Journal

119 openclaw commits: Windows spawn ENOENT fix, agent-personality greetings

voice-generatedtech

Signal

119 openclaw commits, +56,125 / -30,055: biggest commit-count day of the week on openclaw 48 sessions on investor-research, 205 minutes, $48.02: fewer calls than 01-29 or 01-30, higher per-session weight Windows npm-spawn ENOENT resolved (#5815), agent personality now drives /new session greeting (#5706)

Evidence

investor-research: 48 sessions, 205 minutes, $48.02 cost, Opus 4.5 dominant Session count halved from 01-30 (104 to 48), minutes dropped 18%: longer sessions, not shorter ones openclaw: 119 commits, +56,125 / -30,055, net +26,070 lines fix: resolve Windows npm spawn ENOENT (#5815) (thanks @thejhinvirtuoso) and fix(process): resolve npm/pnpm spawn ENOENT on Windows: two commits, contributor PR plus local hardening fix: skip extension append if command already has one: classic Windows-vs-POSIX path-handling bug /new: use agent personality in session greeting (#5706): personality propagates into the first turn instead of a generic hello chore: fix lint, and format after lint to catch reformats triggered by autofixes: two-pass format because autofix kept reformatting what lint had just rewritten

48 sessions at 205 minutes is 4.27 minutes per session, up from yesterday’s 2.4. Sessions are getting longer, which usually means the pipeline is hitting targets that require more context per call rather than more calls per target.

So What

Windows support day. 119 commits, two of which resolve the same ENOENT from different angles, plus the “skip extension append” fix: that is the shape of a multi-root-cause bug where npm spawning on Windows has three failure modes and you have to close all three. The first failure mode is that the command needs .cmd appended; the second is that if it already has .cmd, you must not append it again; the third is that pnpm has the same problem with a different extension. Thejhinvirtuoso caught the npm side, I patched the pnpm extension, and the extension-append check covers the overlap.

The agent-personality-in-greeting fix is small but UX-visible. When a user runs /new, they used to see a generic hello regardless of which agent persona was configured. Now the persona’s own greeting shows up, which means the personality feels consistent from the first turn instead of kicking in later. That is a small change that makes the product feel less like a plumbing layer and more like an identity.

The lint-plus-format commit documents a tooling loop I had to break manually. Lint was rewriting something, format was reformatting what lint wrote, and lint was then flagging the reformatted output. Running format after lint (rather than the conventional order) catches the autofix-reformat collision. Not elegant, but it stops the loop.

The halved session count with only an 18% drop in minutes is the headline I missed on first read. That means the average session is 77% longer than yesterday. Something in the pipeline is asking Opus to think harder per target today; my best guess is the targets that got queued for deep reasoning during yesterday’s Sonnet throughput sweep finally ran.

What’s Next

Windows spawn is hardened for npm and pnpm. Does yarn break here too, or did the PR chain cover all three package managers? Yarn is the obvious next test. If it does not break, the abstraction is sound. If it does, I get a third PR in the chain and a good case for extracting a package-manager-aware spawn wrapper.

The longer-session trend on investor-research is worth tracking: if it keeps climbing, the pipeline is slowly shifting from throughput to reasoning, and the cost profile will follow.

Log

  • Sessions: 48 across 1 projects, 205m total
  • Top projects: investor-research (205m)
  • Commits: 119 across 1 repos (56125 +, 30055 -)
  • Top repo: openclaw
  • Cost: $48.02