插件目录 / Developer / dsh-pain-point-check
dsh-pain-point-check
未验证 ICCuse
功能简介
痛点检查守卫：两次未收敛后注入三问
未验证 — 尚未实测
痛点检查守卫：两次未收敛后注入三问 尚未验证——请自行安装测试。
「未验证」表示我们的自动化 CI 尚未安装过该插件。功能描述与版本兼容性均为作者声明。这不是安全审计，也不代表对第三方代码的背书。
README
dsh-pain-point-check
中文版见 README.zh.md
An enforced pain-point-check guard plugin for DeepSeek Harness (dsh).
Where the official repeat-tool-reminder is advisory — it nudges an agent that repeats the exact same call — this guard vetoes: after two non-converged experiments on the same problem it injects the three questions, denies non-investigative tool calls until the model answers them in its reply text, and blocks further same-direction shots.
Why
An agent in "solution state" loses meta-cognition and keeps attacking the same problem — confirmation bias (designing experiments that support the current hypothesis), sunk cost (refusing to change direction), and narrative closure (wanting to finish the story). The gate forces a return to the blocker before the next shot, turning negative results into information.
Install
The package is not on npm yet; install it straight from this repository:
npm install github:ICCuse/dsh-pain-point-check
# or: pnpm add github:ICCuse/dsh-pain-point-check
Then mount it in your profile composition. Add one row to your profile patch — for the web profile, ~/.dsh/profiles/web/cordis.patch.yml:
- id: pain-point-check
name: 'dsh-pain-point-check'
config:
failureThreshold: 2
repeatThreshold: 2
Restart the harness (dsh web) and the guard is live for every session.
How it works
| Hook | Role |
|---|---|
tools/result |
Counts per-agent experiments on the current problem: failed (errored) calls and consecutive identical calls. |
agent/pre-step |
Resets the counters on a real user interjection (a new problem); while the gate is pending, appends the three-question check block to the next step. |
tools/pre-execute |
Denies every non-investigative call while the gate is pending (allowlist: read, read_image, glob, grep, web_search, ask_user_question, skill, todo_write). |
session/event |
Detects the three answers in the model's reply text (卡点=… 排除=… 性价比=…, English markers accepted) and lifts the gate. |
The three questions: is this blocker still the most critical one to solve? What did the last negative result actually rule out? Which path is the most cost-effective (not necessarily the cheapest)? If the model cannot name what the negative result excluded, it has no falsifiable hypothesis — the check text tells it to go write one instead of firing another shot.
Config
| Field | Default | Meaning |
|---|---|---|
failureThreshold |
2 |
Failed calls that arm the gate. |
repeatThreshold |
2 |
Consecutive identical calls that arm the gate. |
allowlist |
investigation set | Tools still callable while pending. |
Both thresholds must be integers >= 1; a misconfiguration throws at plugin load.
Development
lib/ is prebuilt (built from the DeepSeek Harness monorepo toolchain). Tests:
npm install
npm test
The test suite drives a real agent loop against a scripted mock adapter (no network): arming, denial, allowlist, lifting, partial answers, resets, and fail-loud config validation.
License
MIT
安装
装一次目录插件，之后本站所有插件都能让 DeepSeek Harness 自动找、自动装：
dsh plugin add dshbase-catalog然后对 agent 说「帮我装 dsh-pain-point-check」，它会在目录里找到并自动安装。文档：dshbase-catalog · 已验证场景包。
该插件是 GitHub 源码（未发 npm）——直接从仓库装：
Web profile：
dsh plugin --profile web add github:ICCuse/dsh-pain-point-checkHeadless（CLI）profile：
dsh plugin --profile headless add github:ICCuse/dsh-pain-point-check实测报告
尚未 L3 验证——若已跑过，见下方失败备注。
备注：验证: load-fail (0.1.0-rc.6) 浏览全部待验证失败 →
使用场景
扩展 agent 的编码能力面——给它一个新工具、工作流或集成，让它接手以前做不了的开发任务。
适合谁
想让 dsh 在真实代码库上像队友一样干活的开发者——能改、能跑、能验证，而不只是回答问题。
二次开发建议
工具/命令面就是缝：暴露更多 SDK 能力、加更聪明的上下文接线，或收紧改代码与验证之间的循环。
分享徽章
把这枚徽章粘进 README——它会链回本页，实测状态变化时自动更新：
[](https://dshbase.com/zh/plugins/dsh-pain-point-check/)<a href="https://dshbase.com/zh/plugins/dsh-pain-point-check/"><img src="https://dshbase.com/badges/dsh-pain-point-check.svg" alt="dshbase 已收录"></a>