dsh-thinking-levels

dsh-thinking-levels

Per-round thinking-level (reasoning_effort) control for DeepSeek Harness (dsh): pick Auto (a mask) in the session model selector and the plugin schedules low / high / max from the recent tool-call history before submitting the API effort — or fix a wire level (off / on / minimal / low / medium / high / xhigh / max) manually. Cheap tool rounds stay cheap; heavy work never starves.

• English README
• 中文 README
• 日本語 README
• 한국어 README
• Installation guide
• 中文安装指南
• 日本語インストールガイド
• 한국어 설치 안내
• Changelog
• 日本語 changelog
• 한국어 changelog

Compatibility note: Version 0.6.0 includes Japanese (ja) and Korean (ko) dictionaries and selector entries, but the current official DSH releases expose only zh and en through LocaleRuntime. On stock DSH, selecting ja or ko fails with locale "<id>" is not registered. These languages will work after official DSH adds the locale IDs. Advanced users can use a DSH fork that updates packages/client/locale/src/locale-settings.ts (LOCALE_IDS) and packages/client/locale/src/client/index.ts (LOCALES labels), together with the corresponding core dictionaries and tests, then rebuild and run the forked DSH. Changing this plugin alone cannot extend DSH's global locale list.

In a multi-step tool chain, the model re-thinks before every tool call — and that thinking dominates the wall-clock time (a 50-step agent task can spend minutes reasoning between tools). dsh-thinking-levels plugs into the agent/request waterfall that dsh re-resolves for every step (registered with prepend so the session model-selection assembly cannot overwrite its decision) and injects a thinking level into the next model request.

Preview

Screenshots of the live UI (dsh web):

Native model selector gains Auto — pick it and the plugin schedules low/high/max per step instead of a fixed wire level.

Thinking-level settings card: the auto scheduler plus its boundaries, and llm-pi-ai model-capability mapping (gear → gateway wire values).

Per-model capability card — pairs with dsh-llm-openai-completions: this card detects & writes capabilities, that adapter takes over the wire (compat.thinkingFormat).

Levels

Wire-level facts (verified against the official DeepSeek docs and dsh's llm-deepseek adapter): low maps 1:1 on deepseek-v4-flash / v4-pro, while medium / xhigh collapse onto high. The adapter accepts off | low | high | max and rejects anything else with UNSUPPORTED_REASONING_EFFORT — auto is the plugin's mask layer, never sent to the API, always resolved to a concrete wire level before injection. on is not an effort level: it is advertised only by toggle-only models (Qwen3.6-style), and it only flips enable_thinking true — no reasoning_effort is sent; an effort-capable model never advertises on, so a manual on pick on one is stripped.

Custom wire mapping

For hand-declared llm-pi-ai models the settings card lets you map each level to the exact value your gateway expects (borrowed from dsh-thinking-effort): tick a level and enter its wire value, e.g. high → ultra. The mapping is stored as the model's reasoningEfforts table, so the Composer selection High sends ultra to the gateway. Leaving off empty means "do not send".

• Official preset: Off / High / Max (official DeepSeek style)
• Generic preset: Off / Low / Medium / High

Context-window presets

The settings card's per-model editor now includes a context window limit control: preset buttons 64K / 128K / 256K / 400K / 512K / 1M, a custom integer input, and a clear button. The value is written to the llm-pi-ai model entry contextWindow (integer 2000–1000000).

Upstream, the harness consumes it through resolveModelInfo(...).context.contextWindow for compaction thresholds, context-overflow detection and context-pressure projections. Because llm-pi-ai re-reads the live config on every resolve and the openai-completions takeover does not block model discovery, a settings edit takes effect on the next request without a restart.

The plugin config also accepts models['provider/model'].contextWindow as a validated (integer 2000–1000000) declaration at the composition/config surface.

Model-aware guard (v0.5.0)

The plugin never sends a reasoning_effort to a model that does not advertise one. Custom openai-completions routes (e.g. a local Qwen3.6 without reasoningEfforts) are classified non-reasoning via ctx.llm.resolveModelInfo, and any effort — inherited or scheduled — is stripped instead of sent, so dsh's per-request UNSUPPORTED_REASONING_EFFORT rejection cannot fire. Unsupported fields are never passed to an API that cannot take them.

Version behavior:

The auto scheduler may still pick low for supporting models — the capability guard above is what keeps it away from models that cannot take it.

Model-selector Auto

The session model selector (next to the model) now offers Auto after the wire levels (injected into the model-directory metadata by the plugin):

Auto scheduler

The hub is high (the official default). auto schedules between low / high / max; it never picks off.

The scheduling policy is the same source as dsh-tool-turbo (same simple-tool whitelist / payload thresholds / 75% ratio rule).

