ank-loop

The backlog was planned so this loop would not have to think about architecture. Consume tasks. Never redesign them.

The ank skill is the contract and applies here in full: own worktree, fresh branch from the default branch, ANK_AGENT set, criterion frozen at claim. This file adds the loop policy only.

One pass

ank status                       where you are, what others hold
ank graph                        what is genuinely unblocked
ank find --status open --free    open tasks no live claim overlaps
pick one task                    see below
ank claim <id>
ank log --method loop            record that this policy was loaded, once
ank show <id>                    the body carries the reasoning; read it first
ank context --since              at the top of every turn while the claim is held
work; ank log "<discovery>" as you learn, not when you finish
review your own diff             two axes, below; ank log what it found
ank done                         the close; it measures the tree, you never report it
next pass

Every turn under a claim

Other agents keep working while you do. Once the claim is held, start every turn with ank context --since: the entities whose file changed, and the claims and completions recorded, since your last work on the task. Names by id and never by content, so ank show <id> whatever you need to read whole.

Your last work is the claim's last renewal, so the cursor moves whenever the claim is renewed, and reading --since is one of those renewals: two calls in a row answer differently. Without a claim it is refused, because the cursor is the lease.

It does not replace the full ank context <path> taken before the claim. That one says what binds the perimeter and what is claimable; --since says only what moved, and a constraint never read in full is not one a diff of changes teaches you.

Picking

Three tests, all mandatory: unblocked in the graph, criterion executable as written, scope clear of every live claim. When no open task passes all three, stop and say so. An idle loop is cheaper than two agents rewriting one perimeter.

When a task is not ready

A task is underplanned when its criterion is ambiguous, rests on a false premise, or needs a decision no ADR has made. Do not invent the architecture. Do not soften the criterion.

one clause rests on a false premise    ank log "discrepancy: <assumed> vs <measured>", finish the rest
the whole criterion is wrong           ank release --reason "<why>", move on
a decision is missing                  ank release --reason "<which>", flag it for planning, move on
it names work your scope cannot reach  do the part you can, file the rest, then
                                       ank release --reason "<what is left, and where>"

That last row is not a failure and is the commonest of the four. A criterion that is right about the work and wrong about who can do it is answered by releasing it, not by widening a scope you were not given: say what you measured, what you finished, and which task carries the remainder.

Work you discover is a new task with blocked_by, left for a later pass. Never a detour in this one.

Reviewing your own diff

Before ank done, never after it, read the diff you are about to close on two axes.

against the criterion     every clause answered, and nothing it did not ask for
against the constraints   every ADR that binds the paths the diff touches

The first axis takes the frozen criterion clause by clause and points each one at a hunk. A clause you cannot point at is not done. A hunk no clause asked for is scope you took without being given it, and it comes back out.

The second axis takes the constraints, not your memory of them: run ank context <path> again over the paths the diff actually touches, which is rarely the set you ran it on at the start of the pass, and read the diff against what comes back.

Then ank log the outcome, whichever way it came out: the two axes, what each one found, and nothing found where nothing was. The log is what outlives the session, and a review nobody can read afterwards was not one.

A finding is acted on in the pass that found it. A constraint the diff breaks is fixed before the close; scope the criterion never asked for comes back out of the diff; a clause left unanswered is the release of the table above, and never a smaller criterion.

The review is not a grade and does not replace one. ank done still runs the verifiers and still measures. What the two axes catch is what a green suite cannot: a criterion answered beside the point, and a constraint no test encodes.

Never in the loop

No questions to a human. No new ADRs or specs. No edit to any criterion.

The loop ends when nothing claimable is safe or the human returns. Leave the report: what finished, what was released and why, and what planning owes the backlog.