openclaw - 💡(How to fix) Fix [Bug]: heartbeat session at agent:main:main pins model from creation time, ignores openclaw.json edits across runs (follow-up to #51677) [1 comments, 2 participants]

Summary

The autonomous heartbeat session keyed agent:main:main continues to use the model that was current at the session's creation time, even after agents.defaults.model.primary is changed in openclaw.json and the gateway is restarted. #51677 closed the equivalent issue on the reply path; this report is a heartbeat-path follow-up.

Reproduction

1. Run gateway on v2026.4.15. Config: agents.defaults.model.primary = "google/gemini-3.1-flash-lite-preview", fallbacks include google/gemini-3.1-pro-preview. Heartbeat enabled (HEARTBEAT.md in workspace).
2. Wait long enough for the autonomous heartbeat tick to create the agent:main:main session. The session jsonl opens with a model_change event pinning provider/model. The sessions.json entry caches modelProvider: "google", model: "gemini-3.1-flash-lite-preview".
3. Edit openclaw.json: change agents.defaults.model.primary to openai-codex/gpt-5.4 and clear fallbacks.
4. Restart gateway.
5. Observe subsequent heartbeat ticks. They continue to use gemini-3.1-pro-preview (the fallback baked in at session creation) per the model_change event in the existing transcript.

Logs show:

[gateway] agent model: openai-codex/gpt-5.4
...
{"api":"google-generative-ai","provider":"google","model":"gemini-3.1-pro-preview","usage":{"input":2123,"cacheRead":371407,...}}

i.e. the gateway-level resolved model is openai-codex/gpt-5.4, but the heartbeat session continues to call gemini.

Workaround

Manually reset the session:

mv ~/.openclaw/agents/main/sessions/<id>.jsonl{,.reset.$(date -u +%s)}
python3 -c '
import json
p = "/home/ubuntu/.openclaw/agents/main/sessions/sessions.json"
d = json.load(open(p))
del d["agent:main:main"]
json.dump(d, open(p, "w"), indent=2)
'

Next heartbeat creates a fresh session and picks up the new primary correctly.

Why this isn't covered by #51677

#51677 was closed citing get-reply.ts and model-selection.ts (the reply path). The heartbeat is a different code path: an autonomous cron-style turn loop that resumes the existing agent:main:main session jsonl on every tick. That path replays the leading model_change event as part of the transcript context, and that event is what determines the model for the LLM call.

It's possible v2026.4.22 fixes this incidentally; I haven't verified the heartbeat path on v2026.4.22 yet because I had already reset the session before rolling forward.

Bonus issue in the same session

While investigating I also noticed the heartbeat session had compactionCount: 0 and a 2.27 MB transcript covering 9 days of HEARTBEAT_OK ticks, replayed in full on each tick on top of a 47k-character bootstrap. Cache hits softened the bill but each tick was still pushing 370k+ cached input tokens. Not the same bug, but worth flagging.

Environment

• OpenClaw 2026.4.15 (originally observed); rolled to v2026.4.22 after manual reset
• Linux x86_64, Node 24.14.1, user-level systemd unit
• Channels: telegram (long-poll)
• Heartbeat enabled via workspace HEARTBEAT.md