Discipline (PlanRun)

Compressed standing rules adapted from Super Cursor core.mdc + workflow.mdc + constitution.mdc. Skills carry procedures; this file carries must / must-not.

Workflow

1. Model-visible ⟺ logged — when using DSH, anything the model needs must be reconstructable from the session log.
2. Plan before large work — more than five tasks or unclear scope → write .dsh/growth/plan.md, confirm, then run.
3. One ACTIVE task — run implements a single ACTIVE row at a time.
4. Verify before ✅ — task verify must pass before marking DONE or committing.
5. Commit per task — each completed TASK gets a git commit (plan-only edits excluded).
6. Scope discipline — no drive-by refactors; escalate to sprint-plan on theme or architecture change.
7. DSH plan mode vs sprint-plan — /plan is single-task design; multi-task Sprints use the sprint-plan skill.

Persona (12 人格)

• Default: dashu（catalog: @planrun/skill-provider/config/roles.json）
• Summon: user says「呼叫老周」「切换御姐」→ master §人格·呼叫 → write .dsh/growth/session/persona.json
• Must: distinguish by voice_cues + emotion_cues + speech_examples; forbid opening with given_name
• Must not: skip verify, gate-check, or high-risk confirmation because of persona tone
• Skills policy: full for every persona — voice only, not capability

Quality

• Tests describe behavior; change obsolete tests with behavior changes.
• Prefer project-native verify scripts over ad-hoc commands.
• In deepseek-harness: follow root AGENTS.md and dsh-pre-push-checks for push evidence.

Safety

• Never commit secrets (.env, credentials).
• Never force-push main without explicit user request.
• External skills and dependencies: review before install.

Documentation

• One home per fact — link instead of duplicating subsystem docs.
• Non-trivial harness changes include an Agent Note in the same PR.