[Bug] Anthropic reports stale cache-write cost when message_delta updates cache creation

[Adolanium]

Affected area

AI providers and models

What happened?

The Anthropic provider computes the cache-write rate once from message_start and reuses it. When message_delta has different cache-creation numbers (LiteLLM and other proxies do this), cost.cacheWrite and cost.total drift from actual billing. The 1h vs 5m rate gap under cacheRetention: "long" makes that worse.

Steps to reproduce

1. Stream a request through the Anthropic provider with cacheRetention: "long".
2. Have message_start report only 5-minute cache writes, then message_delta report only 1-hour cache writes. This is what some proxies emit.
3. Compare the final cost.cacheWrite to the 1h rate. It still uses the 5m rate.

Expected behavior

Reprice from the latest event.usage.cache_creation breakdown on message_delta when it is present. Keep the message_start rate when it is not.

Prime Agent version

0.7.1 / source as of 2026-08-08

Environment

Reproduced from source. Shows up behind proxies that re-aggregate usage, such as LiteLLM.

Additional context

Original report: #997

Tested patch, closed in the August 15 backlog sweep: #998

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: The Anthropic provider computed the cache-write rate once from message_start and reused it, failing to recompute it when proxies like LiteLLM updated cache-creation numbers in message_delta.
Fix: Added logic to recalculate cacheWriteCost using getAnthropicCacheWriteCost inside the message_delta event loop when cache_creation breakdown is available on the event.
Validation:

• Ran npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/anthropic-delta-cache-write.test.ts (1 test passed)
• Ran npm run check (Checked 965 files. No fixes applied. All checks passed)
  Upstream discussion: [Bug] Anthropic reports stale cache-write cost when message_delta updates cache creation #1769

Diff: awhite0030/prime-agent@main...jules-681529213854070229-72eab51c
Full pull request: awhite0030#12

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.

[Adolanium]

@awhite0030 The diagnosis and the original tested patch are already here. I filed them as #997 and #998 on Aug 8. This discussion restates both. Your fork PR implements that same fix.

A few things that aren't independent rediscovery:

• Same call in the same place: recompute cacheWriteCost with getAnthropicCacheWriteCost inside the message_delta handler, using the same cacheControl.ttl === "1h" ? "1h" : "5m" ternary.
• Same regression case: claude-haiku-4-5, 1000 cache-write tokens, message_start reports 5m then message_delta reports 1h, cacheRetention: "long", expected cost 0.002.
• createSseResponse / createFakeAnthropicClient were lifted out of anthropic-sse-parsing.test.ts, which is where the original test already lives.

Calling this an investigation that produced a candidate fix leaves out that the writeup and the patch were already mine. If this goes anywhere, credit #997 / #998 and @Adolanium.

Please don't do this on other people's reports either. If you pick up a discussion that already names an issue and a closed PR, credit that work. Don't refile it as your diagnosis.