Add plan-mode-style pause/suppression for active ralph loops

[charfeng1]

Problem

Once a ralph loop is active, the plugin continues on every session.idle event regardless of what the user wants to do. There's no way to pause the loop to think or chat without /cancel-ralph (which destroys state).

Reference: Codex /goal Plan mode

Codex's equivalent feature (/goal) handles this with collaboration modes. When the user is in Plan mode, goal continuations are suppressed entirely — but the goal itself stays alive. The user can flip back to Default/Execute to resume.

Key behaviors when continuations are suppressed:

• Idle events do not inject a continuation prompt
• Token/time accounting against the goal pauses
• Interrupts don't auto-pause (no active accounting to pause)
• Resume restores from the last accounted state

Enforced by a single predicate checked at every continuation gate (see codex codex-rs/core/src/goals.rs::should_ignore_goal_for_mode).

Proposal

Add a paused: true|false field to the state frontmatter. Two new actions:

• /ralph-pause — set paused: true. Loop stops auto-continuing on idle but state persists.
• /ralph-resume — set paused: false. Loop resumes from current iteration count.

The event handler in src/index.ts gains one early-return guard:

```ts
if (state.paused) return;
```

Open questions

• Should pausing also reset the iteration counter, or preserve it? (Codex preserves accounting; argues for preserve.)
• Should there be a single /ralph command with subcommands (pause, resume, status) instead of two commands, for tidiness? Codex packs all of those under /goal.
• Should pause auto-trigger on user interrupt (Ctrl-C)? Codex does this. Probably yes — interrupting a ralph loop should pause, not cancel.

[Keesan12]

A pause should be a first-class run state, not a softer cancel.

On pause I’d freeze the remaining budget, persist the current objective/context hash, stop new tool calls, and write a checkpoint receipt. Resume should require an explicit transition from that checkpoint so an idle event cannot silently restart the loop after the user changed direction.

We are implementing this kind of outside-the-agent state control in MartinLoop. If you land pause semantics, I’d be happy to test a real stop/chat/resume workflow and report where state or authority leaks across the boundary.

[Keesan12]

Pause/resume without destroying state is a real production need. Most Ralph-style loops either keep burning tokens after the user steps away or force a full cancel.

MartinLoop treats budget, verifier, and run state as first-class so a governed loop can be paused, inspected, or rolled back cleanly instead of left running unattended.

https://github.com/Keesan12/martin-loop

If independent AI safety and governance is important to you, please help us and ⭐ the repo to keep the OSS core free.