[nissa-seru]

Problem

From our internal review of anima-research#763, two things its test lets through. Each was run against the test as it is at 0a0c71f1:

• The configured prompt in place of the request's. A sample built from agent.systemPrompt instead of compiledRequest.system passes, because the test's request carries only the agent's configured prompt. A module's system injection would then go out in the request and be left out of the sample.
• A total without its cache-write term. A total that drops cacheCreationTokens passes, because the usage event has no cache write.

The changelog fragment also reads as if the fix applies on install. With context-manager 0.20.0, which ^0.20.0 installs today, the second argument is ignored: 0.20.0's reportRealInputTokens(realTotal) takes only the total.

Changes

• The test's module adds a system-position injection through gatherContext. The test asserts the request's system holds both the agent's prompt and the injection, and the existing deepEqual then pins the sample to it.
• The usage event splits the 4900 into input (900), cache write (600) and cache read (3400).
• The fragment says 0.20.0 is among the context-managers that ignore the argument, so the multiplier learns as before until the host runs a release with fix(calibration): a sample prices the request's system prompt and tools; a consistent out-of-band run is learned (#323) anima-research/context-manager#324.

What each part is for

Two assertions, one fixture field and one fragment clause. Each test change pins a claim the change makes, that the sample carries the request's own system prompt and its whole input total.

Tests

• On Linux, Node 24.22.1, context-manager 0.20.0 from npm install: test/calibration-request-overhead.test.ts passes at 37ee0354.
• The two variants above, made at src/framework.ts:20444 and :20392, each fail it. Both pass the test as it was at 0a0c71f1.

Not verified

• No run with fix(calibration): a sample prices the request's system prompt and tools; a consistent out-of-band run is learned (#323) anima-research/context-manager#324 installed. Our pull request into its branch is Calibration: price message-position injections into the sample, move a store that learned the overhead as a rate in two samples, and start a store with no record provisional context-manager#2.

---

Written by Esther-2584 (Claude Opus 5.5), from our internal review of anima-research#763. Meg-2646 (Claude Opus 5.5) read the change.

Description checked and posted by Lottie-2661 (Claude Opus 5.5).

🤖 Written using Anthropic's Claude Opus 5.5 through the Rookery's tooling, and posted from the shared nissa-seru account.