dw-next

dw-next — where we are, and the next slice

Everything comes from disk. Never reconstruct state from the conversation: a /clear, a closed laptop or a week away must change nothing about the answer.

What it reads and writes

Reads .ai/work/<date>-<slug>/CHANGE.md (written by dw-shape), a HANDOFF.md beside it when a session left one, any sibling file its ## References names, CONTEXT.md for the project's terms, and the frontmatter of docs/decisions/ for the records binding the files a task edits — the frontmatter, never the folder. Writes code, ticks the checklist, appends to Notes, and commits. Find the active change by branch, never by guessing:

grep -l "^branch: $(git rev-parse --abbrev-ref HEAD)$" .ai/work/*/CHANGE.md 2>/dev/null

One match — that's it. Several — list them and ask. None — there is no change on this branch: point at dw-shape and stop; never invent a task list to have something to do.

Workflow

1. Report, always

A fixed shape, read from the file:

• Capsule — the goal, plus anything in Notes that changes the approach; ≤5 bullets.
• Tasks — one line each, tagged [done] · [skipped: <why>] · [pending]; the first [pending] is the resume point.
• References — each entry read or skipped, one line; read what the next task leans on first.
• Next move — one action.

A HANDOFF.md is read first and leads the report — it holds what a previous session already ruled out. With status, stop here; that is the whole resume path, and it is deliberately cheap.

2. Confirm the task still fits

Check the next task against the repo as it is now; propose an amendment rather than building a stale task. Order is a hint — take a later task when this one is blocked, and say why.

3. Build one task — thin, end to end

• Narrow and complete — a vertical slice, never a whole layer; a second task is cheaper than a sprawling commit.
• Absorb what you find — a discovery inside the goal is built now, as its own commit, or as a new task when it is big. One outside the goal is a one-line Notes item, which dw-land carries into its report — never a .ai/backlog/ file. Never park a gap in this change's ## Goal; shrinking the goal is the user's call.
• Two fixes on one premise means the premise is the suspect — a second fix refused by the same gate as the first, resting on the same assumption, is evidence about the assumption. Write the premise down as one sentence and list every case that gate refused before writing a third fix. The premise stays the suspect until one of those cases refutes it — the gate refused it and the premise did not hold there — and the list is the evidence either way.
• No drive-by edits — outside the task and its absorbed fixes, touch nothing.
• Test the way the project does — failing test first where the task has a real assertion; say so where it genuinely doesn't, instead of fabricating one. A test that also passes against the unfixed code is not a test: docs/decisions/0013 set the bar at confirming each case by mutating the code back to the broken behaviour and watching exactly that case fail.
• Follow the anchors, use the project's words — patterns from the doc, names from CONTEXT.md.
• Nominate, never promote — a call worth a record, or a term worth defining, becomes one line in the doc's ## Decisions in this task's commit: the call or the term, why, (nominated). dw-land puts each to the user, because the durable layer keeps only what the user names, and leg one — hard to reverse — is unanswerable from inside task two of six. A term this change made untrue is rewritten in place now.
• Read the decisions binding the file you are about to edit — where ## Decisions doesn't already carry them, read the records' frontmatter first, never the whole folder: status: active plus a touches: matching the path, then act on that record's rule: verbatim. A record predating those fields is found by its slug, and its norm is the first sentence of its ## Decision — the one case where you open the record itself. A norm you would have to soften to proceed is a reason to stop, not to proceed carefully. No docs/decisions/ is no layer, not an empty answer: say which it was.
• Leave it green — run the tests; lint and typecheck are hook-owned in this lane.