Usage is observed through each provider's official telemetry into one provider-neutral record, and reported, estimated and billed amounts are never merged

213. Usage is observed through each provider's official telemetry into one provider-neutral record, and reported, estimated and billed amounts are never merged

Kontext

The owner ordered automatic token and model tracking (10-12) and then a generic fabric rather than three provider subsystems (10-13). docs/ECONOMICS.md already classes every number, and economics::usage already normalises Anthropic and OpenAI usage for benchmark runs. Telemetry events are runtime volume.

Rozhodnutí

Official seams only (Claude Code OpenTelemetry, Codex app-server notifications, Gemini CLI OpenTelemetry), declared per provider in share/providers.yaml. One usage record, distinct from capability executions, and one normalisation. Events are runtime state, never .ai/repo; .ai holds policy, schema and budgets. Reported tokens, estimates and billed amounts are distinct fields; a missing value is absent, never zero; cumulative snapshots are never summed; provider metadata is observed evidence, never canonical state; no transcript or body is read or kept.

Zamítnuté alternativy

• Transcript parsing
• A MITM proxy or client patch
• Per-provider telemetry subsystems
• One 'cost' field
• Storing events under .ai/repo

Důsledky

execution-telemetry loses its dependency on runtime-adapters; the server gains a loopback OTLP/HTTP-JSON receiver with no new dependency; subscription users see estimates labelled as such.

Sources

Promoted on 2026-10-18 from the owner's ChatGPT project (workspace chatgpt-majordomus, issue I3100). The owner has not accepted it; it is proposed for the owner's decision.

• chatgpt-majordomus:c174#18
• chatgpt-majordomus:c234#0
• chatgpt-majordomus:c236#4
• chatgpt-majordomus:c236#7
• chatgpt-majordomus:c253#2
• chatgpt-majordomus:c253#3
• chatgpt-majordomus:c253#4