Preface

When running DeepSeek V4 series with the Standard preset in DeepSeek Harness (dsh), two common startup issues are frequently encountered: V4 Pro tends to drift into a Let me trajectory with diminished capabilities when using the full tool catalog; V4 Flash often exhibits shallow thinking and hasty actions when operating with default unguided settings. Community solutions like two-stage anchoring and Flash persona routing exist, but they typically target single models or single channels, and can easily lose alignment again after long-session compaction or resumption.

Miraculous Standard is an agent-plane preset maintained by rinDBeans. It consolidates the four combinations of V4 Pro/Flash with the official API and opencode-go into a unified routing logic. It uses a Minimal anchor for the first request followed by model-aware catalog promotion, stabilizing the startup trajectory without sacrificing the complete Standard toolset, and manages long-session drift with epoch tracking. This article summarizes installation and configuration points based on the GitHub repository README and the SkillHub catalog page. This project is a community experiment and is not affiliated with DeepSeek.

What This Is

Miraculous Standard is an agent preset plugin for dsh (dsh-miraculous-standard, current version 2.0.0, MIT license). In one sentence: it stabilizes the startup environment of DeepSeek V4 Pro/Flash to the RL-aligned state without sacrificing the complete Standard toolset, and uses epoch management to prevent trajectory drift in long sessions.

It separates “first-request precise anchoring” from “complete Standard tool capabilities”: the first request only exposes the real tools from the official Minimal preset (bash and str_replace_editor), with the system prompt retaining the original Minimal text and stripping all auto-injections. After the first durable tool/call or assistant/message triggers promotion, the corresponding tool catalog is opened based on the model. Path selection relies solely on the routing model ID (/flash/i), not the provider, to avoid behavioral drift caused by differences between the official API and opencode-go channels.

Core Features

Unified Dual Model, Dual Channel

The same preset covers all four combinations of DeepSeek V4 Pro/Flash × official API/opencode-go. forcePath can be set to auto (default), pro, or flash to lock down the path.

Byte-Level Minimal Anchor for First Request

The first request only exposes bash and str_replace_editor, with the system prompt retaining the original Minimal text and stripping all auto-injections. The README describes this as a condition for more stable V4 Pro trajectory anchoring based on extensive community testing.

Model-Aware Promotion Catalog

  • Pro: After promotion, it maintains a minimal resident set (default 5-tool bundle: bash, str_replace_editor, dev_tool_search, skill_search, skill_load). Heavy tools are unlocked on-demand via dev_tool_search, preventing a full catalog dump that could pull the trajectory back to standard-like behavior.
  • Flash: Persona-led, catalog-immune; after promotion, it defaults to opening the full Standard catalog (flashPromotedCatalog: full), which can be changed to resident mode.

Unified Context-Gate

Before promotion, a claimed-baseline gate is used: only user messages and user skill gestures are retained. No third-party plugins, hooks, or runtime contexts can pollute the first request. suppressedContextSources filters out agent-instructions and skill-catalog by default, with unified gating covering unknown sources.

Non-Imperative Instruction-Hint

Post-promotion, the AGENTS.md hint is changed to neutral wording like “refer to documentation, consult as needed.” The README notes that imperative phrasing like “read first and follow them” disrupts the anchored trajectory back to let me; neutral wording preserves the trajectory while maintaining the availability of environmental documentation.

Epoch-Aware Long-Session Stability

  • Automatic rollback to the controlled stage after compaction;
  • dev_tool_search unlock sets are reset by epoch, preventing monotonic catalog inflation in long dialogues;
  • Instruction-hints/discipline hints are re-sent per epoch;
  • Duplicate injection is avoided by scanning durable logs for deduplication after process restart.

rc.7 Readiness and Windows Engineering

  • Synchronized with the official rc.7 subagent configuration (backgroundMode: one-shot);
  • Benefits from the official persistent Bash latency fix (reduced from seconds to hundreds of milliseconds per call on Linux/macOS) and max-token replay fix;
  • custom-bash auto-detects Git for Windows installation location and normalizes Git Bash paths like /e/foo to E:\foo; falls back to session cwd with a prompt when explicit workdir is unavailable.

Zero Extra Runtime Overhead

No warmup rounds, no replay, no third-party runtime dependencies; the plugin is a Node.js ESM .mjs file, makes no network requests, and collects no telemetry.

Installation and Enabling

Environment Requirements

  • DeepSeek Harness 0.1.0-rc.7 (rc.5/rc.6 are also functional, but rc.7 is recommended in the README);
  • Git for Windows required on Windows;
  • DeepSeek official API key or opencode-go subscription.

