The right model for the plan — and knowing when it's a human

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.

The refusal that looks like caution

We built a lot of machinery to stop expensive models from grabbing cheap work. We shipped almost nothing for the opposite failure — and it turns out to be just as costly, because it fails quietly. A mid or local session would open a plan, see Preferred models: mid, Claude Sonnet 5, Grok 4.5, read the stronger names as a locked door, and skip it. An absent or names-only list read the same way: "reserved for a higher tier." The plan sat in the backlog, you waited, and a model that could have finished it went idle.

The picker code never agreed with that — it already classifies those plans as eligible and claims them. The doctrine the models read was stricter than the harness. So we fixed the doctrine to match the code: only the tiers a plan lists set the floor. Product names are extra good-fit hits, never a raised bar. An absent line or a names-only list you don't match is unknown fit, which is eligible after a one-line note. Difficulty you discover after claiming is a per-step routing decision, not grounds to refuse the claim.

The fit check now has a single, narrow trigger: your documented weak column and orchestration-class work. "A better model exists" is true of nearly every task and is not a fit verdict. And when a session does correctly decline, the refusal is now three lines — a verdict and at most two → next-steps — not a briefing.

Let the script decide

Judging fit by reading headers is exactly the reasoning step models get wrong in both directions, so there's now a script that applies the rules for you:

python scripts/plan_fit.py --model "Qwen3 32B" --tier mid --effort high

take: features/scoped.md — good (Preferred: mid, Claude Sonnet 5, Grok 4.5) → effort high→low (overpaying)
skip: features/arch.md — underqualified (Preferred: reasoner, frontier; you: mid)
skip: bugs/manual-release.md — assigned to alice@corp.com (agents don't complete this)
1 eligible · 2 skipped  [you: Qwen3 32B/mid, effort high]

Identify the worker with --tier, --model, or --endpoint (resolved from endpoints.yaml); add --effort and it also tells you when your reasoning dial is wrong for a plan's tier. It's read-only — it never claims or moves anything, so pair it with plan_select.py --next --claim once you've picked. --json for tooling, --next for a bare path in cron guards, exit 1 when nothing fits you.

The effort advice rests on one tested invariant: effort is a cost dial, not a tier promotion. Cranking a mid worker to high doesn't make reasoner-only plans eligible, and running a reasoner at low doesn't disqualify it from its own work. No --effort value ever changes which plans you can take — only whether you're overpaying for the ones you already fit.

Some plans are a human's to finish

Not all work should be auto-claimed. A release sign-off, a manual QA pass, anything gated on human judgment or access — those belong to a person, and until now the only way to keep agents off them was to leave the plan out of the ready lanes entirely.

A plan can now carry an optional header:

- **Assignee:** alice@corp.com — owns the DB migration

The value may be a name, username, email, or the literal human. Any of them means a person completes the plan, and every claim path — /work, plan_fit.py, work_once.py, and the project-orchestrator MCP — auto-skips it, regardless of model fit. The field defaults to ai: leave it off, or write ai / agent / unassigned, and the plan is agent-eligible exactly as before.

Crucially, "a human completes it" is not "agents can't touch it." An agent may still read an assigned plan and edit its body — add a ## Progress note, answer a question in the spec — and commit that change. Only completing it (moving it to in-progress/ then completed/) is reserved. If an operator really wants an agent to take an assigned plan, work_once.py --allow-assigned forces a named claim.

Draft one the usual way — /draft and the plan template document the field — and set Assignee when a person owns the outcome.