How we got here
v0.32.0 made ax's local loop more complete: local Studio, cost routing, team sync, skill hygiene. v0.33.0 widens the loop in two directions. Outward, the community site starts compiling registered profiles, attributing skill authors, and showing adoption. Inward, ax adds more lenses over costly or hidden agent behavior: image context, content-type facets, memory writes, directives, and skill-edit spar runs.
The release is large, but the commit evidence clusters cleanly. It is about making local agent behavior more inspectable, then making the public surfaces that summarize that behavior more credible.
What changed
Community data became a compiled surface (#492, #499, #501, #498, #514, #516). Registered community profiles now compile through a Cloudflare Worker powered by Alchemy, skills can carry author provenance from known collections, trending skills are grouped by canonical identity, and the site has an unlisted `/status` page with adoption stats plus a Cloudflare Web Analytics beacon.
Profile comparison became shareable (#494, #523, #527). The site added challenge share cards, a dedicated head-to-head `/u/<a>/vs/<b>` layout, and a profile reskin using shared recap-deck charts. The leaders page was rebuilt as a unified builder roster (#491).
Heavy context got its own cost lens (#524, #529, #515). Ingest now classifies tool output content types and exposes read-path facets. `ax cost images` surfaces image-heavy context, split by main thread versus subagent, so screenshots that stay in the expensive main context are visible instead of silently rebilled. Cursor and OpenCode file evidence capture filled in the same evidence family.
ax cost images --days=14Dojo can test skill edits (#486). Spar moved beyond task variants to skill edits: plan the baseline, swap in an edited skill, rerun the same task, then score the edited arm against the original.
ax dojo spar-plan --skill release-announcement
ax dojo spar-score <id>The score is intentionally about cost, turns, repairs, episodes, and landed work. It does not pretend to grade writing quality, but it makes skill changes testable instead of purely vibes-based.
Memory, directives, and proposals got more inspectable (#531, #537, #535, #530). `ax memory` surfaces Claude memory writes. `deriveDirectives` turns proactive directives into guidance proposals. The directive-mining spec was rescoped to a v1 MVP, and `ship-checklist` defines what done means for new ax writes.
Studio Desktop moved toward releasable packaging (#526, #528, #519, #520, #511). The desktop app picked up the n/ax namespace icon and tray assets, bundled `electron-updater` to stop launch crashes, staged the onboarding prompt for release builds, and pruned serve-irrelevant dependencies from the bundle by 128 MB.
Why it matters
v0.33.0 is a visibility release. The community can see profiles, leaderboards, skill provenance, and adoption. The local user can see image context cost, memory writes, content facets, directives, and whether a skill edit actually improved a run. Those are different surfaces, but they serve the same goal: make agent work inspectable enough to improve instead of merely accumulate.