dsh-think-intensity

Thinking Intensity Strategy Plugin for DeepSeek Harness (DSH)
Decides reasoning effort levels on top of existing model capabilities — observable, self-healing, and fully reversible.

English | 简体中文

---

What Problem Does It Solve?

In DeepSeek Harness (DSH), thinking intensity is not an arbitrary runtime switch, but a strict capability declaration (reasoningEfforts):

1. Official Catalog Models (such as DeepSeek Reasoner, OpenAI o1/o3) ship with built-in effort declarations. Opening the model selector in DSH natively reveals the second-level menu: Reasoning Effort >;
2. Custom / Third-Party Models (such as StepFun step-5-preview, Zhipu glm-5.3, or proxy gateways) do not have reasoning effort fields exposed in the DSH custom model settings UI, leaving them without reasoningEfforts declarations;
3. The Result: DSH treats them as plain text models and hides the reasoning effort selector; if parameters are forcefully sent, upstream gateways reject the request with UNSUPPORTED_REASONING_EFFORT.

dsh-think-intensity solves this elegantly:

• Native Activation: Backfills effort declarations into hand-declared models to natively reveal DSH's built-in Reasoning Effort > menu;
• Explicit Choice Priority: When a user selects a level manually in the UI, the plugin respects the choice 100% and passes it through untouched;
• Default Strategies: Injects configured levels (defaultLevel, subagentLevel for subagents, or purposeLevels for compaction) when no manual choice was made;
• Self-Healing Fallback: Automatically blacklists rejected levels inside the session to prevent infinite retry loops;
• Sandboxed & Reversible: Zero invasive hacks; clean uninstallation leaves no leftover overrides.

---

Native UI Interaction

This plugin fully embraces DSH's official native UI components rather than adding intrusive floating widgets or custom sliders. In the chat composer, clicking the active model displays DSH's standard two-row menu:

┌──────────────────────────────────────┐
│ Model             step-5-preview / … >│
│ Reasoning Effort  High               >│
└──────────────────────────────────────┘

Clicking Reasoning Effort expands the native choice list:

• Off (Disable thinking / zero parameter)
• Low
• Medium
• High
• Max

---

Strategy Decision Flow

┌───────────────────────┐
│ Agent Request Triggered│
└───────────┬───────────┘
            │
            ▼
    Explicit UI Choice? ──(Yes)──► Pass through untouched (100% User Priority)
            │ (No)
            ▼
   Origin === 'subagent'? ──(Yes)──► Apply subagentLevel
            │ (No)
            ▼
  Purpose Match (compaction)? ──(Yes)──► Apply purposeLevels mapping
            │ (No)
            ▼
   Apply defaultLevel
            │
            ▼
     Capability Guard (resolveModelInfo)
     ├── Target model supports effort? ──► Inject and proceed
     ├── Model is non-reasoning? ──► Strip inherited parameter to prevent errors
     └── Level hit gateway error blacklist? ──► Fall back cleanly

---

Configuration Paths

DSH stores configuration patches in cordis.patch.yml under each profile:

OS	Desktop App (desktop profile)	Web App (web profile)
Windows	%USERPROFILE%\.dsh\profiles\desktop\cordis.patch.yml	%USERPROFILE%\.dsh\profiles\web\cordis.patch.yml
macOS	~/.dsh/profiles/desktop/cordis.patch.yml	~/.dsh/profiles/web/cordis.patch.yml
Linux	~/.dsh/profiles/desktop/cordis.patch.yml	~/.dsh/profiles/web/cordis.patch.yml

---

Configuration

In your profile's cordis.patch.yml, configure the plugin by entry id:

- id: think-intensity
  config:
    defaultLevel: high        # Default effort for unselected requests; empty = pass through
    subagentLevel: max        # Dedicated level for subagent sessions; empty = fall back to defaultLevel
    providers: []             # Whitelisted provider routes; empty = all providers
    purposeLevels:            # Internal task effort mappings
      compaction: low         # Save tokens on memory compaction
      session-title: off      # Disable thinking for session titles
    backfillEnabled: true     # Automatically backfill declarations to activate the native menu

Options Reference

Key	Type	Default	Description
defaultLevel	string	""	Global fallback level (low, medium, high, max); empty = no intervention
subagentLevel	string	""	Level for subagent sessions (origin === 'subagent'); empty = fall back to defaultLevel
providers	string[]	[]	Whitelisted provider routes; empty = all routes
purposeLevels	object	{}	Mapping from internal task purpose to effort level
statsFlushMs	number	30000	Minimum disk flush interval (ms) for $DSH_HOME/storages/think-intensity-stats.json
backfillEnabled	boolean	false	When true, writes missing reasoningEfforts declarations; default is dry-run
backfillLevels	object	{}	Effort to wire value table; empty falls back to DEFAULT_BACKFILL_LEVELS
routeDefaultLevel	string	""	Route default level; written only if all models in route support it

---

Manual Model Declaration (Without Auto-Backfill)

You can also manually declare reasoningEfforts under llm-pi-ai in cordis.patch.yml:

- id: llm-pi-ai
  name: "@deepseek-ai/dsh-llm-pi-ai"
  config:
    providers:
      stepfun:
        models:
          - id: step-5-preview
            name: step-5-preview
            contextWindow: 1024000
            reasoningEfforts:
              off: null       # off sends nothing to the wire
              low: low
              medium: medium
              high: high
              max: max

Restart DSH, and the native reasoning effort menu activates immediately.

---

Installation & Uninstallation

Option 1: Via npm (Recommended after release)

# Desktop profile
dsh plugin --profile desktop add dsh-think-intensity

# Web profile
dsh plugin --profile web add dsh-think-intensity

Option 2: Via GitHub Release Tag (Pinned version)

dsh plugin --profile desktop add github:fengzhiyushui/dsh-think-intensity#v0.1.0

Option 3: Local directory install

dsh plugin --profile desktop add /path/to/dsh-think-intensity

Uninstallation

dsh plugin --profile desktop remove dsh-think-intensity

# Verify cleanly removed
dsh --profile desktop --dump-config

---

Development & Testing

# Install dependencies
pnpm install

# Build single ESM bundle (lib/index.js) and types
pnpm run build

# Run 124 unit tests with Vitest
pnpm run test

# Type check
pnpm run typecheck

Core Verified Scenarios (124 Tests)

1. Explicit Priority: Manual UI choices are never overridden;
2. Default Injections: defaultLevel, subagentLevel, and purposeLevels applied correctly;
3. Capability Guard: Models without reasoning skip injection; non-reasoning models strip inherited parameters;
4. Wire Value Mapping: Custom wire values (e.g., high: ultra) correctly delivered to upstream gateways;
5. Self-Healing: UNSUPPORTED_REASONING_EFFORT triggers in-session blacklisting to prevent request crashes;
6. Legality Checks: Rejects invalid effort dictionaries and non-standard keys;
7. Concurrency & HMR Protection: Handles hot-reloading transactions and config revision conflicts;
8. Built-in Standard Levels: DEFAULT_BACKFILL_LEVELS complies with catalog schema.

---

Known Boundaries

• In-Memory Blacklist: Gateway error fallbacks are stored in memory and reset upon DSH restart to prevent transient network issues from persisting;
• Non-Reasoning Stripping: If an unreasoning model carries an inherited effort parameter, the plugin strips it to prevent gateway rejections (the only exception to "Explicit Priority").

---

License

MIT © 2026 fengzhiyushui and contributors.