Skip to main content
Back to All Projects
Open Source Verified Architecture

dsh-anchored-standard

dsh-anchored-standard is a two‑phase preset for the DeepSeek Harness that first runs a minimal‑aligned bootstrap and then switches on the full Standard toolset (Project2 98/99). It lets you get a lean initial environment and automatically expand to the complete suite without manual reconfiguration.

3,826 Stars 118 Forks

Core Technologies & Frameworks

JavaScript

Technical Architecture & Specifications

### System Architecture & Under the Hood The `dsh-anchored-standard` project is an experimental agent preset framework built for the DeepSeek Harness. Its core objective is trajectory stabilization: constraining an LLM session's initial execution context using a lightweight "Minimal" tool schema before promoting the session to a broader catalog of Standard tools once context durability is established. In standard agent harnesses, presenting a massive tool catalog upfront can introduce unnecessary noise into early context windows. `dsh-anchored-standard` mitigates this by gating tool schemas through specific state transitions during the session lifecycle. ``` +-------------------------------------------------------------------+ | BOOTSTRAP PHASE | | Starts with Minimal schema (2 tools) or 0 tools (No context inject)| +-------------------------------------------------------------------+ | v [ Promotion Signal Trigger ] +-------------------------------------------------------------------+ | PROMOTED PHASE | | Unlocks Standard Tool Catalog / Resident Catalog On-Demand | +-------------------------------------------------------------------+ ``` ### Operational Modes Breakdown The repository structures its behaviors across distinct directory presets, each handling context anchoring and promotion signals differently: * Anchored Standard (`preset/`): Launches the context with a minimal pair of 2 tools. It listens for a durable `tool/call` or `assistant/message` (`promoteOn: either`) before unlocking the full tool set. This incurs zero additional API overhead. * Zero-Anchored Standard (`zero-anchored-standard/`): Begins with 0 tools exposed. It forces a single fixed anchor turn and triggers promotion upon receiving the anchor's `assistant/message` reply. Cost: +1 model call. * Whoami Standard (`whoami-standard/`): Operates like Zero-Anchored, but uses a explicit `"你是谁"` (Who are you) self-introduction prompt turn as the anchoring trajectory. Cost: +1 model call. * Prefab Anchored Standard (`prefab/`): Bypasses active promotion turns by loading a pre-seeded, successfully rolled trajectory history. The session starts already promoted, requiring 0 extra API calls to instantiate. * Eternal Minimal (`eternal-minimal/`): Enforces a static schema where the visible catalog never expands beyond 2 tools. Heavy tool execution is routed externally through the `dshx` bash gateway. * Wire Think-Execute Standard (`wire-think-standard/`): Keeps tools present in the context payload but enforces `tool_choice: none` on the wire via a sibling provider route. ### Status and Engineer Review Active development on the project is currently frozen (as of August 2026) due to official API price shifts that made large-scale evaluation loops (such as Project2-class runs and multi-trial probe rollouts) cost-prohibitive. The codebase is strictly in maintenance mode for critical bug fixes and harness compatibility updates. Despite the freeze on preset development, the mechanics are model-agnostic. The underlying context-gate framework, the probe evaluation suite, and the prefab trajectory pipeline provide valuable reference patterns for developers implementing custom context-gating or phased tool delivery systems inside their own agent architectures. To explore the implementation details directly, inspect the configuration schemas located within the `preset/`, `prefab/`, and `eternal-minimal/` directories.
Reviewed by DevTechPulse Editorial Board

All listed blueprints, repositories, and case studies are verified against public documentation and LTS container environments. For inquiries or updates, view our Editorial Policy.