An expensive session hitting /work used to have two options on small/mid Preferred work: skip it and leave the backlog stuck, or quietly do it anyway at full cost. There's now a third: check whether something cheaper can actually take it, and say so explicitly.

Operators can now keep an app on current Anchor, register the CLI, and size lean local models — from the coding agent, without memorizing flag soup.

Multiple agents on one repo no longer have to fight over a single HEAD. Each worker can own a git worktree under var/worktrees/<agent-id>/.

Two refinements to the .plans/ workflow: ready plans can now carry an explicit Priority that orders them ahead of the coarse Value field, and fresh drafts are private by default so a half-baked plan never lands in a commit before you mean it to.

Coding agents can now list, claim, and complete plans for one project through a limited MCP surface — without full shell, silent promote, or unbounded spend.

Mythos-core rule 7 says an executor may only touch the files listed in its task spec. Until now that was a promise the model made to itself. scripts/scope_gate.py makes it a check: any change outside the spec's ## Files in scope is rejected before tests run.

Anchor makes less-powerful models behave like a Mythos-class model — through structure, not capability. Plan on the smartest thing you can afford, cut the plan into task specs that fit one context window each, execute on the cheapest model that passes your benchmarks, verify with tooling, review with a fresh-context critic, and escalate after two failures instead of retrying forever.

Anchor now treats tracked plans, durable fleet pullers, and honest inference economics as first-class operator surfaces — not just doctrine on a shelf.

Cross-model handoff is now a git-tracked tree, not chat archaeology. Ready plans live under .plans/; executors start with /work, honor each plan's Preferred models, and archive only when Done when holds.