ArtificialNotImbecile/dsh-context-taxonomy
Context Taxonomy is a learning and debugging companion that shows how DSH builds the context for every ordinary agent call: it turns the provider-neutral logical request at the public llm/stream dispatch layer into an inspectable per-call taxonomy — complete system prompt, conversation history, current prompt, tool definitions, model options, token composition, cache usage and reasoning evidence — kept beside the conversation in a dedicated Context Taxonomy tab. It adds estimated composition by category, MessageSource.kind / ContextForm provenance, unclassified-field surfacing, a logical reasoning-retention check, and sanitized canonical JSON. It complements (not replaces) the official Trajectory view, and it only observes ordinary agent-loop calls after installation — auxiliary calls are excluded and pre-install calls cannot be reconstructed.
Install
dsh plugin --profile web add @artificialnotimbecile/dsh-context-taxonomy@0.1.0npm @artificialnotimbecile/dsh-context-taxonomy 0.1.0 verified 2026-09-02 (repository field → github.com/ArtificialNotImbecile/dsh-context-taxonomy; README EN primary). Install per README: dsh plugin --profile web add @artificialnotimbecile/dsh-context-taxonomy@0.1.0 (an npx @deepseek-ai/dsh@0.1.0-rc.6 variant is also shown). The demo screenshot in the README is a real DeepSeek-V4-Flash run on the official 0.1.0-rc.6 web profile.
Compatibility
DSH web profile (recorded against official 0.1.0-rc.6); observes the public llm/stream dispatch layer; keeps a separately retained, sanitized sidecar.
Details
- Repo: ArtificialNotImbecile/dsh-context-taxonomy
- Category: Agent Capabilities
- Stars: 1
- Version: npm @artificialnotimbecile/dsh-context-taxonomy 0.1.0
- Last push: 2026-08-14
- First seen: 2026-08-13
Recent updates
Per-call context taxonomy from llm/stream; System/Conversation/Current-prompt/Tools/Options/Unclassified sections; provenance + token/cache breakdown; Context Taxonomy tab; complements Trajectory.
FAQ
- How is it different from Trajectory?
- Trajectory answers 'what happened during the run, in what order'; Context Taxonomy answers 'what context made up this particular logical model call' as a per-call snapshot.
- Does it see every call?
- No — it observes ordinary agent-loop calls that reach its llm/stream listener; auxiliary calls are excluded and calls made before installation cannot be reconstructed.
- Where is the data stored?
- It keeps a separately retained, sanitized sidecar so each observed logical call stays independently inspectable.
Alternatives
030611/dsh-context-provenance · bowenliang123/dsh-context · GooodWei/context-vista