Install

See INSTALL.md for the full official-CLI guide (profile discovery, upgrade, migration, verification, troubleshooting). Quick start:

# 1. install the plugin into a profile from npm (web shown; any profile works)
#    (the web profile is a pnpm workspace root, so -w is required)
dsh plugin --profile web add dsh-thinking-levels -w
#    GitHub alternative:
#    dsh plugin --profile web add https://github.com/drscrewdriver/dsh-thinking-levels.git -w
#    local-path alternative (no network needed):
#    dsh plugin --profile web add /absolute/path/to/dsh-thinking-levels

# 2. restart dsh web (a running instance does not hot-load new bundle layers)
dsh web

Note: the dsh runtime uses pnpm 11, whose minimumReleaseAge supply-chain policy may block a

freshly published version with ERR_PNPM_MINIMUM_RELEASE_AGE_VIOLATION — add the version to

minimumReleaseAgeExclude in ~/.dsh/profiles/web/pnpm-workspace.yaml to lift the cooling period.

Manual link: registration (alternative to dsh plugin add):

#    ~/.dsh/profiles/web/package.json dependencies:
#      "dsh-thinking-levels": "link:<absolute path to dsh-thinking-levels>"
#    ~/.dsh/profiles/web/cordis.patch.yml:
#      - insert:
#          - id: thinking-levels
#            name: dsh-thinking-levels
cd ~/.dsh/profiles/web && pnpm install && dsh web

Configuration

Two surfaces share one schema:

• Assembly — the plugin row's config: in the profile composition (e.g. cordis.yml):

  config:
    level: auto            # off | on | minimal | low | medium | high | xhigh | max | auto — the default level when the session picks nothing
    allowDowngrade: true   # let the scheduler drop below `high`
    allowUpgrade: false    # forbid the scheduler lifting to `max`

• Runtime — the dsh-settings namespace thinking-levels (level, allowDowngrade, allowUpgrade, enabled, models): changes apply to the next model request, no restart needed. A visual editor is available under Settings → Plugins → configurable plugins.

Per-model capability overrides (models, keyed provider/model) confirm what auto-detection finds; the configurer has the final word:

config:
  level: auto
  models:
    llm-pi-ai/Qwen3.6-35B-A3B:   # non-reasoning thinking model (thinking toggle + budget)
      vision: false
      thinking: true
      efforts: false             # never send reasoning_effort (stripped at request time)
    llm-pi-ai/Qwen3.8-27B:       # effort-capable model (rc.6-era adapter without low)
      efforts: [low, high]       # confirm low → advertised in the selector + passed through

For Qwen thinking on/off + budget, configure the llm-pi-ai route instead:

compat.thinkingFormat: qwen (→ wire enable_thinking + thinking_budget via

thinkingBudgets), or qwen-chat-template (→ chat_template_kwargs.enable_thinking) for

effort models like Qwen3.8-27B.

Defaults: { enabled: true, level: 'auto', allowDowngrade: true, allowUpgrade: false, models: {} }.

Semantics: the model-selector pick outranks the plugin's default level. Pick auto (mask) → plugin schedules; pick a wire level → applied directly; pick nothing → the plugin's level default is used. allowDowngrade / allowUpgrade constrain auto scheduling only.

Auto-takeover of dsh-llm-openai-completions (v0.5.2)

Custom gateways (vLLM / LM Studio / self-hosted OpenAI-compatible proxies) must be served by dsh-llm-openai-completions once they declare thinking (reasoningEfforts table in llm-pi-ai) — otherwise pi-ai sends role: "developer" (400) or drops enable_thinking. This plugin maintains the takeover list automatically:

• Scans llm-pi-ai.providers for providers that are custom openai-completions gateways

(api: openai-completions or a non-official baseURL) and declare a reasoningEfforts table on any model;

• Merges them into llm-openai-completions.providers with enabled: true (existing manual

entries are preserved, deduplicated);

• Triggers on plugin start, llm/adapters-updated, and settings changes to llm-pi-ai or the

takeover list — no manual config editing;

• Soft-coupled: skips the write silently when llm-openai-completions is not installed (its

namespace is unregistered).

The takeover mechanism as a whole — control-plane contract (who is taken over, how a control-layer plugin decides and injects) and transport-plane wire contract — is standardized in the Takeover Control Spec; this plugin is the reference control-layer implementation of it.

Dependency note

The host half does not value-depend on @deepseek-ai/dsh-settings (settings registration goes through the cordis settings service provided by the dsh runtime) — no need to install official packages into the profile manually. dependencies is just @deepseek-ai/schemastery (installed automatically with the package).

Development

npm run lint        # eslint (typescript-eslint flat config)
npm run typecheck   # tsc --noEmit
npm test            # vitest — 46 tests

Test coverage: level policy (manual pass-through incl. the extended levels, on clamping, auto scheduler, validation, simple-tool boundary), the model-capability guard (reasoningEffortSupported, resolveEffortInjection stripping/passthrough), session-event parsing (guards, window cap, malformed records), the config schema (defaults lockstep, out-of-band rejection, models overrides), and takeover-sync (identification, dedupe merge, soft-coupling).

License

MIT