Token Saving Guide

日本語

Use token-saving changes after measuring a baseline. Lower token counts are useful only when they also reduce total retries, loop count, wall time, or provider cost.

Measure First

Enable phase-level usage events:

observability:
  enabled: true
  usage_events_phase: true

Build TAKT and analyze one or more runs:

npm run build
npm run analyze:usage -- .takt/runs/<run>

Compare by step, phase, provider, and model. Focus on repeated review loops, large phase1_execute prompts, and missing usage records before changing workflows.

Reduce Workflow Cost

• Choose lighter builtin workflows such as *-mini when a task does not need the full planning/review loop.
• Use provider_routing to send cheaper or faster providers and models to low-risk steps.
• Route expensive models only to steps that need them, such as final review or difficult implementation steps.
• Split oversized tasks before running TAKT so each workflow has a smaller planning and review context.
• Shorten output contracts when a report does not need detailed prose.

Keep Context Focused

• Keep personas, policies, knowledge, and instructions narrow to the step that uses them.
• Avoid shared-context bloat: do not attach broad knowledge files to every step if only one step needs them.
• Prefer task-specific facts over long copied background sections.
• Remove stale workflow-local facets when they are no longer referenced.

Evaluate External Compression Proxies

External compression proxies should be evaluated only after baseline usage is known. Compare TAKT usage events before and after the proxy, and compare provider-side or proxy-side stats when available. A proxy is a loss if it lowers raw tokens but increases retries, review rejections, or total cost.