draining defaults: auto mode · resume summaries · prompt suggestions · verbose output

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

A traffic audit of one short coding-agent session: hidden support calls, two independent prompt-cache universes, auto-mode overhead, and invisible thinking tokens.

privacyAll identifiers, paths, and account metadata were removed from the underlying capture before analysis. Token counts and timings are retained because they carry the story. Cost figures are list-price equivalents, not an invoice.

API requests

13

for ~5 visible interactions

Request amplification

2.6×

6 of 13 were probes or auxiliary

Total token accounting

544,266

78% of it cache reads

Warm cache hit

98.6%

main loop · classifier 99.7%

Hidden thinking

38.6%

of output tokens on text answers

List-price equivalent

~$1.20

dominated by cache creation

request waterfall

One conversation, several kinds of work

Each bar is one API request; length is model time. Only the blue calls are the conversation you see in the terminal; everything else is orchestration.

> A coding-agent session is an orchestration pipeline, not a one-prompt-to-one-response exchange.

token accounting

Where the tokens went

Per-request token accounting. Cache reads dominate volume, but they are the cheap part. The expensive part is the orange cache creation, led by the two cold starts (turns 3 and 8).

cache continuity

Two cache universes

Claude Code kept two independent prompt-cache chains: the main conversation (~50–53k tokens) and the auto-mode security classifier (~40k). Each warm request read exactly the frontier the previous one created.

auto-mode latency

What auto mode adds to a shell command

For every observed shell action, a separate security model classified the command before it ran. Both harmless settings reads scored severity 2 of 100, and still cost a model round trip.

redacted reasoning

Hidden thinking: paid, redacted, sometimes skipped

Five text-answer turns carried 638 redacted thinking tokens out of 1,653 output tokens (38.6%). The tool-selection turn 11 and the greeting reply used zero; adaptive thinking skips straightforward dispatches.

recommended settings

Settings worth changing before a long run

For a long, supervised agentic session, disable the auxiliary traffic up front. User-wide settings file (~/.claude/settings.json):

{
  "promptSuggestionEnabled": false,
  "awaySummaryEnabled": false,
  "disableAutoMode": "disable",
  "permissions": {
    "defaultMode": "acceptEdits"
  },
  "outputStyle": "Concise"
}

caveatNote disableAutoMode takes the string "disable", not a boolean. Avoid bypassPermissions as a performance optimization: it removes approval checks, not just classifier overhead. Restart or /clear after changing the output style. Keep stronger safeguards for sessions touching production, credentials, or unfamiliar repositories.

epistemics

What the evidence does and does not prove

Strongly supported 4

13 API requests; 6 were probes or auxiliary operations
Auto mode inserted a separate security-model call before each observed shell action
Main loop and classifier used independent cache chains, both with excellent warm continuity
Redacted thinking was a meaningful share of output on text-answer turns

Not yet proven 4

The user's actual billed amount
Exact savings from Concise mode or from disabling suggestions and recaps
Whether disabling auto mode removes all classifier traffic in every configuration
Needs a controlled before/after run, not another observational trace

raw data

The full per-request log

Every number on this page, in one place. Orange cells are the two expensive cache cold starts; blue is the slowest request; tinted thinking cells are turns that paid for redacted reasoning.

> Don't count chat messages. Count the full system of model calls behind them.