# You typed 5 messages. Claude Code made 13 API calls.

> An interactive traffic audit of one short Claude Code (coding agent) session, reconstructed
> from a sanitized 13-request capture. It documents hidden support calls, two independent
> prompt-cache chains, auto-mode security-classifier overhead, and redacted thinking tokens,
> then recommends settings for long agentic runs.

## Key findings

- 13 API requests served ~5 visible user interactions (2.6x request amplification).
- 6 of 13 calls were a quota probe or hidden auxiliary work: session-title generation,
  a next-prompt suggestion, a resume/away recap, and two security-classifier calls.
- Two independent prompt-cache universes: the main conversation (~50-53k cached tokens,
  98.6% warm hit rate) and the auto-mode security classifier (~40k tokens, 99.7%).
- Auto mode inserted one classifier request (~1.7s) before each observed shell command;
  both harmless settings reads scored severity 2 of 100.
- Redacted thinking was 638 of 1,653 output tokens (38.6%) across five text-answer turns.
- Token accounting: 9,892 fresh input, 105,740 cache creation, 426,591 cache read,
  2,043 output; 544,266 total. List-price equivalent roughly $1.20, dominated by cache creation.

## Recommended settings for long supervised agentic sessions

promptSuggestionEnabled: false, awaySummaryEnabled: false, disableAutoMode: "disable"
(string, not boolean), permissions.defaultMode: "acceptEdits", outputStyle: "Concise".

## Caveats

Actual billing, exact per-setting savings, and classifier behavior across all permission
configurations remain unproven; a controlled before/after experiment is proposed.

## Pages

- [Interactive report](https://draining-defaults.sivaramp.com/): charts, per-request log, and copyable config.