v0.3.0 release notes

v0.3.0 is the skill-library keystone. It changes what a producer is: no longer a holder of a frozen skill list, but a model endpoint that checks out whatever a task needs from a shared library at run-time. The capability gap that used to stop a task cold — “no producer holds the required skill” — is dissolved.

This is the producer-collapse thesis reaching its conclusion. A producer was never the skills themselves; it was always the model executing them. The library just removes the artificial freeze.

A producer is now a model endpoint

You no longer assign skills to producers. In the setup wizard you set up your API keys, then add up to eight producers by assigning an LLM and confirming what it’s good at — a quick capability tag, smart-defaulted for known models — and that’s it. There are no fixed roles and no skill lists to curate.

Skills are composed per task from a shared library, not held per producer. So any producer can run any task: when a task needs a capability the producer doesn’t already have loaded, it checks it out of the library instead of blocking.

This is the first working brick of the skill library: a producer can discover skills it doesn’t hold (search_skills), check one out to read its guidance (load_skill), and drop it to reclaim context (drop_skill), drawn from a shared pool. A cheap resident index (names + one-liners, no bodies) makes discovery nearly free. The full lazy checkout/drop library and the conversational skill-creation flow are specced and land in later bricks.

Dispatch routes on capability + availability — and never blocks

With skills out of the routing path, dispatch picks a producer on two axes:

Availability first. The least-loaded producer wins, so a wave of independent tasks spreads across idle models instead of serializing onto one. You don’t pay for eight producers and watch seven sit idle.
Capability floor — soft. Producers whose model meets a task’s capability floor are preferred; but if none do, the task runs on the best-available model anyway, and the shortfall ships as a Product Quality Report reservation — never a block.

This is the honesty thesis applied one layer down. A goal that can’t be perfect still ships with an honest caveat; now a task that can’t find its ideal producer still runs on the best available one, with the caveat in the PQR. The only hard gap left (ROSTER_GAP) fires when no producer exists at all — a setup error the wizard prevents, not a per-task gap. A skill the plan references that isn’t in the library yet is advisory, never a blocking ticket.

Capabilities now come from the model, not from assigned skills: Modulatio infers a model’s tier, cost class, and strengths from its id for known families (overridable per-model in the wizard), and routing reasons over those.

A fix the live runs flushed out: self-contained goals

Validating the keystone on a live run surfaced a real planning bug. The goal-decompose step could write a goal that referred to its subject symbolically — “covers all three requested topics” — without ever naming them. Because each task runs in a fresh producer that sees only its own task text (not the objective, not sibling goals), that dangling reference was unbuildable; the producer correctly refused to fabricate and shipped a gap-note.

Two fixes ride along: decomposition now names the concrete subject (never “the three topics” / “the above”), and the project objective is threaded into the producer prompt as a north-star so a producer can resolve anything its task left implicit instead of stalling. Re-run with a stronger Leader on the same shape, the goals named their parts concretely and the run fetched dozens of real sources into a substantive deliverable.

What it means for you

Setup gets simpler: pick your models, tag what each is good at, done — no skill bookkeeping. Runs get more robust: a missing capability is a checkout or a best-available run with an honest note, never a dead end, and independent work spreads across the models you’re paying for. Configurations from older versions keep working unchanged — a roster that still lists skills per agent parses and routes fine; those skills are now advisory.

For the full engine contract and the honest ceilings, the Beta calibration page remains the source of truth. The CHANGELOG has the precise commit-level delta.