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

As suggested by the bot, I'm opening a V2 issue — I was in #16626, carrying this over since that closed when the bot treated it as V1-only and ignored the follow-ups — including #54487. This is V2-only (dev branch). I've got that PR open with a working implementation.

description

experimental.idle fires after Complete() would return, with only { sessionID, background } — no vote, no bound. A plugin that wants to keep a session alive has to self-bound its loop.

Want experimental.stopping fired before Complete() in packages/core/src/session/runner/llm.ts (maybeContinueStopping), with { sessionID, step, stepsLimit, continuations, decision }, vote shouldContinue + text, capped at 2 continuations/turn.

Types: packages/plugin/src/effect/session.ts + promise mirror. Example consumer contrast: packages/core/src/plugin/goal.ts (uses experimental.idle today).

What it's for

pre-terminal veto, not idle notification. Not a replacement for packages/core/src/plugin/goal.ts: that owns the multi-round objective loop via experimental.idle. stopping is the bounded single-turn gate idle can't express:

• sql check before success: validate the SQL the turn produced, on failure force one more step with a short fix instruction. Needs a bound so an always-fail validator can't loop, and default-allow-stop when checks pass. idle is too late because Execution.Succeeded is already published.
• typecheck before end: run check-types and inject errors so the agent fixes them in the same session. idle re-prompt shows as a visible user message instead of a silent continuation.
• verified-claims gate: scan the final message for high-risk claims and re-enter once with a correction. Emulated today via idle → messages() → prompt(), but idle also fires on abort, the correction pollutes the transcript, and the user can act on the answer during the round-trip.
• opencode-dev-framework (my plugin, downstream waiter): completion gate runs today on experimental.idle → session.synthetic, bounded by gate.max_blocks, plus a deduped event.subscribe net for the boundaries the hook skips. With stopping the gate becomes a vote with runner-owned budget instead of self-accounting + net.

Without this, every gate re-implements its own bound on top of idle, after the runner already decided to stop.