ggfgfgf-on/dsh-trajectory-anchor ↗★ 1
@dsh-ext/trajectory-anchor
Self-contained DeepSeek Harness bundle: first-round trajectory anchoring, EWMA trajectory scoring, adaptive drift rollback, anchorGate promotion gating, bootstrap context suppression, trajectory-log export, and Layer-4 training-data aggregation. Zero community-plugin dependencies. 适合需要优化Agent执行轨迹、防止漂移并进行数据聚合的开发者。
설치
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:ggfgfgf-on/dsh-trajectory-anchorKey configuration (cordis.patch.yml)
| Key | Default | Meaning |
|---|---|---|
bootstrapTools | [bash, str_replace_editor, pwsh] | intersected with the tools the agent scope can actually see; on Windows the bash row is usually disabled |
bootstrapMaxTokens | null (uncapped) | opt-in: capping the first response below its real size truncates the opening plan and kills the turn (community issue #85, reproduced locally); when enabled, the cap is explicitly stripped on promotion (seed-inheritance trap) |
gateEnabled | true | anchorGate promotion gating |
maxBootstrapSteps | 5 | gate fallback |
suppressContextOnBootstrap | true | agent-scope context suppression during bootstrap |
lexicon / ratioWeights | monitor-calibrated | weighted lexicon scoring (we/let's/we'll/"we need"/our vs "let me"; neutral i will/i'll/i need/check/verify) |
specMax / reactMin | 0.2 / 0.5 | persona_ratio band boundaries (spec/mixed/react) |
baselineMinSamples / rollbackPercentile | 10 / 25 | baseline-relative drift: rollback only fires when the current ratio's percentile within the session's own history is < 25 |
rollbackEnabled | true | calibrated rollback switch (safe to enable under baseline-relative semantics) |
leanDenyPatterns | 61 entries | drift-rollback deny set (* prefix wildcards) |
rewardAnnotator | default | Layer-4 process reward; pluggable extension point |