hermes - 💡(How to fix) Fix Standing goals: agent should be able to signal completion and close the goal [1 pull requests]

Problem

When a standing goal is set via /goal, the agent works on it across turns. After each turn, a judge evaluates whether progress was made. If the judge detects no progress, it forces a retry — up to goals.max_turns (default 20) times.

The problem: when the agent has fully completed the goal and says so explicitly (e.g. "目标已完成" / "goal complete"), the judge does not recognize this as a signal to close the goal. Instead, it sees "no new progress actions" and keeps retrying, flooding the user with identical "done" replies until max_turns is exhausted.

Real example

1. User sets a standing goal
2. Agent does all the work, then says "已完成。停止。"
3. Judge: "no progress detected, retry"
4. Agent again: "已完成。停止。"
5. Judge: retry again
6. ... repeats 15+ times before max_turns stops it

Expected behavior

The agent itself should be able to signal "this goal is complete, close it". When that signal fires, the goal is closed immediately — no judge loop, no retries.

Option A: Agent-manageable goal tool

Add a tool (e.g. goal_manage similar to cronjob) that the agent can call programmatically:

goal_manage(action="complete") → closes standing goal, stops retry loop

This gives the agent explicit agency over goal lifecycle, just like it already has with cron jobs via the cronjob tool (pause/resume/remove).

Option B: Judge recognizes explicit completion signals

The judge/evaluator should recognize explicit goal-completion statements in the agent's final response (e.g. "goal complete", "已完成", "done"). When detected, close the goal and stop retrying.

This is simpler but fragile — relies on phrasing patterns and the judge's NLU.

Additional context

• Config key: goals.max_turns (default 20)
• Slash commands: /goal status|pause|resume|clear — none are callable by the agent itself
• The agent already has full lifecycle management for cron jobs via the cronjob tool — standing goals lack equivalent tooling
• Related: #28640 (ACP goal parity) — different scope (ACP transport, not agent lifecycle signaling)

Suggested scope

Minimum viable: a tool or slash command that the agent can invoke to mark a standing goal as complete. The judge retry loop should immediately stop when the goal is closed.

Lowest-effort observation: /goal clear already exists as a user subcommand. If the agent could invoke this programmatically through a structured tool, the loop would end. But the agent has no way to do this, and the system prompt does not inform the agent that it can.