Changelog.

The harness doubles down.

• Claude Code is the harness — native rails: a Stop-fence, schema-checked verdicts, and the fleet exposed over MCP.
• Intent, line by line — agents file the reason for every changed line as they write it; the viewer shows intents in the gutter, and conformance is checked against them.
• Dispatch supervisor — dispatch as a lease: resume-respawn, failure ladders, mid-flight model switch.
• Command bar — a real ⌘K palette inside the app.
• Law theme, at the primitives — mono, square, hairlines, ink; enforced at the component library, not per screen.

STAGED ON MAIN · SHIPS AS THE NEXT SIGNED BUILD

Cloud, Console, and the measured run.

• Loom Cloud — register a mini-server from the app; the composer grew a real Local/Cloud machine picker.
• Loom Console — remote fleet control at loom.mingllm.com/console, proven end-to-end.
• Terminal-Bench 2.1: 82.0 — the same-model run that beats stock Claude Code (78.9) and Terminus (74.6). The receipt lives at /bench.
• Verification certificates — interactive certificates and verdict cards on every finished mission.

The dark-work month.

• 216 commits landed in one pass — missions merge-queue: worktree features through a test-gated queue.
• Chorus went live — order-independent structural merging; six writers, one file, byte-identical in any order.
• Autonomy passes — autoRun missions, goal verification as a follow-up, the brain supervisor switched on.
• MCP connect-anything — research → setup → test → save, for real servers.
• Accounts — Claude OAuth sign-in bridged to real sessions.

The Conductor is born.

• One chat, six Claude Code sessions — the session rail, the conductor, the inspector.
• The plan docket — briefs become task DAGs with owned globs and acceptance checks.
• Any model, routed by cost — the docket picks the cheapest model that clears the bar.