Summary

Adds time-based compaction that triggers when a session has been idle beyond a configurable threshold (idleCompactMinutes). This addresses stale context in sessions that go through periods of inactivity — overnight, between work sessions, or during long-running background jobs.

Motivation

Currently, LCM only compacts when context exceeds the contextThreshold (turn-based trigger). Sessions that accumulate significant context but never quite hit the threshold — or sessions that go idle for hours/days — can resume with stale, unorganized raw context. This is especially common in agentic workflows with periodic cron/watchdog messages.

Idle compaction ensures returning users get lean, organized context instead of a wall of raw messages from hours ago.

Enables practical use of large context windows (1M tokens). With models like Gemini offering 1M-token context at no pricing premium, the turn-based threshold trigger alone isn't sufficient — sessions can accumulate hundreds of thousands of tokens of raw conversation over hours without ever hitting the 75% threshold. Idle compaction provides the missing time-based dimension: after periods of inactivity, context is proactively organized into the summary DAG regardless of size, keeping large-window sessions navigable and performant over long lifespans.

How it works

1. Config: idleCompactMinutes (env: LCM_IDLE_COMPACT_MINUTES, default: 0 = disabled)
2. Detection: getLastActivityTimestamp() queries the most recent user or assistant message. Heartbeats (never ingested) don't reset the timer. SystemEvents and sub-agent completions do.
3. Leaf pass in assemble(): When idle threshold is exceeded, runs a single compactLeaf() with force: true before assembling context. Protected by a 15-second timeout to bound first-message latency.
4. Full cascade in afterTurn(): Flags the session so afterTurn() forces a full compaction sweep regardless of context threshold, ensuring the leaf pass doesn't leave the DAG in a partial state.

Changes

Configuration

{
  "plugins": {
    "entries": {
      "lossless-claw": {
        "config": {
          "idleCompactMinutes": 180
        }
      }
    }
  }
}

Or via environment: LCM_IDLE_COMPACT_MINUTES=180

Idle timer policy

• User/assistant messages → reset timer ✅
• Heartbeats → do NOT reset (never ingested) ✅
• SystemEvents (cron, watchdog) → reset timer (genuine new information) ✅
• Sub-agent completions → reset timer ✅

Testing

Tested in production OpenClaw deployment with idleCompactMinutes: 180 across multiple concurrent sessions (main, autoresearch, data collection). Gateway restart successful, config persisted, no errors in logs.

Breaking changes

None. Default is 0 (disabled). Existing behavior unchanged unless explicitly enabled.