dsh-trajectory-governor
A DeepSeek Harness policy plugin that governs agent task trajectories with revision-aware verification, finish gating, and optional benchmark controls.
Install
$ dsh plugin --profile web add .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.
Key Capabilities
- Tracks task episodes, relations, phases, artifacts, and information novelty across agent turns.
- Creates revision-aware verification debt after recognized mutations and requires applicable readback or executable verification evidence.
- Temporarily restricts selected write/edit tools before evidence is gathered, while retaining mixed-tool safeguards described by the project.
- Guards configured explicit finish tools and natural turn completion when evidence blockers remain.
- Optionally requires parseable, current-revision benchmark results and compares QPS only under matching benchmark specifications.
- Provides the read-only trajectory_policy_status tool for per-agent policy state.
- Writes a local append-only JSONL decision ledger, with hashing of original messages rather than storing their text.
- Supports native tool events and Code Mode dispatch events.
Useful For
- Applying evidence-based completion controls to coding or editing tasks.
- Requiring validation after workspace mutations before an agent completes a task.
- Running benchmark-gated performance experiments with configured quality and comparability thresholds.
- Inspecting agent trajectory state, open verification debt, and completion blockers.
- Starting in shadow mode to observe policy decisions before enabling active request intervention.
Who It Fits
- DeepSeek Harness users configuring agent execution policies.
- Developers running code-editing agents that need revision-aware verification controls.
- Teams evaluating performance-oriented agent workflows with explicit benchmark gates.
Documented Limitations
- The relation engine is an explainable rule baseline, not a learned classifier.
- Artifact tracking currently relies mainly on paths and tool parameters.
- Shell-command semantics are recognized conservatively.
- Benchmark passing requires explicitly named metrics in tool output; free-text performance claims are not accepted.
- It cannot perform automatic rollback because no official workspace checkpoint API is available.
- It does not provide an automatic subagent evaluator, contextual bandit, or workspace counterfactual fork runner.
- It does not replace tests, sandboxing, approval, or human review.
DSH Compatibility
Version-specific runtime evidence collected by DSH Plugin. A missing result means we have not tested that combination yet.
Security Signals
Objective signals discovered from package metadata and source inspection. These are not a guarantee that a plugin is safe.
package.json declares 1 runtime, 13 development, 6 peer, and 0 optional dependencies.
package.json declares DSH bundle metadata.
GitHub reports the repository license as MIT.
A root package.json was captured and can be inspected by the registry.
Public GitHub source metadata is available for this registry snapshot.
Source & Registry Notes
Public provenance, Registry classification, and the latest source check for this entry, kept separate from runtime verification.
- Source repository
- chunsi-w/dsh-trajectory-governor
- Registry source
- GitHub · dsh-plugin topic
- Registry classification
- Plugin
- Source checked
- ce53de0 · 2026-08-31
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
- 13GitHub stars
- Forks
- 0
- Open issues
- 0
- Last commit
- 2026-08-18
- Last release
- No release detected
Related DSH Plugins
Ranked by overlapping capabilities, use cases, plugin type, categories, and DSH profile.
DSH undo and crash-recovery plugin with snapshots, rollback, Safe Mode, and offline Windows tools.
View plugin →A DeepSeek Harness notification plugin with a model-facing notify tool, automatic session-event alerts, and multi-channel delivery.
View plugin →Schedule and run standalone coding tasks in fresh Agent sessions with Web and Agent management.
View plugin →Governance-powered AI-agent task execution framework with a DeepSeek Harness plugin integration.
View plugin →