Session Goal

Session Goal

Session Goal lets you set a persistent objective for a single conversation. After each turn, an independent judge model decides whether the goal is complete; if not, a continuation prompt is injected automatically until the goal is met, the turn budget is exhausted, or the user pauses/clears.

Commands

Command	Description
/goal <description>	Set goal and start first run (default 20-turn budget)
/goal status	View goal, subgoals, turn count, judge reason
/goal pause	Pause auto-continuation (keep goal state)
/goal resume	Resume auto-continuation
/goal clear	Clear goal
/subgoal	List subgoals
/subgoal <condition>	Append sub-condition
/subgoal remove <N>	Remove subgoal N (1-based)
/subgoal clear	Clear all subgoals

On Chat, terminal commands (/goal status, pause/resume/clear, /subgoal list, etc.) run via Habitat RPC conversation.command and show in a result panel or toast — they do not flash through the message stream. Setting a goal (/goal <description>) and /retry still use message.send so the LLM turn streams into the conversation. Discord/Weixin keep the existing slash → stream reply path.

Workflow

1. User runs /goal … → writes conversations.goal and triggers engine run.
2. After each assistant reply → goal judge reads goal, subgoals, recent dialogue, last reply.
3. Judge outputs strict JSON: {"done": boolean, "reason": "..."}.
4. done: false → inject user-role continuation (e.g. ↻ Continuing toward goal (3/20): …), continue next turn in same SSE stream.
5. done: true → mark completed, stop continuation.
6. User message mid-run preempts current continuation; after that turn, re-judge; continues if not paused.

Judge conservative policy

• Pass: assistant clearly confirms completion, shows final deliverable, or states blocker needing user input (reason explains).
• Fail: vague progress, plan only without evidence; implied completion does not count.
• Judge unavailable: call or parse failure pauses the goal (status: paused), logs a warn, and appends a visible status line to the conversation (not a continue prompt). Use /goal resume to retry. Check /goal status for last_judge_reason.

Configuration

Optional goal_judge in Habitat runtime llm.profiles (Shell Settings → Habitat 服务 → 服务配置). Falls back to llm.default_profile when unset.

Judge calls use OpenAI-compatible response_format: json_object and thinking: { type: "disabled" } (via params.extra), plus maxOutputTokens: 2048. Prefer a model/gateway that supports JSON mode; unsupported endpoints will fail the judge call and pause the goal.

vs ACP

• Goal: in-conversation synchronous continuation loop; orchestrated by platform turn lifecycle.
• ACP: external agent async task; callback triggers separate turn on completion.

See architecture.md.