dsh-newwindows

dsh-newwindows

中文说明 | English

dsh-newwindows is a performance-oriented extension plugin for DeepSeek Harness (DSH/Cordis). Inspired by the modern OpenAI Codex runtime context architecture, it delivers Summary-Free Context Rollover, Model-Authored Notes, and Scoped Raw History Lookup.

Eliminate sluggish, lossy LLM recursive summarization and empower your agents with microsecond-level, deterministic context relay during long-horizon tasks.

---

✨ Key Features

• ⚡ Zero-Token Summary-Free Rollover: Bypasses costly recursive LLM summarization. Leverages DSH surface replacement for sub-millisecond context reset, completely eliminating 5~20 second stalls and token billing overhead.
• 🛡️ Node 0 Absolute Invariant: Permanent retention of initial system prompts (startSeq = 1). Never loses core identity, system rules, or global tool specifications across window transitions.
• 📝 Model-Authored Notes: Agents actively curate architectural decisions and pending tasks during reasoning. Active notes are automatically formatted and injected as deterministic seeds into new windows.
• 🔒 Concurrency & Tool Safety Barrier: The new_context tool declares exclusive execution. Latched state machines postpone compaction until the step boundary, guaranteeing all concurrent tool calls and results pair completely before rollover.
• ⏱️ Single-Shot Budget Reminder: Triggers a gentle reminder when approaching budget boundaries (e.g. 75%). State-latched de-duplication prevents context spamming.
• 🔍 Scoped Raw History Lookup: Immutable physical logs remain intact. Provides a bounded, read-only inspection tool with strict pagination and per-item character truncation.

---

💡 Comparison Matrix

---

🚀 Installation & Setup

Option 1: Via Cordis Bundle Patch (cordis.patch.yml)

Declare the plugin in your DSH bundle configuration:

- insert:
    - id: dsh-newwindows
      name: '@deepseek-ai/cordis-plugin-group'
      group: true
      isolate:
        newWindows: true
      config:
        - id: dsh-newwindows-runtime
          name: 'dsh-newwindows'

Option 2: Programmatic Loading

import { Context } from '@deepseek-ai/cordis'
import plugin from 'dsh-newwindows'

const ctx = new Context()

// Load plugin with custom configuration
await ctx.plugin(plugin, {
  auto: true,
  reminderThresholdRatio: 0.75,
  overflowThresholdRatio: 0.90,
  maxNotes: 30,
})

// Access registered services from context
const { notesStore, windowChain, engine } = ctx.newWindows

---

⚙️ Configuration Options

The plugin accepts the following parameters during initialization:

---

🤖 Contributed Agent Tools

Once activated, dsh-newwindows equips the agent with three specialized tools:

1. note_action

• Purpose: Structured state scratchpad.
• Actions: create, update, delete, list.
• Use Case: Allows the agent to persistently retain technical milestones, constraints, and pending todos across window rollovers.

2. new_context

• Purpose: Explicit window rollover trigger.
• Mechanism: Declares exclusive execution mode. The agent initiates rollover upon phase completion or high context usage, specifying next_goal to cleanly start a fresh window.

3. lookup_raw_history

• Purpose: Safe diagnostic probe over archived interactions.
• Mechanism: Provides bounded, read-only pagination over shadowed raw events with optional role filtering (user, assistant, tool).

---

📖 Architecture & Developer Documentation

• Core architecture & invariant design: System Design Spec
• DSH internal interface reconciliation: DSH Integration Guide
• Test suite & verification matrix: Validation Matrix

---

📄 License

Licensed under the Apache-2.0 License. Core mechanisms inspired by the OpenAI Codex context management architecture.