[FEATURE] Make /goal model-invocable (let the agent set its own Stop-condition)

Summary

/goal <condition> registers a session-scoped Stop hook that blocks the agent from ending its turn until the condition holds. Today it is user-only — typed into the prompt. The agent itself cannot set, update, or clear a goal.

Requesting a model-invocable path: a tool (or a model-invocable form of the command) that lets the agent register its own Stop-condition mid-task.

Precedent

This was in the original design and was cut:

• #56085 (closed/completed) — the request that shipped /goal. Its proposal included set_goal/update_goal as model tools so the agent could manage its own objective. Shipped user-only; the model-tool portion was dropped and never separately tracked.
• #46524 (closed/stale) — broad "make every slash command agent-invocable" ask. Covers this by scope but stalled.

This issue re-files the narrower, higher-value slice: /goal specifically.

Motivation

A goal/Stop-condition is uniquely valuable to set from inside a task, where the agent — not the user — is the one who knows what "done" means:

• Self-enforced completion. Agent commits to an exit criterion before a long multi-step task, then can't prematurely declare victory — the same discipline /goal gives users, applied where the agent has the context.
• Decomposition. On entering a sub-task, the agent arms a condition for it and clears it on completion, without bouncing control back to the user.
• Loop/skill authoring. Skills that drive iterative work could install their own convergence condition rather than relying on prose instructions the model can rationalize past.

Proposed shape

A model-invocable tool, e.g. SetGoal { condition, action: set|update|clear }, subject to the same Stop-hook evaluator /goal already uses. Reuses existing machinery; only the invocation surface is new.

Guardrails worth considering:

• Cap concurrent self-set goals to avoid runaway/self-deadlock.
• Make self-set goals user-visible and user-clearable (so the human can always override).
• Gate behind the existing model-invocation flag system (cf. disable-model-invocation, #29226/#50075).

Acceptance criteria

• Agent can register a Stop-condition without user typing /goal.
• Self-set goals are surfaced to the user and clearable by them.
• Same Stop-hook evaluation as the user-typed path.

---

Out of scope but adjacent: #46524's general slash-command-invocability ask — flagging in case maintainers prefer to solve the broad version.