Token estimation

Implementation lives in application/internal/cost/token_counter.go and estimator.go. The estimator exists so AgentFlow can show order-of-magnitude token and cost figures before it shells out to agents—useful for comparing against budgets and for logging what a plan roughly implied, without pretending to be a hosted provider's billing API.

How it works

The flow is deterministic: measure character counts for files or text blobs, classify each slice by extension into a content kind (code, markdown, json, default), then divide by the configured token_estimation.*_chars_per_token values and round up.

token_estimation:
  default_chars_per_token: 4.0
  code_chars_per_token: 3.2
  markdown_chars_per_token: 4.2
  json_chars_per_token: 3.6

Cost display

Estimated tokens are multiplied by pricing.models.<model-id> rates to produce an estimated EUR amount—or whatever budgets.default_currency declares. That multiplication is transparent YAML math on your side, not a hidden rate table in the binary.

Commands

agentflow estimate billing-v2 --task task-003
agentflow work "develop billing-v2" --estimate-only
agentflow cost billing-v2

Golden tests

The estimator's shape is pinned by golden tests in application/internal/cost/. Update goldens with UPDATE_GOLDEN=1 only when you intentionally change the algorithms—otherwise you risk masking regressions in how prompts are sized or costed.

Related

• Budgets
• Context optimization