Mission#

Pick one low-stakes recurring agent task — heartbeat email check is a good candidate — and run it on the local Qwen model for at least 24 hours. Measure: latency, success rate, and qualitative output quality vs Opus.

Suggested target#

Heartbeat email check: scheduled every 30 min, looks at unread inbox, decides whether anything needs attention. Currently runs on Opus. Cheap, repetitive, perfect for a local model.

Alternative: cron health summarizer (read ~/.openclaw/cron/runs/ for the last hour, summarize in one line).

Steps#

1. Pick the target task (whichever is most active and least sensitive).
2. Create a duplicate cron job that runs the same payload but with model: "qwen" (the alias from t-039).
3. Run both side by side for 24h.
4. After 24h, compare:
   • Success rate (did it complete cleanly?)
   • Latency (Ollama is local — should be faster on small calls)
   • Output quality (read 5–10 random responses from each, assess whether the local model is good enough)
5. Decision: if local is good enough, kill the Opus version and route the production cron to local. If not, document what failed and what would need to change.

Acceptance criteria#

• Target task chosen and documented
• Side-by-side cron jobs running
• 24h of data collected
• Comparison report posted (Discord channel, or workspace memory log)
• Decision made + executed (kill one, keep the other)

Notes#

• Don't pick anything that delivers to Maxime's WhatsApp directly. Reliability bar is too high for a first test.
• If the local model fails partway, don't crash — it should fall back to the original behavior (or fail loud and clear, not silent).