cache_read=0/128 - Claude Code

Quick fix

No verified fix exists for cache-read drops under concurrent requests. Keep concurrency at 1 for now and follow the original issue.

Symptom

Error outputCopy

并发数大于 1：更容易出现 cache_read=0/128；

English translation

Concurrency greater than 1: cache_read=0/128 is more likely to occur;

Cause

This is an unresolved report involving Claude Code on CC Switch 3.20.1 and Windows. The report says the cache is generally stable with concurrency set to 1; after enabling subagent, side_query, or other concurrent requests, cache_read_tokens may occasionally drop to 0 or 128 and may recover in the next round.

The original post speculated that requests in the same session were not serialized, that the main session and subagent/side_query used cache keys differently, that a dynamic attribution header or Responses conversion changed the prefix, or that concurrent requests reached the upstream before cache writing finished. None of these were verified, and no verifiable fix was provided.

Fix

1. 1

   The original post observed no obvious cache loss when concurrency was 1. Until a verified fix is available, keep concurrency at 1 as a temporary workaround and follow replies on the source issue.

Affected Versions

ToolClaude Code

Version3.20.1

PlatformsWindows

Source Issues

This page is distilled from 1 real issues

• #6951并发超过1后会出现掉缓存的问题

FAQ

(No FAQ yet)

Related problems

Still shows 1M after switching to a non-1M model, Claude CodeRetest on the latest CC Switch; if it still shows 1M and cannot auto-compact, submit the model configuration and actual model ID.1 issues merged/v1/models returns {"models":[]}A user restored 1M with a Python proxy returning model IDs with the [1m] suffix and context_window: 1048576; no reusable code or official fix was provided.1 issues mergedContext compression keeps failing — CodexUpdate CC Switch from 3.17.0 to the latest version and retry. If it still fails, provide system details, reproduction steps, and logs or screenshots.1 issues merged