dsh-context-provenance: An Observe-Only Evidence Ledger DSH plugin for DeepSeek Harness
030611/dsh-context-provenance
Observe-only provenance ledger for DeepSeek Harness, comparing adjacent request evidence without exposing sensitive data.
Provides observe-only context evidence labeling for DeepSeek Harness, helping developers understand the boundaries of Agent behavior differences without exposing sensitive content.
dsh plugin --profile web add github:030611/dsh-context-provenance- Category
- Data & Knowledge
- Platform
- DSH-Plugin
- Author
- 030611
- Distribution
- Plugin
dsh-context-provenance Key Features
dsh-context-provenance Repository
dsh-context-provenance Repo Summary
dsh-context-provenance is an observe-only evidence ledger plugin for DeepSeek Harness (DSH), designed to label and compare adjacent request runtime evidence without exposing prompt text, message content, tool schemas, or raw paths. It addresses the question of what public runtime interfaces can actually prove about an Agent's context, categorizing evidence as Observed, Estimated, or Unavailable to clarify the boundaries of behavioral differences. Maintained by 030611 under the MIT license, written in TypeScript, this DSH plugin is CPU-only, local-only, and performs no file, network, subprocess, GPU, persistence, session, permission, tool, model-routing, or request mutation.
What evidence matrix does the dsh-context-provenance DSH plugin provide?
The plugin collects evidence through public interfaces, such as observing the effective request provider/model from llm/stream GenerateOptions (gated by isAgentLoopRequest()), retrieving adapter context window from Session.requestContext(), and detecting system presence/change and tool catalog changes via GenerateOptions.system and GenerateOptions.tools without returning text or digests. For items with no public interface, like tool owner mapping, hidden policies, or private prompts, it explicitly marks them as Unavailable. It also leverages the official pluginInventory/list service to observe loader order and uses contextBreakdown and contextPressure projections to estimate system/tools/messages breakdown and prompt pressure.
How to install dsh-context-provenance?
Use the following command with DSH's plugin manager:
dsh plugin --profile web add github:030611/dsh-context-provenance
After installation, the plugin retains only the two most recent ordinary agent-loop request observations in memory and exposes them via the existing Cordis inspect query mechanism, without mutating any requests.
What is the audit verdict and boundary of dsh-context-provenance?
The plugin received a conditional pass against official DSH commit 47f943859bef60e4160492346772ded9b24f765a, acknowledging that its evidence ledger is deliberately incomplete and cannot explain all behavior differences. Public collision review found no overlap with official pluginInventory/list, token-meter projections, Cordis inspect, or the community dsh-context-doctor; its distinct surface is strict evidence labeling plus adjacent actual request comparison.
What are the dependencies and compatibility of dsh-context-provenance?
The plugin depends on several official DSH packages, including @deepseek-ai/dsh-agent, @deepseek-ai/dsh-cordis-host-runner, @deepseek-ai/dsh-host-plugin-inventory, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-session, and others, ensuring seamless integration with the DeepSeek Harness runtime. It is community-maintained and not an official DeepSeek project; related trust-layer plugins include Telemetry Redactor, Verification Receipt, and Evidence Audit.