get(ns) call path no longer exists in this DSH version (the settings service only offers describe/update/replace/mutate/configure/prepareDocument).
  • Startup sweep now retries until the llm-pi-ai settings namespace is active, and falls back to app-boot/config-reload / llm/adapters-updated triggers.
  • force: true now skips entries already equal to the target tiers, so the verification sweep always converges (no write loop).
• Added a dispose() teardown that clears pending timers.

v0.2.0 (2026-08-20)

• BREAKING: Default reasoning tiers upgraded from 3 (off/high/max) to 4 (off/low/high/max), aligning with the official llm-deepseek adapter since rc.7.
• Models with existing reasoningEfforts are no longer overwritten (default force: false).
• Use force: true to forcibly update all models.

v0.1.0 (Initial Release)

• Auto-add off/high/max reasoning efforts to custom provider models.
• Respect explicit declarations (do not overwrite when present).

License

MIT