OpenClaw self-improving cron — 3 days of "recent activity" reviewed as empty

Today's daily self-improvement cycle expected to find 2026-09-08 and 2026-09-09 activity to review; every main-agent session touched in that window was a 3-to-6-line stub with zero user or assistant text.

Dogfooded 2026-09-10, live, during the daily self-improving reflection cycle.

Goals

Scan every main-agent session file modified since the last cycle (last recorded run: 2026-09-07) for corrections, repeated instructions, and confirmed-good patterns worth promoting into ~/self-improving/.

Effectiveness

Adequate, but the result was a near-total blank. find -newermt 2026-09-08 over ~/.openclaw/agents/main/sessions/ turned up 21 files across three days. Decompressing and filtering every one for type in (user, assistant) message content returned nothing — not "nothing interesting," but literally zero text events. Every file was a bare cron-init stub.

Cross-checking regular-reviews' own git log for the same window showed the identical shape: no commits 09-08 or 09-09, only routine job-hunt/apartment-hunt commits landing in the coding workspace instead. Real work clearly happened — the coding repo's commit history proves it — it just never touched a session file this cycle could see.

What made it effective

Not stopping at "the sessions are empty" and treating that as license to fabricate findings. Cross-referencing a second, independent signal (the regular-reviews and coding repos' own commit timestamps) confirmed the gap was real rather than a decompression bug, and confirmed real work still happened elsewhere in the same window — just routed around the main agent's own session log entirely (delegated straight to coding subagents per the existing cron pattern).

Friction, pain points, surprises

A three-day stretch of "nothing to learn from" is indistinguishable, from inside a single cycle, from "the review missed something." The only way to tell them apart was checking a second data source (git commits in an unrelated repo) that happens to exist for unrelated reasons. Without that, an honest "nothing new" report and a silent blind spot look identical.

Confirms and extends the existing 2026-08-13 finding (openclaw-session-rotation-adjacent: cron output lands in the destination channel's session, not the cron session's own file) — but this time even the destination sessions were stub-empty, meaning the actual work (job-hunt/apartment-hunt sankey updates) is fully delegated to coding subagents whose own trajectories never surface in any main-agent session this cycle scans. The self-improving skill's scanning scope (~/.openclaw/agents/main/sessions/) structurally cannot see that work at all.

Verdict: adequate for its stated job, but the job's scope is narrower than it looks. No corrections were lost — there weren't any human-agent conversations in the window to correct. But a reviewer relying solely on main-session content would have no way to audit the actual autonomous work happening in delegated subagents unless it independently thinks to check sibling repos' commit history.