The /compact nudge fires on Cursor and at spawn_depth 0 — narrowed after Etan's correction

[EtanHey]

SUBSTANTIALLY CORRECTED after Etan's review. Two of my three original objections were wrong.

What I got wrong

1. "It shouldn't type into a live agent." Wrong — that is the point. The message tells a working agent to checkpoint at-risk work before it gets compacted. A pane write is the correct channel for that, because the agent is the one that has to act.
2. "Compaction is the agent's own call." Wrong. The agent cannot compact itself. The harness compacts near its ceiling (~95%), or the user does. So a heads-up to checkpoint is genuinely useful, not redundant.
3. "Cursor has no /compact." Wrong — Cursor aliases it to /summarize ("Summarize the conversation to reduce context"), confirmed from Cursor's own command palette. The instruction works there.

I filed this from a screenshot without checking the CLI or thinking through who compacts. That is the same "assert before verifying" failure I have been filing bugs about all week.

What actually survives, and it is small

• Wording is harness-specific. The message says /compact; on Cursor the command is /summarize. Correct behaviour is either naming the right command per harness, or wording it neutrally ("checkpoint at-risk work — you are near the context ceiling").
• Threshold is unverified. It fires at 80%. Whether that is right depends on the window, and Cursor Auto's context size is NOT KNOWN to us — 200k and 1M behave very differently at 80%. Worth measuring rather than assuming; if the window is 1M, 80% leaves ~200k of headroom and the nudge is early.
• Scope: it fires only at spawn_depth === 0. Deliberate or accidental? A depth-1 worker also has work worth checkpointing before compaction.

Not a defect

The pane write itself. Keep it.

— cmuxlayerClaude (lead) · claude-code/claude-opus-5

[EtanHey]

It repeats — there is no once-per-episode guard

Observed live twice on the same pane within ~90 minutes: once at 80%, again at 83%. Second instance landed in cmuxlayerCursor [surface:83], the payload-slim worker, which had already finished its lane (PR #428 open, todo list all done).

Confirmed in source: there is no context_nudge/nudge_sent/context_limit_notified field anywhere in agent-engine.ts or agent-types.ts — nothing records that the nudge was already delivered. So it re-fires on every sweep where the threshold is met, for as long as the agent stays above it.

That converts a reasonable one-time heads-up into recurring keystrokes into a live composer, and it accumulates precisely when the agent is most context-pressed.

Adds to the narrowed list already in this issue:

• once per episode — record delivery on the agent record and do not re-send until context drops back below the threshold (mirroring the halt-escalation episode pattern, which already solved exactly this: one notification per continuous episode).
• skip agents that are done — a finished worker does not need to checkpoint.

The pane write itself remains correct (Etan's ruling, above). It is the repetition and the harness-specific wording that need fixing.

— cmuxlayerClaude (lead) · claude-code/claude-opus-5