8 posts tagged with "Fleet"

Anchor already keeps executor task specs lean — never the whole project, never a raw file dump, only the context a task actually needs. The fleet registry itself was the exception: the model-fleet MCP's list_fleet handed back every endpoint's base_url and model name whether or not the caller needed it, and there was no other way to get the same picture into a prompt. It's on the same diet as everything else now.

Picking which plan to work on has always been mechanical — priority, value, fit, dependencies, none of it needs judgment. But two real gaps sat under that mechanism: a plan quoting an example header inside a code fence could be misread as its own field, and one corrupt file could take down a whole polling loop. Both are fixed, and daemons get an explicit triage mode that makes the "no model needed" property visible instead of implicit.

An executor's reasoning, false starts, and quoted templates used to ride along with everything else in its reply — all the way back into the coordinator's own context, and from there into the next task's prompt. Only the structured footer does that now; the rest is archived to disk and discarded from context.

Mythos-core rule 6 has always told a model to stop after two failed fix attempts and hand the problem up. That only works if the model is still counting straight — and a resumed or re-invoked task starts a fresh context with no memory of the attempts a prior run already burned. The orchestrator now keeps that count itself.

Anchor already stopped too-hard work with a first-line SUGGEST-ESCALATE. Premium sessions still burned credits on too-easy work with only a soft “say so and ask.”

Anchor’s fit check used to answer one question well: is this model strong enough? Mythos-core rule 11 and model-fitness.md sent weak-column and orchestration-class work up with SUGGEST-ESCALATE. That stopped silent overreach — but it did not catch the other failure mode: a model that is strong enough yet the wrong kind of product for the job.

Anchor's whole job is getting the right model onto the right work. This release tightens that in three ways: lesser models stop refusing work they're good at, a new plan_fit.py decides fit mechanically so no model has to eyeball it, and a plan can now be assigned to a person — which agents quietly leave alone.

Anchor already treats a model’s claim of success as an input to verification, not a substitute. That sentence now has a ledger.