budget-diminishing-continue

SKILL: Budget Diminishing Continue

Domain: tool-orchestration
Trigger: Apply when you have a task-level token budget and must decide between continuing and stopping while respecting diminishing returns. Source Pattern: Distilled from reviewed tool-loop and result-shaping patterns.

Core Method

Track the cumulative tokens consumed per turn, issue a continuation nudge while the usage stays below completionThreshold (90% by default), and stop once token deltas shrink across repeated continuations. Diminishing-return detection keeps the loop from looping forever: once two cells in a row add fewer than DIMINISHING_THRESHOLD tokens, or the tracker reached several continuations, stop and emit the final completion payload. The tracker also guides budgetDecision.action so callers can differentiate between gentle nudges and hard stop signals.

Key Rules

• Initialize BudgetTracker at turn start and update continuationCount, lastDeltaTokens, and lastGlobalTurnTokens after each continuation decision.
• Allow continuations only when under budget * COMPLETION_THRESHOLD and the last token deltas are meaningful; otherwise treat it as diminishing returns.
• When stoppage occurs due to diminishing returns, include the completionEvent that records counts, percentages, and elapsed duration so callers can log or surface why the turn ended.
• If the budget or agent ID is missing, return an immediate stop with completionEvent = null to avoid undefined behavior.
• Always compute the continuation message via getBudgetContinuationMessage so the user sees consistent phrasing and the tracker returns the tidy pct/turnTokens metrics.

Example Application

If a query advances toward a 100k token budget, continue streaming until usage hits 90% or token growth slows; when the stop decision triggers, log the diminishing-returns flag so both user and telemetry know the stream ended intentionally.

Anti-Patterns (What NOT to do)

• Do not keep continuing just because the earlier logic returned continue; re-run the decision each iteration to respect the latest token deltas.
• Do not drop the completionEvent data when stopping; it contains the telemetry the backend uses to tell budgets apart from user cancellations.