Skip to content

Define EIT (Effective Input Tokens) and cross-subtree cost summary #15

Description

@EightRice

Goal

A cross-provider cost scalar so an agent subtree's economic footprint is comparable regardless of which provider each agent uses, plus a summary tool that reports it.

Why

Today, Usage carries input_tokens, output_tokens, cache_read_tokens, cache_creation_tokens per provider. The numbers are not directly comparable: an Opus 4.7 input token is not the same cost as a DeepSeek-Flash input token. A subtree of mixed-provider agents has no single cost figure.

Proposed design

  • Define EIT (Effective Input Tokens) as the unit. Conversion factor per (provider, model, token-type) — fresh-input, output, cache-read, cache-creation each get their own factor based on published pricing.
  • Store factors in atn/model_specs.py (or a sibling file). Same longest-prefix lookup pattern.
  • Add AgentDefinition.budget_eit (optional). When set, runtime totals subtree EIT and refuses turns that would exceed it.
  • New tool get_cost_summary accessible from any agent: returns subtree EIT total, per-provider breakdown, cache-hit rate.

Acceptance

  • A two-agent test (one on Claude Max, one on DeepSeek) reports a single subtree EIT figure that matches manual calculation.
  • An agent with budget_eit set refuses to start a turn that would push it over budget, and the snapshot shows the budget state clearly.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    track:opsCaching, telemetry, cost tracking, observabilitytype:featureNew capability

    Type

    No type

    Projects

    • Status
      Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions