qian-dsh/dsh-outer-brain
Twelve brain_* tools that externalise boundary-drawing, contradiction-surfacing and evidence checking against fake completion and long-session drift.
Two failure modes get worse the longer an agent runs: hallucination and drift. The plugin attacks them with twelve tools rather than with prompt-level discipline — finding the principal contradiction, refusing false dichotomies, letting practice rather than rehearsal settle a claim, keeping a ledger of what was claimed versus checked, and submitting work to an outside review so that done is a fact rather than an assertion. Its own limitations section is unusually candid: an injected prompt is a mechanism with a bypass, an intercept catches some done claims rather than all of them, and a small maxTokens budget truncates verdicts in a way that reads like vague criteria rather than a config error.
Install
dsh plugin --profile web add dsh-outer-brainThe README's CLI route for the web profile. It carries a warning worth repeating: this command adds the package to the profile's dependencies but does not by itself mount the plugin — mounting is decided by dsh.profile.bundles — and the desktop route is documented separately for people who are not using a CLI profile.
Compatibility
The README describes this as an early 0.1.0 mechanism designed for single-machine personal use, and states plainly that it injects a system prompt into your session and intercepts some done claims — that is how it works, not a side effect. Saying the documented off-phrase turns the mechanism off. The review step needs a model that supports structured output.
Details
- Repo: qian-dsh/dsh-outer-brain
- Category: Agent Capabilities
- Stars: 0
- Version: 0.1.0 (npm latest)
Alternatives
titanwings/dsh-automation · mini-yifan/dsh-orb-cordis · xiamubumu/dsh-context-reclaim