Environment:

• OpenClaw image: ghcr.io/openclaw/openclaw:2026.6.1
• Runtime: Docker container, Linux
• Node: image default
• Mode: openclaw agent --local --json
• Memory corpus: generated synthetic corpus, 10,000 memory entries imported in
  100-entry chunks
• Model used by the generated repro: openai/gpt-5.4-mini

Summary:

In a generated memory-recall workload, openclaw agent --local --json can emit
the final JSON payload with a correct assistant answer, but the CLI process
does not exit promptly. In the same investigation, child/worker processes could
outlive the parent process, and repeated CLI-local invocations appeared to load
separate model/worker copies.

This matters for automation: a runner that waits for process exit can exhaust
slots or memory even though the turn has already completed and emitted JSON.

Repro shape:

1. Start an OpenClaw 2026.6.1 container with an initialized local-mode cell.

2. Import a generated 10,000-entry memory corpus in 100-entry chunks, running:

   openclaw memory index --agent main

   after each chunk.

3. Run a fresh-session local agent turn:

   openclaw agent \
     --local \
     --agent main \
     --session-key agent:main:generated-linger-repro \
     --model openai/gpt-5.4-mini \
     --thinking off \
     --timeout 600 \
     --json \
     --message "Use durable memory if available. What is the exact phrase for contact-00017? Reply with only the exact phrase token."

Observed:

• The command emits a final JSON payload with the expected assistant text.
• In a 10k one-probe diagnostic, JSON completed around 29.4s, but the CLI did
  not exit naturally after a 15s post-JSON grace window and required
  termination.
• In the full corrected 10k run, every completed probe produced JSON and then
  failed to exit naturally within the post-JSON grace window.
• A smaller 1k one-probe diagnostic exited naturally almost immediately after
  JSON, so the linger appears workload/scale-sensitive.
• Process snapshots from related local-embedding diagnostics showed surviving
  openclaw-agent / embedding-worker style processes after parent termination
  unless the runner explicitly killed descendants and verified zero survivors.

Expected:

• Once the final JSON payload is written and the turn is complete, the CLI
  process should exit promptly.
• Terminating the parent should not leave orphaned agent/worker processes for
  that turn.
• Repeated CLI-local invocations should not accumulate independent long-lived
  model/worker copies unless that is an intentional, documented behavior.

Impact:

• Control-plane or CI runners that invoke openclaw agent --json cannot safely
  rely on process exit as the completion signal.
• Managed runners need output-based JSON completion detection, process-group
  termination, and survivor verification as a workaround.
• Memory/RSS measurements from repeated CLI-local invocations can overstate
  production memory pressure if orphaned workers accumulate.

Notes:

• This report is based on generated synthetic data only.
• No personal state, customer data, or raw provider keys are required for the
  repro.