Goal mode for v0.8.43
Goal mode should replace the narrow /goal command over time with a persistent objective/workflow surface. It is not another meaning of "auto" and it is not a new TUI execution mode — Plan / Agent / YOLO remain the three execution/approval modes. /model auto remains model/thinking auto-routing.
Design principles
- Persistent objective surface. Goal mode is a durable objective + workflow tracker that the agent can see and work toward across turns. It outlasts a single command.
- Clarifying questions. When the objective is underspecified, the agent should ask clarifying questions before starting — not guess and thrash.
- Autonomous progress. The agent should keep working toward the goal until complete, blocked, or explicitly stopped by the user.
- Fin as wakeup/verifier. A cheap
deepseek-v4-flash thinking-off pass (Fin) acts as a lightweight wakeup and completion verifier: it checks whether the work is done from a binary perspective — tests, build, changed files, docs, install state, release gates, and user acceptance criteria.
- Plan / Agent / YOLO preserved. These remain the execution/approval modes. Goal mode influences what the agent works toward, not how tools are approved.
/model auto preserved. Model auto-routing stays separate from Goal mode. The two concepts serve different purposes and should never be conflated.
Relationship to existing /goal
The current /goal slash command sets a session objective with an optional token budget — it is goal tracking, not a mode. Goal mode would subsume this over time, giving the /goal command a richer backend surface.
Out of scope for v0.8.43
- New TUI execution modes. Plan / Agent / YOLO are the three.
/model auto changes — that stays model routing, not workflow management.- Full spatial workbench (deferred to v0.8.44).
Related
Goal mode for v0.8.43
Goal mode should replace the narrow
/goalcommand over time with a persistent objective/workflow surface. It is not another meaning of "auto" and it is not a new TUI execution mode — Plan / Agent / YOLO remain the three execution/approval modes./model autoremains model/thinking auto-routing.Design principles
deepseek-v4-flashthinking-off pass (Fin) acts as a lightweight wakeup and completion verifier: it checks whether the work is done from a binary perspective — tests, build, changed files, docs, install state, release gates, and user acceptance criteria./model autopreserved. Model auto-routing stays separate from Goal mode. The two concepts serve different purposes and should never be conflated.Relationship to existing
/goalThe current
/goalslash command sets a session objective with an optional token budget — it is goal tracking, not a mode. Goal mode would subsume this over time, giving the/goalcommand a richer backend surface.Out of scope for v0.8.43
/model autochanges — that stays model routing, not workflow management.Related