Claude Code “Run Until Done” Still Needs a Reviewable Finish¶
Ralph Workflow is a free and open-source AI agent orchestrator built around a simple core loop inspired by the original Ralph loop. That simple core composes into a stronger workflow system for serious repo work, and the default workflow is already strong enough to start with before you customize anything.
If you are searching for Claude Code run until done or the newer /goal-style finish mode, the useful question is not whether the session can keep going longer.
The useful question is this:
When it stops, do you get something you can actually review and decide to merge?
Ralph Workflow is the operating system for autonomous coding: a free and open-source composable loop framework and AI orchestrator that runs the coding agents you already use on your own machine.
It is for developers and technical teams with work that is too big to babysit and too risky to trust blindly.
What makes it different is the finish state: Ralph Workflow hands back a strong software result — a real diff, checks that ran, artifacts you can inspect, and clear open questions — instead of a longer session plus another confident done claim.
Why use it now? Because if Claude Code can now push further on its own, the next bottleneck is not raw autonomy. The bottleneck is whether the morning-after handoff is trustworthy enough to act on.
“Run until done” solves persistence, not the human review¶
A longer-running Claude Code mode can help when you want the agent to:
keep iterating without constant nudges
stay on one task longer before handing control back
reduce stop-and-start friction in an otherwise promising workflow
push a bounded task further while you are away from the keyboard
That is real progress.
But it still does not automatically answer the harder engineering questions:
What changed?
What checks actually ran?
What still needs human judgment?
Would you merge this?
If the run lasts longer but the morning still starts with transcript archaeology, the workflow is not finished just because the session was.
The real problem is trust in the finish state¶
Most developers do not actually want maximum autonomy.
They want a result that is cheap to inspect and boring to review.
When “run until done” still feels risky, the missing pieces are usually:
one bounded task instead of an open-ended session
acceptance criteria before code starts
checks that run during the workflow
a fail-closed handoff when the result is weak
repo-local artifacts that make re-entry easy the next morning
That is the gap Ralph Workflow is built to close.
What Ralph Workflow adds on top of Claude Code¶
Ralph Workflow does not replace Claude Code.
It wraps the agent you already use in a repo-native workflow that makes the finish state easier to trust:
what changed
what checks ran
what was fixed before handoff
what still needs human judgment
whether you would merge it
That is more useful than merely extending the session.
What a trustworthy morning-after handoff looks like¶
If the run went well, the handoff should look roughly like this:
Task: Add empty-project-name validation to the CLI create flow
Changed files:
- cli/create.py
- tests/test_create.py
Checks run:
- unit tests for create flow
- lint / formatting checks if applicable
Open questions:
- should reserved names be rejected too?
- should whitespace be trimmed before validation?
That is the standard to hold unattended work to: proof of completion, not just a stronger done mode.
If you want to inspect that artifact shape first, open the public example review bundle.
When Ralph Workflow is the better fit¶
Ralph Workflow is the stronger path when:
Claude Code is already useful, but you want a cleaner morning-after review path
/goalor “run until done” reduces interruptions but still leaves a fuzzy human reviewthe task is clear enough to hand off overnight
you care more about the result summary than the live transcript
If you need live exploration on a fragile task, stay interactive. If you want a bounded overnight handoff you can judge honestly in the morning, Ralph Workflow is the better fit.
Best next step if “run until done” still leaves a fuzzy morning-after handoff¶
Do not stop at a longer session.
Use Codeberg as the main public home:
Inspect the primary repo on Codeberg: https://codeberg.org/RalphWorkflow/Ralph-Workflow
Star or watch on Codeberg if the finish state earns trust: https://codeberg.org/RalphWorkflow/Ralph-Workflow
Report finish-receipt or first-run friction on Codeberg: https://codeberg.org/RalphWorkflow/Ralph-Workflow/issues/new
Use GitHub only as the mirror: https://github.com/Ralph-Workflow/Ralph-Workflow
Then use After Your First Ralph Workflow Run so a promising trial becomes a visible primary-repo trust signal and a rough one becomes a useful issue.
Fastest honest first run¶
Keep Claude Code already installed and authenticated on your own machine
Pick one real backlog task with clear acceptance criteria
Run Ralph Workflow overnight
Review the diff, checks, and artifacts in the morning
Ask: does the implementation hold up?
If you want the shortest path, start with Getting Started.
If the blocker is still approval babysitting, read claude-code-approval-mode.md.
If the blocker is the broader overnight automation path, read run-claude-code-overnight-without-babysitting.md.
If you want the sharper product comparison, read ralph-workflow-vs-claude-code.md.