[FEATURE]: Add Codex / Claude Code style /goal autonomous loop command #49205

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Request a native /goal command for opencode with the same semantics as OpenAI Codex and Claude Code, for long-horizon autonomous work.

Reference behavior:

1. Codex /goal (experimental, goals = true under [features]):

• /goal <objective> sets one durable objective, /goal shows status, /goal pause / /goal resume / /goal clear controls the run.
• Stays active across turns for multiple hours until achieved, paused/cleared, blocked, or budget-limited (pursuing, paused, achieved, unmet, budget-limited).
• Intended for code migration, large refactors, prototypes where the contract defines objective, non-goals, validation commands, and stop condition.
• Docs: https://developers.openai.com/codex/use-cases/follow-goals

2. Claude Code /goal (native since v2.1.139):

• /goal <condition> / /goal / /goal clear (aliases: stop, off, reset, none, cancel). One goal per session, ◎ /goal active indicator, restored on --continue / --resume, cleared on /clear.
• After each turn a small fast model (Haiku by default) judges the condition against the conversation transcript only (no independent tool calls). Not yet met -> continue with reason as guidance; Met -> clear + achieved entry; Impossible -> clear + failed entry.
• Does not change permission mode; use auto-accept for unattended runs. Works in interactive, -p, and remote control.
• Docs: https://code.claude.com/docs/en/goal , https://code.claude.com/docs/en/commands

Current gap in opencode:

Today a 200-file migration or full test-suite fix requires typing "keep going" every few minutes. Custom slash-command templates cannot do persistence, runtime continuation after turn end, token/time accounting, evaluator loop, or TUI/desktop status sync.

I am aware of #27167 (native session goals, open, 200+ reactions). This request explicitly asks for parity with the Codex/Claude evaluator-loop semantics above, not just goal text storage:

• /goal <objective> / /goal / /goal pause / /goal resume / /goal clear
• session-scoped single active goal with status projection (turns elapsed, tokens spent, last evaluator reason)
• per-turn evaluator using a configurable small/fast model against transcript, with met / not-yet-met / impossible handling
• idle continuation in TUI + opencode run non-interactive parity, human input always wins, /clear clears goal
• optional per-goal token budget + wall-clock cap and budget-limited state (cf. Codex token_budget discussion)
• model-visible goal tools with restricted mutation (complete only after verification), shared server API so TUI/desktop stay in sync

Benefits:

• Removes the human-enter bottleneck for long runs.
• Verifiable "done" contract instead of model claiming done.
• Competitive parity: users coming from Codex / Claude Code expect /goal to work the same way.