← DeepSeek Harness Plugins
DeepSeek Harness PluginManifest Valid

dsh-trajectory-anchor

Trajectory anchoring and drift-rollback bundle for DeepSeek Harness agents with EWMA scoring and automated lexicon calibration.

Terminal & TUIDeveloper ToolsSecurity & PolicyRemote Execution
1GitHub Stars0ForksUpdated2026-09-26

Install

No clear install command was found in the current repository evidence.

Plugin Overview & Capabilities

AI-assisted organization based on the public repository snapshot. The content must be grounded in source evidence and does not replace compatibility or security verification.

source-grounded
dsh-trajectory-anchor is a self-contained DeepSeek Harness plugin bundle designed to prevent agents from drifting off-task. It locks in disciplined initial behavior during the bootstrap phase, suppresses early runtime context, continuously scores trajectory via EWMA and dynamic lexicons, triggers adaptive rollbacks on behavioral degradation, and exports annotated trajectory logs for Layer-4 PRM/DPO training data.

Key Capabilities

  • First-round trajectory anchoring and bootstrap context suppression
  • AnchorGate promotion gating based on reasoning style recognition
  • Continuous EWMA trajectory scoring with baseline-relative drift detection
  • Automatic output-fingerprinted lexicon discovery and calibration across diverse model styles
  • Adaptive drift rollback using configurable deny patterns
  • Process-reward trajectory annotation and Layer-4 training dataset export

Useful For

  • Preventing multi-turn subagents from degrading into reactive or hallucinated loops
  • Enforcing strict initial reasoning plans before unlocking unconstrained tool execution
  • Collecting structured process-reward (PRM) datasets and pairwise trajectory logs for post-training

Who It Fits

  • DeepSeek Harness agent developers
  • AI researchers evaluating multi-agent trajectory stability
  • LLM fine-tuning engineers collecting PRM and RLVR process datasets

Documented Limitations

  • Requires Node.js 22 or newer for built-in zstd decompression during training data export
  • Lexicon auto-calibration n-gram engine currently covers CJK, Latin, Kana, Hangul, and Cyrillic scripts; non-supported scripts like pure Arabic require manual profiling
  • Downstream outcome attribution for step-level quality proxying can be noisy in multi-tool execution turns

DSH Compatibility

Version-specific runtime evidence collected by DSH Plugin. A missing result means we have not tested that combination yet.

Not tested yetNo runtime compatibility tests have been published yet.

Security Signals

Objective signals discovered from package metadata and source inspection. These are not a guarantee that a plugin is safe.

Dependency Counts

package.json declares 0 runtime, 0 development, 1 peer, and 0 optional dependencies.

info
Dsh Bundle Declared

package.json declares DSH bundle metadata.

info
License Declared

GitHub reports the repository license as MIT.

info
Package Manifest Available

A root package.json was captured and can be inspected by the registry.

info
Source Available

Public GitHub source metadata is available for this registry snapshot.

info

Source & Registry Notes

Public provenance, Registry classification, and the latest source check for this entry, kept separate from runtime verification.

Source repository
ggfgfgf-on/dsh-trajectory-anchor
Registry source
GitHub · dsh-plugin topic
Registry classification
Plugin Bundle
Source checked
7d52cb0 · 2026-09-26

This project is independently indexed from public source information. DSH Plugin is not affiliated with DeepSeek or the plugin author. Always check the author repository before installation.

Repository Activity

GitHub stars
1GitHub stars
Forks
0
Open issues
0
Last commit
2026-09-26
Last release
No release detected
For maintainers

Maintaining this plugin?

This listing is generated from public repository data. If you maintain this project, you can review the information and share this listing with your users if you find it useful.

Add to README
Listed on DSHPlugin.app