Skip to content

Stage Budgets

Stage budgets define different operational limits for different phases of work.

Purpose

ORCA Framework should not treat planning, implementation, QA, and unattended background progress as one undifferentiated pool.

Suggested Dimensions

  • wall-clock time
  • step or iteration count
  • retry count
  • optional token or cost estimate when available
  • optional cache-hit or cache-read signals when available
  • risk level

Default Bias By Stage

  • onboarding: moderate read budget, low rewrite budget, end with a compact orientation artifact
  • spec creation: allow enough reasoning to produce a clear contract, but avoid repeated full-context restatement
  • milestone planning: prefer concise milestone and goal artifacts over sprawling plan narration
  • implementation: stricter retry and output growth limits; spend tokens on execution, not status prose
  • QA: allow evidence gathering, but cap repeated re-analysis of the same failure
  • regression follow-up: stay narrow and scenario-specific
  • review: prioritize diffs, findings, and receipts over full transcript replay
  • goal execution: require milestone-sized contracts, explicit retry ceilings, and receipt emission at pause points

Soft Cap Behavior

At soft cap, ORCA Framework should usually:

  • warn in status or receipt artifacts
  • summarize the current state
  • tighten scope
  • downgrade to smaller context slices
  • prefer receipts, shared-state summaries, or targeted excerpts over raw history replay

Hard Cap Behavior

At hard cap, ORCA Framework should usually:

  • stop the current stage
  • emit a receipt or metrics artifact
  • recommend continue, revise, split, or escalate
  • avoid one more broad retry unless a specific new input changed the situation

Background Connection

Background mode should use stage-specific budgets:

  • planning can usually tolerate more read-only iterations
  • implementation should have stricter write and retry limits
  • QA prep should bias toward artifact generation and evidence gathering
  • goal-oriented background runs should stay milestone-sized and stop at contract boundaries
  • long unattended runs should favor cached or stable prompt prefixes instead of reshaping context every cycle

Budget Rule

If a stage budget is exhausted, the run should stop and emit a receipt instead of continuing optimistically.