Repository navigation
Description
AWF's budget notices warn the agent when a run's AI credits or time are nearly spent. Up to AWF v0.29.11 they are off unless the AWF config turns them on. From the next AWF release they are on by default (github/gh-aw-firewall#11486), and only an explicit false turns them off.
gh-aw already turns notices on for every workflow with a credit budget, and writes enableTokenSteering: false when a workflow sets sandbox.agent.token-steering: false. When a workflow sets max-ai-credits: -1, gh-aw leaves the setting out. With AWF releases after v0.29.11, those workflows will therefore get notices too: time notices against the agent deadline, and credit notices against AWF's hard cap. That's intended, since notices should be on by default. token-steering: false stays the way to turn them off. No code change is needed, but several docs describe the old behaviour.
What we'd like
token-steeringdescriptions. The "Token steering (sandbox.agent.token-steering)" section ofdocs/src/content/docs/reference/sandbox.md, the Token Steering entry indocs/src/content/docs/reference/glossary.md, and thetoken-steeringdescription inpkg/parser/schemas/main_workflow_schema.json(which generatesfrontmatter-full.md) all say that turning the setting off preserves "the explicitly configured provider and model". Steering never changes the provider or model. They should say that the setting controls AWF's budget notices, that the notices are on by default, and thattoken-steering: falseturns them off.- What the notices do. The sandbox.md paragraph describing the notices, and the
max-ai-creditssection ofdocs/src/content/docs/reference/frontmatter.md("budget-warning messages at 80%, 90%, 95%, and 99% of the budget"), need updating for AWF releases after v0.29.11. Each conversation in the run, sub-agents included, gets each notice once. The notice states how much budget or time is left, then says what to do. A further notice comes when about five requests' worth of credits, or about five minutes, remain. max-ai-credits: -1. frontmatter.md says a negative value disables "both budget enforcement and steering". With AWF releases after v0.29.11, it disables only the budget, and notices continue unlesstoken-steering: falseis set. The same applies to the comment inpkg/workflow/awf_config_build.gothat says-1disables token steering.
gh-aw can keep writing enableTokenSteering, which AWF keeps as an alias of the new apiProxy.steering.budget.
Related: github/gh-aw-firewall#11442, github/gh-aw-firewall#11486
Activity
[-]Write `enableTokenSteering: false` when a workflow turns budget notices off, and fix the `token-steering` docs[/-][+]Update the `token-steering` and `max-ai-credits` docs for AWF's on-by-default budget notices[/+]github-actions commentedon Nov 12, 2026
Assigned to
@pelikhan: they most recently changed the max-ai-credits check script (.github/scripts/check-max-ai-credits.cjs), the budget area these docs cover.