drscrewdriver/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

v0.7.0-beta.1 (2026-09-06): the short-circuit route is retired. This release no longer depends on dsh-llm-openai-completions — custom-gateway fixes ride the official llm-pi-ai compat surface (requires dsh ≥ v0.1.2-alpha.1); keep the adapter plugin uninstalled. See the CHANGELOG.

• 中文 README
• 日本語 README
• 한국어 README
• README en français
• Deutsche README
• README in italiano
• README на русском
• README en español
• Installation guide
• 中文安装指南
• 日本語インストールガイド
• 한국어 설치 안내
• Guide d'installation en français
• Installationsanleitung auf Deutsch
• Guida all'installazione in italiano
• Руководство по установке на русском
• Guía de instalación en español
• Changelog
• 日本語 changelog
• 한국어 changelog
• Changelog en français
• Changelog auf Deutsch
• Changelog in italiano
• Список изменений на русском
• Changelog en español

▼ DSH version support

This release (4.0.0) supports DSH v0.2.0-rc.1 to < 0.2.1 only.

DSH version	Status	Notes
≥ 0.2.0-rc.1	✅ Supported	This release (4.0.x): same declarative settings surface (.volatile() schema fields rendered by the host; cross-plugin reads/writes through the configForms service), with the peer gate retargeted to the 0.2.0-rc segment
≥ 0.1.7-rc.1 to < 0.2.0	✅ Supported	Use the 3.x line (3.4.3, npm dist-tag dsh-0.1.7): declarative settings — the host renders the Plugins form from the plugin's .volatile() schema fields; cross-plugin reads/writes go through the configForms service
< 0.1.7-rc.1	⚠️ Not supported	DSH 0.1.7 removed the imperative settings registration and the per-plugin card seat the earlier lines (3.0.x and below) relied on — stay on plugin 3.0.2 for 0.1.2–0.1.6 hosts.

The boundary is 0.1.7-rc.1, where DSH removed the imperative settings registration (settings.register / installSettingsSection) and the client settingsScope service. Since that boundary the runtime-adjustable config fields are marked .volatile() in the schemastery schema, the host generates the settings form from that schema alone (no registration call, no client settings card), and the plugin reads the live values per request, driven by loader/volatile-update. The 3.1.x–3.4.x lines target the 0.1.7 declarative surface; the 4.0.x line is the same surface retargeted to the 0.2.0-rc segment.

Version-range policy: every compatibility line pins its host segment tightly. Lines targeting 0.1.x hosts follow >=0.1.x-rc.1 <0.1.(x+1)-0 (3.1.x: >=0.1.7-rc.1 <0.1.8-0; 3.0.x: >=0.1.5-alpha.1 <0.1.6-0; 2.0.x: >=0.1.2-alpha.1 <0.1.3-0; 1.0.0-beta: >=0.1.0-rc.8 <0.1.2-alpha.1); lines targeting the 0.2.x segment follow >=0.2.0-rc.1 <0.2.1-0 (4.0.x: >=0.2.0-rc.1 <0.2.1-0). No line ever declares an open upper bound, so a compatibility resolver can never match a plugin line against a newer host segment it was not built for. 0.4.0–0.6.0 carried no dsh peer declarations at all and are effectively compatibility-untyped — do not install them.

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):

Levels

Level	Meaning	Where
off	thinking disabled (manual only — never auto-picked)	model selector / default level
on	thinking enabled (toggle-only models only): sends enable_thinking, never a think effort	model selector / default level
minimal	least effort (very light tasks)	model selector / default level
low	manual pick for simple chat tasks (cheap rounds stay cheap)	model selector / default level
medium	medium effort	model selector / default level
high	the official default effort	model selector / default level
xhigh	extra high effort	model selector / default level
max	heavy work	model selector / default level
auto	mask: schedule per step from the recent tool-call history, resolved to a wire level before submission	model selector (injected by the plugin) / default level

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, 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 in the llm-pi-ai config, 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

The visual editor for this mapping rode the plugin's settings card, which the DSH 0.1.7 migration removed (the seat no longer exists). Edit the reasoningEfforts table through the official Models settings surface instead — the host-side detection and injection read it live either way.

