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