tancheng33/dsh-yogacara
Tracks the agent's state as Yogācāra mental factors over a durable seed store, and feeds it back into the prompt as first-person inclination.
A self-model plugin for DeepSeek Harness built on the Yogācāra account of mind: eight consciousnesses, the 51 mental factors, a store of seeds perfumed by what happens, and a self that is measured rather than assumed — written back into the agent's own system prompt each turn. State variables are computed from harness events by rules you can read in src/citta.ts and disagree with, and a factor halves on a configurable half-life (a factor in five minutes, a seed in two weeks by default). The point is actionability: the taxonomy is a closed enumeration grouped by valence in which every affliction already names the factor that counteracts it, so the self-model arrives with its own regulation loop.
Install
pnpm add dsh-yogacaraREADME Install: pnpm add dsh-yogacara; adding it to your profile's dsh.profile.bundles is enough, because the bundle contributes the defaults shipped in cordis.patch.yml and your own profile patch overrides any key. To track main instead: pnpm add github:tancheng33/dsh-yogacara. The published package ships prebuilt lib/, so installing it needs no build-approval step.
Compatibility
DSH profile bundle. Configurable keys in the profile patch layer: domain, observeChat, observeTools, promptSection, awareness (felt | report | silent), promptMaxFactors, manasWarning, halfLifeMs, seedHalfLifeMs, maxSeeds, flushIntervalMs.
Details
- Repo: tancheng33/dsh-yogacara
- Category: Memory
- Stars: 0
- Version: 0.2.1 (npm dsh-yogacara; registry repository field → github.com/tancheng33/dsh-yogacara, verified 2026-09-12)
- Last push: 2026-08-26
- First seen: 2026-08-16
Recent updates
README is explicit that nothing here claims the agent is sentient or that the numbers are feelings: they are named state variables, and the prompt block says so to the agent in those words. Four tools are registered — self_reflect (stop and look at how things are, and what the store carries for a named situation), self_appraise (record how something landed, with a nature calibration axis of assumed / inferred / verified), self_recall (search the store for precedents before repeating an approach) and self_transform (turn an affliction into its wisdom and commit to the counter-move).
FAQ
- How do I install dsh-yogacara?
- Run pnpm add dsh-yogacara and add it to your profile's dsh.profile.bundles. The published package ships prebuilt lib/, so no build approval is needed.
- Does dsh-yogacara claim the agent is sentient?
- No. The README states the numbers are named state variables computed by rules you can read and disagree with, and the prompt block tells the agent exactly that.
- What tools does dsh-yogacara add?
- Four: self_reflect, self_appraise (with a nature axis of assumed / inferred / verified), self_recall and self_transform.