Installation

Method 1: Clone the repository and copy the preset directory.

git clone https://github.com/rinDBeans/dsh-miraculous-standard.git
# Windows PowerShell:
#   Copy-Item -Recurse .\dsh-miraculous-standard\preset "$env:USERPROFILE\.dsh\.agent-presets\miraculous-standard"
# Linux / macOS:
#   cp -r dsh-miraculous-standard/preset "${DSH_HOME:-$HOME/.dsh}/.agent-presets/miraculous-standard"

Method 2: Install via bundle (official command from the README).

dsh plugin --profile web add github:rinDBeans/dsh-miraculous-standard

After installation, fully restart dsh, and select Miraculous Standard (Pro/Flash unified, experimental) in the new session.

Model Configuration

Specify the provider, model, and default preset in the dsh configuration:

# DeepSeek official API or opencode-go subscription
agent-default-model:
  provider: deepseek            # or opencode-go
  model: deepseek-v4-pro        # or deepseek-v4-flash
  reasoningEffort: max          # Recommended: max; at least high
agent-presets:
  default: miraculous-standard

Configuration Details

Configuration is located in preset/agent.cordis.yml. Common keys are listed below.

Key Default Value Description
bootstrapTools [bash, str_replace_editor] Tools visible in the first request
promoteOn either Promotion trigger: either / tool-call / assistant-message
bootstrapMaxTokens Not set Max output tokens for the first request (opt-in, stripped after promotion)
residentTools 5-tool bundle (see above) Pro’s resident catalog after promotion; can add read, grep, etc.
compactionTools [read, edit, glob, grep] Controlled toolset after compaction and before re-promotion
forcePath auto Path lock: auto / pro / flash
flashGuidance true Flash divine-mode persona toggle
flashPromotedCatalog full Flash’s catalog after promotion: full / resident
includeSubagents true Whether subagents also go through the bootstrap anchoring phase
proDisciplineHint false Inject a long-task discipline hint per epoch after Pro promotion

Request Lifecycle

User's First Message
    │
    ▼
┌ Request #1 ─ Bootstrap Phase ─────────────────────────────────────┐
│ tools   : bash + str_replace_editor (Minimal real schema)         │
│ context : context-gate blocks unknown injections by default       │
│ persona : Pro = Minimal original text / Flash = divine-mode five anchors │
│ budget  : adapter default (bootstrapMaxTokens optional)           │
└───────────────────────────────────────────────────────────────────┘
    │ First durable tool/call or assistant/message
    ▼ Promotion (derived from durable events, safe for resume/reload)
┌ Request #2+ ─ Resident Phase ─────────────────────────────────────┐
│ Pro   : Resident 5-tool bundle + unlocked tools (configurable via residentTools) │
│ Flash : Full Standard catalog (default; configurable to resident) │
│ context : Injection resumes; neutral instruction-hints prompt AGENTS.md as needed │
└───────────────────────────────────────────────────────────────────┘
    │ Compaction/End
    ▼ Rollback to controlled stage: bootstrap pair + compactionTools, until new promotion signal

Typical Usage and Verification

For daily use, configure the provider and model as described above and start a new session; Pro and Flash are automatically routed based on model ID, with no need to manually switch presets.

The repository includes smoke tests, which can be executed locally after cloning:

node test/smoke.mjs

The README states that tests cover dual-path anchoring, promotion catalogs, unlocking, compaction rollback, epoch unlock isolation, per-epoch hints, context-gate injection blocking, resume deduplication, and degraded path; output ALL PASS when all pass.

Suitable Scenarios and Notes

Who It’s For

  • Users running V4 Pro on dsh who want to avoid the Let me trajectory drift caused by Standard’s full catalog;
  • Users running V4 Flash on dsh who desire a deeper planning persona while retaining full Standard capabilities;
  • Scenarios requiring epoch-aware catalog and hint management in long sessions, compaction, subagents, and resume;
  • Windows environments needing Git Bash path normalization and custom-bash fallback.

Known Limitations

  • The README claims strong reproducibility for the trajectory anchoring mechanism, but capability gains in independent reproductions remain inconclusive; results should be judged by metrics like task completion, tokens/time, etc.
  • On Flash’s closely related multi-turn modification chains, static guidance may be counterproductive; set flashGuidance: false if needed.
  • Windows PTY: The rc7 node-pty upgrade does not address the lack of a Windows process inspector; it still uses custom-bash (new process per call); adding read/grep to the resident set can reduce bash calls.
  • Compaction summary quality, token accounting, etc., are host capabilities of dsh and not within the scope of this preset.

Security Notice

The plugin runs with the privileges of the current dsh process. Before installation, it is recommended to review the GitHub source code, the MIT license, and the derivative source notes in the NOTICE file. SkillHub is a community catalog site and has no official affiliation with DeepSeek or High-Flyer.

Conclusion

Miraculous Standard unifies the startup trajectories of V4 Pro/Flash on the official API and opencode-go into a consistent two-stage anchoring and epoch management system, addressing first-request drift and long-session drift while preserving the complete Standard toolset. If you are running the DeepSeek V4 series on dsh and encountering the above startup or long-session issues, you can install it per the README and adjust the promotion catalog and Flash persona in agent.cordis.yml as needed.