A coding agent can move fast while quietly guessing what you meant. The next session may guess again. Tests can pass while the result is still wrong.
Workspine is a workflow for coding agents. It brings decisions to you and keeps development consistent across agents and sessions.
npx -y workspine setupSetup adds Workspine to the current repository: durable project state in .work/ and workflow skills in
.agents/skills/work-*. Runtime-specific adapters are optional layers on top of that shared skill surface.
For a self-contained task with clear decisions and checks, your agent's normal workflow may be enough. Use Workspine when you want the plan, decisions you approve, verification, and next-session handoff kept with the repository.
Start in your coding agent's chat with one small planned change:
- Ask the agent to use
work-planand describe the outcome and constraints. - Review the checked plan and answer any material questions. Give explicit owner approval in chat when it matches what you want; the agent records the approval against that plan and confirms it. You do not need to create an approval ID or edit a record.
- Run
work-executeto implement the approved plan. - Run
work-verify, then inspect the change, check results, and any remaining gaps.
Verification is evidence-based. If the approved plan makes a UI claim, it can require real browser evidence; passing code tests alone does not establish that rendered behavior.
Use work-quick when the change is already understood and needs less ceremony.
Use work-new-project when the project or milestone itself still needs shaping.
For example, suppose a docs sidebar should reveal a newly selected page while respecting manual collapse
on the current page. You choose that behavior and approve the plan. Workspine keeps that decision, the
implementation summary, and verification evidence in .work/ in your repo. Inspect the diff and
reported checks against the behavior you chose. A fresh session can recover the decision and unfinished
work from those records.
The chat can end. The decisions stay with the work.
Stopping mid-work? Run work-pause and check that it saved the decisions, unfinished work, and next
action. In a fresh session in the same repo, run work-resume to continue or work-progress to see
the current state and next action.
Ask your coding agent to run work-plan. Use its slash command or skill reference when it discovers the installed skills.
If discovery is unavailable, open .agents/skills/work-<workflow>/SKILL.md and follow it directly.
After you approve the plan, run work-execute, then work-verify.
You normally interact through those work-* skills. The generated .work/bin/ helper runtime is for workflow
mechanics and advanced recovery, not a command catalog you need to learn before using Workspine.
Use npx -y workspine health to check the installation and npx -y workspine update to repair
generated files. Both are network-free.
If you already have a legacy .planning/ workspace, setup offers a byte-preserving move to .work/
and explains why. It never migrates silently, and declining leaves the old files untouched.
Install it, try one real change, and report any friction.
An AI development harness for coding agents. Humans normally interact through the work-* skills; the deeper
lifecycle and helper machinery stays in the repository for the agents and for inspection when needed.
- User Guide for every workflow and command.
- Runtime Support for tool-specific installation and invocation.
- Verification Discipline for what the checks do and do not establish.
- Changelog for migration and compatibility notes.
Workspine requires Node >=22. The compatibility command workspine init remains available throughout
the 0.35.x release line. The gsdd binary alias is removed in the next minor release.
MIT. See LICENSE.