feat: auto-abort zombie runs idle > 60 min via gateway chat.abort

[rjcloudsigma]

Summary

Extends the stuck-run detector (PR #10) with a zombie auto-abort background task.

What it does

• Every 60s (configurable), scans agent sessions for zombie runs (idle ≥ 60 min)
• When TAAS_AFFINITY_AUTO_ABORT_ZOMBIES=true (off by default), aborts via injected abortRun(sessionKey) callback
• Tracks aborted sessionKeys in an in-memory LRU-bounded array (cap 1000) for idempotency
• Dry-run mode via TAAS_AFFINITY_AUTO_ABORT_DRY_RUN=true — logs what it would abort without calling abort

Env vars

Variable	Default	Description
TAAS_AFFINITY_AUTO_ABORT_ZOMBIES	false	Enable auto-abort (must be set to true)
TAAS_AFFINITY_AUTO_ABORT_THRESHOLD_MS	3600000 (60 min)	Idle threshold for zombie classification
TAAS_AFFINITY_AUTO_ABORT_CHECK_INTERVAL_MS	60000 (1 min)	Background check interval
TAAS_AFFINITY_AUTO_ABORT_DRY_RUN	false	Log candidates without aborting

⚠️ BLOCKED

The OpenClaw plugin SDK does not expose dispatchGatewayMethod, api.runtime.chat.abort(), or any equivalent for calling gateway RPC methods from inside a plugin. The default abortRun function is a no-op that logs a warning. The detection and logging infrastructure is complete and tested — when the SDK adds a gateway dispatch capability, wire it via setAbortRunFn() in the register() method.

Tests

• AC-AUTO-ABORT.1: Auto-abort enabled → calls abortRun for each zombie
• AC-AUTO-ABORT.2: Auto-abort disabled (default) → logs candidates without aborting
• AC-AUTO-ABORT.3: Idempotent — same zombie only aborted once across ticks
• AC-AUTO-ABORT.4: LRU cap — 1100 zombies processed, set stays ≤ 1000
• AC-AUTO-ABORT.5: Dry-run mode — logs but never calls abortRun

Commits

1. feat: detect zombies and emit candidates list (no-op default)
2. feat: add BLOCKED note — no plugin SDK dispatch for chat.abort yet
3. test: zombie auto-abort acceptance coverage