haksolot/ank

ank は何をしますか？

Tasks and architecture decisions live as files in .ank/, attached to the code they constrain. Two modes: executing work, and shaping the work that exists. context starts both.

Loop:      ank context -> ank claim <id> -> ank show <id> -> ank log "<msg>" -> ank done
Off-loop:  ank new, ank find, ank release --reason "<r>"
Planning:  ank new adr, ank amend, ank review, ank graph, ank check

The loop

ank context [path] — run it first, always. With no claim it orients you: what binds this perimeter, what is claimable. With a claim held it switches to execution and gives you the criterion and the constraints in full. Constraints are never truncated; if something is cut, it says so.

ank claim <id> — takes the task and freezes its done_criteria by hash. It refuses, with the reason and the next command, when the task is held, blocked, finished on another branch, has no criterion to be measured against, or when you already hold a live claim on another task.

ank show <id> — the entity whole, frontmatter and body, byte for byte. context gives you the criterion and the constraints; the body is where the reasoning lives, and it is the one thing nothing else serves.

ank log "<message>" — what you learned, what you tried, what you rejected. Renews the claim. Log when you discover something, not when you finish.

ank done — runs the declared verifiers itself and records what actually ran, hashed. Never edit status: by hand, and never report your own result: an agent that grades itself can simply be wrong.

Off-loop

ank new task --title "<t>" --scope "<glob>" — a scope is mandatory; an entity attached to nothing is invisible. A subtask you discover is a new task with a blocked_by, never a softened criterion.

ank find <query> — search titles, scopes and criteria. ank find --status open lists what remains, with no query to invent.

ank release --reason "<why>" — stuck, or wrong about the approach. Say so and hand the task back rather than letting the claim lapse in silence.

Planning

Deciding which tasks should exist, in what order, under which constraints. This is what produces the work everyone else loops on, so a badly shaped backlog costs more downstream than any single task does.

ank new adr --title "<t>" --scope "<glob>" --constraint "<rule>" — a decision, not a task. Write one when you hit something the corpus should be held to afterwards, rather than something you merely have to do now. It lands proposed, which binds nobody until it is ratified.

ank amend <id> — the fields a plan actually changes: blocked_by, scope and --criteria. It adds and removes explicitly and never takes a replacement list, so nothing is dropped by being forgotten. It will not touch a done_criteria a live claim has frozen — under a claim, that is release --reason.

ank review — the ratification queue and the health of the corpus: what is proposed and waiting, and which scopes have gone dead and want closing.

ank graph — the blocked_by DAG as text, indented under what blocks it. What is genuinely a root, and what only looked like one in a flat list.

ank check — the mechanical invariants: parse, round-trip, references, frozen fields, orphaned claims. Exit 8 means findings, and findings are for reading, not for silencing.

accept is not yours to run. It is what turns a proposed ADR into a binding one, and it is a human act: signed, on the default branch, with no way around it. Propose the decision, then say it is waiting. Knowing where your authority ends is part of planning well.

Rules that are not negotiable

• The criterion is frozen at claim. Editing it to unblock yourself unblocks nothing: the hash is held where you cannot reach it, and check reports the divergence. If it is wrong, release --reason and say why.

• constraint in an ADR is binding, not advice. Read the ones covering the files you are about to touch — that is what context hands you.

• .ank/ is opaque, like .git/. Never read or write those files directly. ank show <id> gives you an entity whole, ank find lists, ank context binds — the CLI knows the budget, the freeze and who holds what; the files do not.

• One agent, one working tree, one identity. The nominal case is a tree per agent — a clone or a git worktree — each on its own branch. ANK_AGENT names the session and falls back to <user>@<hostname>, so two sessions in one tree are one agent to the refs: they share a claim instead of arbitrating over it, and the second one is refused work it should have been given. Set it per session. Several agents in one tree runs, and is a degraded mode rather than the design.

• What you read is never styled. Colour is emitted only when a human is at a terminal, never into a pipe, a file or --json, so the bytes reaching you are plain: there is nothing to configure and no second surface to prefer.

Exit codes carry meaning: 4 unavailable, 6 a frozen field diverged, 8 findings, 9 environment. Errors always name the exact next command.

Flags beyond these live in ank help, loaded when you need them.

Install

npx skills add haksolot/ank

The skill says how to use ank; it does not install the binary. Releases carry one for Linux, macOS and Windows.