dsh-plugin-auto-blame: A Critical Follow-up Suggestions Plugin for DeepSeek Harness
huanlinoto/dsh-plugin-auto-blame
DSH plugin generates 3 critical follow-up suggestions after a model turn, click to send.
Generates 3 critical follow-up suggestions for each model turn in DeepSeek Harness, clickable to send, encouraging users to think deeper or challenge the model's output.
dsh plugin --profile web add github:huanlinoto/dsh-plugin-auto-blame- Category
- Sessions & Messages
- Platform
- DSH-Plugin
- Author
- huanlinoto
- Distribution
- Plugin
dsh-plugin-auto-blame Key Features
dsh-plugin-auto-blame Repository
dsh-plugin-auto-blame Repo Summary
dsh-auto-blame is a session enhancement DSH plugin for DeepSeek Harness, maintained by HuanLinOTO, written in TypeScript, and last updated in 2026-08. It addresses the lack of critical feedback after a model's single-turn response: when the model completes a turn, the plugin calls an LLM to generate 3 critical follow-up suggestions, displayed as clickable chips above the input box. Clicking one sends it directly, encouraging users to think deeper or challenge the model's output. Core capabilities include non-blocking fire-and-forget LLM calls, event folding and pushing via session/projection, click-to-send reusing the InputBar path, and silent failure with automatic clearing on the next turn.
How does the dsh-auto-blame DSH plugin generate critical follow-up suggestions?
After the agent/turn-stopping event, the plugin sends the last 3 messages to the LLM, generating 3 critical follow-ups via ctx.llm.stream(). The call is fire-and-forget, not blocking turn closure, so suggestions appear a few hundred milliseconds later. Results are written via session.append('auto-blame/suggestions', ...), then folded by a projection unit into session/projection frames; the client receives them via useProjection('autoBlame') and renders chips. These suggestions are not part of the model-visible history and do not interfere with the agent loop. If the LLM call or parsing fails, nothing is shown.
How to install and run dsh-auto-blame?
Installation from npm is recommended, using the following command:
dsh plugin --profile web add github:HuanLinOTO/dsh-plugin-auto-blame
For local development with hot reload, you can use a link path. After installation, restart dsh web and hard-refresh the browser (Ctrl+Shift+R) to load the plugin.
What are the development and test commands for dsh-auto-blame?
The project provides complete development scripts: typecheck uses tsc --noEmit, tests use vitest run (28 unit tests), and build uses tsc + tsdown to produce the lib directory. It depends on @deepseek-ai/dsh-* workspace packages; during development you need to manually create Junction links to a sibling dsh checkout, or copy node_modules from dsh-spur and add the required package Junctions.
What are the interaction experience features of dsh-auto-blame?
While generating, the "leadership view" label above the input box shows a DeepSeek blue shimmer sweep (same effect as Deep diving...), and chips fade in sequentially when suggestions arrive. Clicking a chip sends it via inputActions.setDraft(text) + submit(), the same path as manual input. On a new turn, the projection resets to zero and old chips disappear immediately, keeping the UI clean.