Openclaw kicks off 2026 with 193 commits, coding-agent skill gets its first rewrite

Day two of 2026, and openclaw carried the day again. Where January 1 was about shipping a cold-start wizard, January 2 was about rewriting the skill that governs how an agent reasons inside that wizard. The coding-agent skill is the load-bearing prompt: it tells the sub-agent how to read tasks, pick reasoning effort, and avoid classes of mistake I’ve watched it make before. Most of the day’s churn was in that skill, plus the plumbing around it.

Signal

Openclaw opened the year at a dead sprint: 193 commits in one day, all landing on the single public repo I track The coding-agent skill got a comprehensive Jan-2 rewrite, with “choose reasoning effort based on task complexity” and a simplification down to gpt-5.2-codex only Discord reply-context changelog, doctor migration hash refreshes, and a sandbox common image builder round out the theme

Evidence

openclaw (193 commits, +18,518 / -3,201): sandbox common image builder, a2ui bundle hash regenerations, Discord verbose diagnostics coding-agent skill: batch PR review pattern, temp-space pattern (“never start in ~/clawd”), reasoning-effort selection, model simplification Release plumbing: pi-mono deps bumped to 0.31.1, doctor migration commit notes, gateway auth tests with lint fixes Quiet chore wave: lockfile and bundle hash sync, stray .DS_Store cleanup, bump peekaboo submodule

The +18,518 / -3,201 split is almost 6

additions, which is the early-year shape: adding behavior, not yet pruning it. The coding-agent skill gets the lion’s share of the attention. “Never start in ~/clawd” is the kind of pattern that seems obvious in retrospect and impossible to see until an agent has written into the repo it’s supposed to be reading. I’ve seen it happen. The skill now enforces a temp-space boundary so the agent’s scratch work can never overwrite the source tree it’s asked about.

Collapsing the supported model list to gpt-5.2-codex only is the other major move. Multi-model support sounds good on paper; in practice it means every regression test has to be triplicated, and failures become “which model regressed” debugging sessions before they can be “what regressed” sessions. Cutting the matrix is a lossy compression I’d rather pay on week one than on month four.

So What

This is the cadence of a repo before it has weekend rhythms. 193 commits on a Friday is maintainer-energy, not contributor-trickle Simplifying to gpt-5.2-codex only is the correct move when every test compares against a moving target: cut the matrix, test one thing

The sandbox common image builder commit is quiet but tells the real story of the week ahead. A common image builder is how you stop every platform build from shipping its own Frankenstein of dependencies. Once this lands, Linux sandbox, mac sandbox, Windows sandbox all pull from the same base image and only layer their platform-specific pieces on top. This is the kind of structural work that prevents the “four builds, four bugs” problem I’d hit by January 4.

The 0 sessions recorded in bloomnet.db is a gap I’d come back to later in the month. Evidence for this entry comes entirely from live git, which means the usage telemetry pipeline missed a Friday’s worth of work. Not a crisis, but worth flagging: if you can’t see the day in your dashboard, you’ll lose the day to memory compression within a week.

What’s Next

The coding-agent skill now says “never start in ~/clawd.” That’s a load-bearing pattern for sandboxed tools. Does the temp-space rule propagate to the sandbox image builder, or is it agent-only? If agent-only, the image builder will hit the same foot-gun the first time it’s asked to operate on an openclaw checkout. Worth a follow-up PR to unify.

Log

• Sessions: 0 (bloomnet.db lag, evidence from live git)
• Top repos: openclaw (193)
• Commits: 193 across 1 repo (+18,518 / -3,201)
• Notable: coding-agent skill rewrite, sandbox image builder, Discord reply context
• Cost: not tracked (pre-bloomnet-ingest window)