Effort slider (per model line)

The model panel's per-line effort <select> became a segment slider: click a line's effort chip (it shows the effective level, or Provider default) and a full-width slider expands under that line. The stop count adapts to what the model advertises — two-stop toggle-only models, the official off/low/high/max set, custom gateway wire values; auto (when the directory carries the mask) is pinned leftmost, then off/on, then the strength gradient, unknown wire values last.

Interaction follows the approved dsh-reasoning-effort visual baseline: the thumb follows the pointer continuously and snaps on release (one route write per gesture), the plain thumb is pure white in every theme, the thumb stays fully visible at both endpoints, and ←/→/Home/End step stop-by-stop. The select's Provider default reset survives as the ↺ button of the slider row.

DeepSeek lines (the official route, or any gateway model whose id/name says deepseek) run the whale-girl runner as the thumb: an 8-frame side-run strip, ping-pong looped (720 ms per direction at rest, 420 ms while dragging), frozen under prefers-reduced-motion. Every other model gets the plain knob.

The strip is community whale-girl artwork sourced from HanaAyane/dsh-reasoning-effort (assets/chibi-runner-strip.png); regenerate the inlined asset with python tools/whale-mascot.py.

Context-window presets

The composer tool-row quick control (next to the model/effort select) edits a context window limit: preset stops 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) — or to the llm-deepseek entry for official DeepSeek models.

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 compat sync 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:

dsh version	low handling
rc.6 (old)	not native: the selector only shows it when a configurer-confirmed models override names it; the level is then advertised (selector + request validation) and passed through verbatim
rc.7+ (new)	native: the plugin neither rewrites nor re-injects it; a manual low pick passes through unchanged

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):

Model-selector pick	Behavior
Auto	plugin schedules via tool history + the upgrade/downgrade toggles, resolves to low / high / max before submission
off / on / minimal / low / medium / high / xhigh / max	manual choice wins — plugin does not intervene (on stays on on toggle-only models, never lifted to an effort; effort-capable models strip it)
unset	the plugin's default level applies (below)

Auto scheduler

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

Recent tool calls	Level
none (fresh prompt, pure chat)	low
≥75% simple tools, small args, downgrades allowed	low
mixed / heavy tools	high
very heavy payloads, upgrades allowed	max

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 plugin's .volatile() config fields (enabled, level, allowDowngrade, allowUpgrade): DSH 0.1.7 renders the Plugins settings form from the declared schema, and committed changes reach the plugin as live config references (loader/volatile-update) — they apply to the next model request, no restart needed. (models stays a configurer-level field: edit it in the profile composition.)

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.

Official compat surface: the short-circuit tool is retired (0.7.0-beta.1)

Once custom gateways (vLLM / LM Studio / self-hosted OpenAI-compatible proxies) declare thinking, this plugin writes the fixes into the official llm-pi-ai compat surface (introduced in dsh ≥ v0.1.0-rc.8, commit 884f7b9c41) — dsh-llm-openai-completions is no longer needed and should stay uninstalled:

• Scans llm-pi-ai.providers for routes that are custom openai-completions gateways (api: openai-completions or a non-official baseURL) and declare a reasoningEfforts table on any model (including modelOverrides), then writes:
  • route-level compat.supportsDeveloperRole: false — the system prompt goes out as system, fixing the vLLM / SGLang Unexpected message role 400;
  • model-level compat.thinkingFormat: 'qwen-chat-template' on toggle-style thinking rows (thinking table without row-level supportsReasoningEffort) — pi-ai then sends chat_template_kwargs.enable_thinking (bare vLLM servers ignore the top-level enable_thinking of the plain qwen format);
• Writes go through the official settings channel (read → pure transform → whole-section settings.update('llm-pi-ai', …)), so dsh's schema validates the write where it is written: a dsh older than rc.8 rejects the fields with a log warning — no silent misconfiguration; explicit values on any layer are never clobbered;
• Triggers on plugin start, llm/adapters-updated, and llm-pi-ai config changes — no manual config editing;
• Response-side inline <think> splitting remains a gateway concern: bare vLLM needs --reasoning-parser qwen3 (pi-ai parses only reasoning_content / reasoning / reasoning_text).

Dependency note

The host half does not value-depend on @deepseek-ai/dsh-settings — since the DSH 0.1.7 line there is no settings registration at all: the settings form is generated by the host from the plugin's declared schemastery schema (.volatile() fields), and the client half talks to the configForms 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 the official-compat sync (identification, explicit-value respect, identity idempotence, write-time schema gating).

License

MIT