What SprintBias does.

A board that lives in your repo

Every idea, task, bug, and plan is a markdown file in git. The folder a task sits in is its status. The agent reads the board like any other file, and you see each change in git status.

docs/tasks/{backlog,next,doing,blocked,review,done}/

Six folders are the whole board. Move a file to change its status.

./sprint.sh newtask "…"

Add a task with its own ID. Also newidea, newfeature, newbug, and newplan.

./sprint.sh status

See where all the work stands. No AI runs.

./sprint.sh search <keyword>

Find tasks by keyword.

Grounded in your docs before work starts

A senior engineer reads first. SprintBias maps your project once, then checks each task against your own docs. The agent asks questions until the task is clear. Only then does it reach the sprint.

./sprint.sh profile

Map the project: stack, code, commands, and the docs that are the source of truth. Every AI command starts from this map.

./sprint.sh chat 42

Talk one task through until it is clear, then add it to the sprint.

./sprint.sh gate

Check that tasks are ready. Each one gets a Grounding section that names its sources and any conflicts.

./sprint.sh split <path>

Break a large task into smaller ones.

One fixed process, from build to proof

Every task runs the same steps: build, test, fix, and prove it. Each call the agent makes is written down as "Decided: X, because Y", so you can read it and change it. Work lands in review with a proof line. The agent never commits. You do.

./sprint.sh work

Build the next ready task, or one by ID. It lands in review/ with its proof.

./sprint.sh loop

Autopilot: start ready plans and work through the sprint.

./sprint.sh polish --code

Audit the code. Without the flag, it judges finished work in review/.

./sprint.sh promote

Run each task's tests and move it to done/ when they pass. With no tests, you approve the move.

Plans group related work

A plan is one file that lists related task IDs. It starts its tasks together and retires once they are all done.

./sprint.sh newplan "Auth" 12 13

Make a plan from task IDs, a parent task, or the backlog.

./sprint.sh plan think 5

Improve the plan and line up its tasks.

./sprint.sh plan start 5

Check every task in the plan and add them to the sprint in one step.

./sprint.sh plan check 5

A health report. No AI, no moves.

./sprint.sh plan done 5

Retire the plan once every task is done.

A crew that ships a plan for your sign-off

A crew is a team of AI agents built from your project. Its lead takes a plan and runs the tasks in parallel waves. Each member works inside the rules you set. What reaches you is tested work and one approval.

./sprint.sh crew setup

Build the crew from your project, your agent files, and your standards docs. Each choice is recorded as a Decided: line.

./sprint.sh crew lead plan:5

Run a whole plan from one terminal. The lead starts members in waves and routes the work each landing frees up.

./sprint.sh promote plan:5

Your one sign-off. The lead runs the full test suite until it passes, then hands you this command.

./sprint.sh crew

List the members and the tasks each one holds. No AI runs.

• Your policy is written down. Each member's file says what it may touch, so only one member owns each kind of work.
• A member claims a task by moving its file into doing/. A file move is atomic, so two members never take the same task.
• Each edge case the crew finds is added to the project map. Later tasks test the ones their change touches.
• Want to see one first? ./sprint.sh crew --demo plays a crew run. Nothing in your project changes.

The frontier agent you choose

SprintBias drives the agent you already pay for. Claude Code and Grok Build are fully supported. Switch for one run, or pin a model to each kind of work.

./sprint.sh config

Pick Claude Code or Grok Build and a default model.

./sprint.sh -c work · ./sprint.sh -g work

Run one command on Claude Code (-c) or Grok Build (-g) without changing config.

./sprint.sh model set work <model>

Pin a model for one role, such as work or gate.

• Other agent CLIs run through a generic profile that sends the prompt only.
• Budget caps work on Claude Code only. There is no cap unless you set one.

A shared dev server for the crew

Run your crew on one server you already have: one checkout, one branch, one tmux window per member. SprintBias gives you the setup guide and one view of every agent session. You bring the server.

./sprint.sh shell install

Add a prompt and tmux bar that show the machine, folder, git branch, and clock. The machine name turns red over SSH.

./sprint.sh crew <name>

Start each member in its own tmux window.

./sprint.sh agents

List every AI session on this machine and your servers: state, project, branch, model, and last prompt.

./sprint.sh agents label

Name each terminal tab or tmux window after its session.

• Give the server the fewest keys that work. Git uses a deploy key for the one repo, and a person commits and pushes.

Keep the project healthy

Checks that keep task IDs, links, and the project map in good shape.

./sprint.sh validate

Check task IDs and the links between tasks.

./sprint.sh profile check

Flag where the project map has drifted. No AI runs.

./sprint.sh deps

File a task to audit old or unsafe dependencies.

./sprint.sh sync

Push task changes to GitHub, where they show as Issues.

./sprint.sh learn

Watch a demo of each command. Nothing in your project changes.

What SprintBias is not

• Not a SaaS. It needs no account, login, or database.
• Not an AI model. You bring your own Claude Code or Grok Build.
• Not a host. It runs on your laptop or on servers you already control.
• Not hands-off shipping. Agents plan, build, and check. A person approves, commits, and deploys.

Common questions