Costs, limits, and troubleshooting

Who pays for what

A run spends two things, and Repo Fleet is on neither meter.

• GitHub Actions minutes come off your own GitHub plan, exactly as your tests do. A public repository runs free; a private one draws the plan's included minutes, then pays per minute.
• Model tokens bill the key you added as a secret, priced per million by the model the run uses.
• Repo Fleet runs no model, meters no task, and keeps no copy of your code.

What one run reports it spent

• The result the action writes at the end of a run names its cost in dollars, its turns, and its duration.
• execution_file is the step output holding that execution log, for a later step to read, upload, or comment.
• display_report: true on the step writes the report, cost included, to the job summary, so the run's own page carries the number.
• The Console totals what every run has spent on the key; GitHub's billing page does the same for minutes.

The four caps that bound one run

• --max-budget-usd in claude_args stops Claude Code once the run's API spend reaches that number. It is the only cap counted in the thing Anthropic bills.
• --max-turns caps how many times Claude calls the model, so a run going in circles ends before it reaches the budget.
• timeout-minutes is GitHub's hard stop on the job, and it caps the minutes the job can bill. Left out, GitHub allows a job six hours.
• A concurrency group keeps runs to one at a time, so the worst hour costs one run rather than a queue of them.
• A cap stops a run, never a spend already made. A run that fails still spent what it spent.
• Claude Code in GitHub Actions sets all four in one file, running on a schedule is where they matter most, and timeouts and concurrency covers the two GitHub keys on their own.

Keeping a run cheap

• Narrow the trigger. A label, or a path filter, runs the agent on what you meant rather than on every push.
• Read the first few runs before you widen anything.
• Cap it outside the file too. A spend limit on the Console workspace and a budget on GitHub Actions catch what the file cannot.

What each plan reaches

• Free takes every public repository on every account you link, and one private repository in total.
• Pro takes every repository, public and private.
• A repository past the limit says so where you meet it. Pricing has the rest.

When something goes wrong

What you see	Usually
The workflow never runs	Nothing triggers it, or a schedule was edited on a branch rather than the default one.
The Claude step fails at once	ANTHROPIC_API_KEY is missing, empty, or spelt differently in the file.
The push at the end is refused	The token has no contents: write, or the branch is protected.
No pull request appears	The token has no pull-requests: write.
A run queues and stays there	Another run holds the same concurrency group.
The run stops mid-task	It reached --max-turns, --max-budget-usd, or the job's timeout-minutes.
Runs lists nothing	Nothing has run in the last seven days, or the app cannot see that repository.

Where to look

• The log on github.com is the whole truth: every command, its output, and the step that stopped.
• The editor's findings catch what GitHub would refuse, before a run spends anything.
• Still stuck? Contact us with the repository and the run.