[0xacee]

Why this matters

Anthropic message deltas can carry cumulative input and cache counters, in addition to output tokens. Token Spy currently retains the initial input/cache values and only updates output, under-reporting a response whose final counters differ.

Normalize the present cumulative counters in one helper used by the deployed Messages proxy and provider adapter. Replace previous values rather than adding them; retain prior observations for omitted/null fields and honor explicit zero values. Stream content and event framing are unchanged.

Contract references: Anthropic streaming usage is cumulative and MessageDeltaUsage fields.

Overlap check

Searched open/closed titles and changed files. #2511 changes Claude model pricing; #2519 fixes OpenAI cached-token accounting; #3595 prices routed telemetry. Inspected #3806: it handles nullable usage containers but does not record cumulative input/cache deltas. This changes counter extraction, not rates or nullable-container policy.

Validation

• Authenticated POST /v1/messages regressions on public-beta: 2 fail, 2 controls pass.
• Full Token Spy suite: 26 passed, 1 skipped (PostgreSQL integration unavailable).
• Same scenarios exercise the provider adapter after the HTTP assertions.
• Multiple deltas, revised input/cache totals, explicit zeros, omitted/null individual fields, exact passthrough text, one logged entry, and stop reason are checked.
• Changed-file Ruff with CI rules and diff whitespace checks pass.
• Mock upstream transport; no live provider calls or billing charges.

Review / risk

Review required before merge: independent review of proxy telemetry behavior and the live proxy validation lane. Pricing tables and requests are unchanged.

AI assistance: implemented and validated with Codex.

Combined validation: c08b6695ff14 on public-beta base 9fc9f61. Token Spy: 64 passed, 1 skipped.

Merge guidance: tested Token Spy main.py order is #4561, #4569, #4583, #4634, #4640. The #4583/#4634 poller conflict must retain await asyncio.to_thread for status and reset, include_session_id=True on assessment, session_id=status["_reset_session_id"] on deletion, and cooldown updates only after action == "killed".

Exact PR-head CI remains red: Windows dashboard compact-draft assertion in RemoteProvider. See run 34709252447. #4570 addresses the overwrite; the combined dashboard suite passes all 906 tests. The contributing account lacks repository rerun permission (HTTP 403 on the same workflow).