dsh-context: Inspecting Context, Tools and Agent Collaboration
Purpose, mechanisms, getting started and boundaries
Source: bowenliang123/dsh-context · bowenliang123
User Problem
Positioning and task fitWhen a long session slows down, becomes expensive or drifts, a single context percentage reveals little. dsh-context is an observability plugin for DeepSeek Harness. It connects the current window, request history, tool execution and agent relationships. Its value is diagnostic: establish what entered the context and when compaction occurred before changing task boundaries, tool output or skill configuration.
- Diagnose sudden growth in tool results
- Compare usage and context composition across sessions
- Inspect parent and child tasks and communication
Product Mechanism
Capabilities and execution- From overview to a requestContext Insights supplies the cross-session view, while the Context tab exposes session details. Filter for an unusual session, then inspect request trends to turn a vague cost concern into a question about a particular turn. Source
- Inspect assembled contextContext Browser expands prompts, tool schemas, user input, reasoning and tool results, with a separate category for skill injections. Large tool definitions and large tool responses become distinguishable problems requiring different remedies. Source
- Align changes with eventsRequest-level trends align window growth with compaction and pruning. File activity and command records lead back to tool results, creating a path from a change in occupancy to the operation that caused it. Source
- Observe the agent familyFleet connects the relationship graph, member list, tasks and communication. In collaborative work, it helps investigate a member that keeps running, a dependency-blocked task or an unhandled message. Source
Workflow
Start with a small task- Install in the active profileSearch for dsh-context in the Harness plugin UI, or use the official CLI command below. Install into the profile used by the web interface, then open a real session.
- Establish a baselineObserve Current Context and request trends on a small task before reading large files. Do not subtract current-window tokens from cumulative billed tokens: they describe different quantities.
- Locate the incrementAfter a file read or tool call, inspect added content, tool output and compaction events. Expand the underlying text to investigate repeated loading or oversized responses.
- Compare after a changeRepeat a comparable task after narrowing reads or reducing irrelevant tools. Compare request structure and final results while accounting for model and task differences. Do not attribute performance gains to the plugin without a controlled test.
The trial sequence and checks are editorial guidance. Entry commands follow official documentation; this project was not executed for the article. Source
Official entry example
dsh plugin --profile web add dsh-context
dsh webComplete the prerequisites above first. Examples contain no real credentials.
Limits and Boundaries
Constraints and selection| Boundary | Meaning | Suggested action |
|---|---|---|
| Billing and estimates | Cost uses mapped prices, and context reconstruction can contain estimates. Source | Read the pricing mapping and use provider billing for settlement. |
| Historical reconstruction | Pre-compaction steps can be reconstructed from archives rather than preserved as exact requests. Source | Check approximation notices and preserve execution records for important investigations. |
Scroll horizontally to view the full table
It fits existing Harness users who need to understand context costs. A dashboard can reveal an unclear task or a model limitation, but cannot resolve either by itself.
How to Evaluate Adoption
Trial scenario and verificationA useful exercise is to list a directory, read one long file and summarize it. Compare tool schemas, file content and assistant replies, then locate each in Context Browser. Repeated content suggests a reading-strategy problem; several agents repeating the same research suggests a task-division problem visible in Fleet. This tests observability, not automatic compression performance.
Installation is verified when the entry point appears, the selected session has data and underlying content expands. For empty charts, check the session and profile before investigating compatibility.
Evidence: Pinned README · Supporting official documentation · Current repository。
Source fetched (UTC): 2026-11-02T07:26:37.778Z；snapshot commit: 6d5682870e36。The source date denotes the repository snapshot commit, not the project’s original launch.