0.57.0 · bb-plugin 2026-09-28

release 0.57.0 on GitHub →

Features

• Optional UI design reference MCP, with per-CLI detection. A worker can now study real shipped sites before writing an interface instead of reconstructing a layout from memory — the archetype library is interaction philosophy and carries no visual reference, which is where generic-looking UI comes from.

Detection is honest about the thing that actually breaks: registration is per agent CLI, and a vendor installer writes only the configs that exist when it runs. A CLI installed afterwards is permanently unregistered, and that is invisible from inside a worker turn. The About panel and the Build/Explore onboarding re-read on every open and name the exact CLIs bb can run that lack it.

The client list maps to bb's own provider registry rather than a hardcoded set, so acp-opencode and acp-cursor resolve to their CLIs, an installed CLI bb does not offer is not reported as a gap, and a bb provider with no known config path surfaces as a declared blind spot. Registration is explicit-click only, matching the doctrine already stated for the sibling host dependency: a plugin install changes the host, not this card. "Registered" claims config presence only, never a working handshake, and the verdict is always re-read from disk rather than taken from an exit code.

Also fixes a real defect in the decision router: the settings slider read "Act at confidence >=" for every point, which is backwards for auto-continue. routeAt floors the confidence that a turn did progress, so raising it makes the veto fire more and the worker spend fewer turns — an operator setting 0.9 to be strict would have got the opposite. Each point now carries its own thresholdLabel in the registry, beside the point it describes. Display only; no stored value or behavior changed.

Bug Fixes

• A boundary question renders as the moment it records. An Interface Contrast boundary carries a kind — reaction or confirmation — and the host read neither, so both rendered identically. A card could ask "what was your first reaction?" while listing the options the agent had already generated: the exact inversion the method exists to prevent, because it replaces the recorded first reaction with the agent's framing.

Reaction boundaries now withhold their options and name the withholding, so nobody wonders where a list went; confirmation boundaries keep theirs. The kind is read from the [Stelow boundary <id>] marker the run already requires, and the framing is decided once in lib/ and sent whole, the convention splitAction already follows.

• A receipt naming a human decision reported no stale artifacts. The list was read off a field that additionalProperties: false makes no receipt able to carry, so it was always empty. It now derives from the route table that owns it, and a receipt smuggling its own list cannot narrow it.

• The runtime validator now covers every field the published schema declares. nextAction was the consequential one: a receipt with no next move passed.

• A running run says where its live progress streams; a paused or finished one points nowhere, rather than at a chat token no card can render.

• A design-reference status read from the ambient PATH. The register-then- report case installed a claude stub into a temp bin dir and expected the verdict to see it, which only happened on a machine that already had that CLI on PATH. Green locally, red in CI. The handlers now take the environment they answer from.

Chores

• Sync stelow 0.71.6-alpha (calionauta/stelow#27), which carries the design-reference skill. The four skills/ files this release introduced now exist upstream, so the next sync can no longer delete them. Going up changed two things deliberately: the reference names the host's dependency panel rather than one host's About panel and the CLIs the host can run rather than one host's list, and its cross-skill link uses the ../../../<skill>/references/... form that resolves in the repo layout, the flat hub layout, and the vendored tree alike.
• Pins and debt ledger left behind by a rebase. The preset-onboarding pin looked for a literal helper call where the code now invokes it through a local name; the behaviour never changed, and the replacement is a stronger pin (binding and call site with exact arguments).

Documentation

• A squash-merged PR is typed by its title, not its commits. The five commits of #157 collapsed into one whose subject is the PR title, and the subject is the only line release-please reads for the change type. This release is 0.57.0 rather than the 0.56.5 first computed, corrected through the workflow's release-as input, and the rule is now written down so it cannot recur.

(#159)