dsh-habit

Self-learning habit engine for the DeepSeek Harness — observes user-correction signals from session events, judges habits with a low-cost model on threshold, and settles candidates behind a two-level human gate. No new agent role: the judgment is an event-driven plugin, immune to context decay.

The loop

① observe   session/event → correction-signal detection (deterministic, zero-token)
② judge     >=3 signals in one session → one flash call (evidence slices + existing habits)
③ settle    candidate zone → user confirms → dsh-memory remember() (suggested)
            → user confirms again → auto → recall injection

Compose

- id: habit
  name: '@max-null/dsh-habit'

Requires storage and llm in the host composition (dsh-base ships both). Installs as a bundle: dsh plugin --profile <name> add @max-null/dsh-habit.

Service

• ctx.habit — the engine:
  • snapshot() → candidates (newest first)
  • confirm(id) / discard(id) → first-level human gate
  • (the second gate is dsh-memory's own suggested→auto confirmation)

Config

Field	Default	Meaning
signalThreshold	3	Correction signals before one judgment call
provider	deepseek-official	Judgment model provider
model	deepseek-v4-flash	Judgment model (cheap, deterministic)
storageRoot	$DSH_HOME/storages/habit	JSON storage root

Design notes

• Deterministic observation, LLM on demand: correction detection is a fixed phrase list + length cap (task descriptions are not corrections); the LLM only runs when a session accumulates enough signals.
• Two-level human gate: candidates must be confirmed in the UI AND then pass dsh-memory's own suggested→auto gate. The model can never promote its own habits.
• Narrow input for quality: the judgment call gets at most 5 evidence texts plus the existing habit list — judgment quality comes from precise context, not volume.

Develop

npm install --legacy-peer-deps
npm test
npm run typecheck
npm run build