BLOG
What's new in 2.6: one plan, built in slices
Several agents can now build one plan at once and still open one merge request, and in the default chain a reviewer checks every gate before it asks you.
Kraft 2.6 is out. Several agents can now build one large plan at once and still open a single merge request, and a reviewer agent checks every gate in the default chain before you see it.
Build a plan in slices
Until 2.6, one implementer built each work item. A plan with three parts that never touch each other still went through one implementer, one worktree and one review.
Now a plan can name the parts that stand alone. End it with a ## Slices section, one bullet per slice:
## Slices
- **Slice 1: Parser.** Tasks 1-3.
- **Slice 2: CLI verb.** Tasks 4, 6.
Task 5, in no slice, is shared groundwork.
When you approve that plan, Kraft adds the split to the chain revision and asks you to approve that too. Then:
- Shared groundwork comes first. Kraft builds, tests and reviews every task in no slice in the item's own worktree, so each slice starts from tested code.
- Each slice becomes a work item of its own, with its own worktree, branch, implementer and review, and the slices run in parallel. You can steer, pause, retry or drop one by its ID; a slice that fails or needs you doesn't stop the others.
- The slices join. When every slice has ended, Kraft replays each completed slice's commits onto the parent's branch in slice order. Verification, review and the merge request then run once, on the combined code.
The item page draws the whole group: the groundwork, a box per slice that opens its item, and the join.
If two slices conflict at the join, an agent tries to resolve the conflict. Kraft carries on only once every commit of the slice is on the branch. Otherwise it puts the branch back and stops the item, with the exact steps for you.
The item's dollar and token caps cover it and all its slices together, so slicing doesn't raise the item's cap. A cap stops the next task from starting, so slices already running can each finish their current one.
When not to slice. Each slice pays for its own worktree, setup, implementer and review. That's worth it only when each part stands on its own, with its own tests. For small independent tasks, one implementer with sub-agents has far less overhead. A plan without a ## Slices section runs as one item, as before. Split work into slices has the rules.
A reviewer at every gate
In the default chain, a reviewer agent now checks each gate (spec, plan, chain revision, local review and final review) before you're asked. It reads the document the gate is about, sends work that isn't ready back with its reasons, and hands you the rest marked ready for you, with a note. It never approves; the gate then waits for you.
The gate's page also keeps its history: every round, how it ended, who decided and the note, on the desktop and the phone.
Each check is a paid agent session, at every gate an item reaches. Turn it off for one item with kraft item create --no-auto-gate, or for one gate with an override; Change or turn off the gate reviewer shows how.
See what each sub-agent is doing
When a Claude Code task starts sub-agents, they used to run out of sight. Now the item, the chain graph and the phone list each one with what it's doing, its tokens and its tool uses, and the log can show one sub-agent's lines alone. A sub-agent says what each field means.
Also in 2.6
- Kraft's own chains ship with Kraft. As the built-in add-on
kraft-default, they update when Kraft does, andkraft admin updateshows what changes in them before it installs. See the built-in kraft-default. - Re-run CI that failed outside the code. When jobs fail on a download timeout or a lost runner, a fix-loop repair can re-run them with
kraft item rerun-ciinstead of the item getting stuck. - Heavy test suites take turns. A test scope's
concurrency_groupruns its suite one at a time, across work items and repos, on your machine and on each remote one. - Easier review. The review page counts unresolved threads, and
]and[step through them. - Blue means running, amber means you. The web UI and the console draw running work in blue and keep amber for what needs you.
- The console and VS Code reach a remote Kraft after a password sign-in. See the console from another machine.
uvx kraft-sdlcstarts Kraft, and Kraft runs on Python 3.15, without vector search for now.
Before you upgrade
Check these if you have scripts or chains of your own:
- A
kraftcommand run from a Claude Code session's shell is now recorded as an assistant's, not a person's. SetKRAFT_CLIENT=clifor a command you mean to make as yourself. kraft view listprints the status the board shows (running,needs you,failed). A script that reads the STATUS column should read--json'sstatusinstead.- A home from before 2.6 keeps its copied chains, which don't follow
kraft-default's updates. Move your edits onto an extending chain to get them. - A chain that extends
defaultand gives one of its gates anauto_reviewwith anid,kindorextendsis now refused. Delete those keys and keep anyprompt,skillorharnessit sets.
Upgrade with kraft admin update. Upgrade Kraft has the steps and how to roll back, and the 2.6.0 release notes list every change and fix.