Skip to content
CC Switch WikiUnofficial Community Wiki

The thinking process is not shown after using the Claude Code proxy feature - cc-switch

Quick fix

Switch back to the Anthropic protocol to restore the thinking display; on openai_chat, change the block conversion in streaming.rs.

Symptom

Section titled “Symptom”
Error output
我使用cc-Switch的代理功能，让claudecode使用openai协议的api来调用模型，但是在使用中无法显示模型的思考过程，使用anthropic协议却可以正常显示
English translation

I use the cc-Switch proxy feature to make claudecode call the model through an openai-protocol api, but the thinking process is not shown in use — with the anthropic protocol it displays normally.

Cause

Section titled “Cause”

Both issues point at `create_anthropic_sse_stream` in `src-tauri/src/proxy/providers/streaming.rs`. An OpenAI-compatible vendor (Qwen, for instance) may send `reasoning_content` and `content` alternately, while Anthropic SSE content blocks are sequential. The current implementation tracks only `current_non_tool_block_type`, so every switch between `reasoning` and `content` emits a `content_block_stop` and a new `content_block_start`.

That produces two variants: in #2009, the thinking content of models like Qwen is not mapped to a displayable `thinking` block, so the user cannot see the thinking process; in #5692, the alternating stream turns into 24 `thinking` blocks and 23 `text` blocks, and Claude Code renders each `text` block as its own bullet, producing a fragmented reply. To be clear: #5692 proposes ignoring late `reasoning` once text has started, and its local normalizer verification shows displayed thinking blocks at `0`, which does not match what #2009 wants — the thinking process displayed normally. Neither issue is closed, and there is no merged fix yet.

Fix

Section titled “Fix”

  1. Temporary check: if the provider also offers the Anthropic protocol, switch back to it and confirm the thinking process displays normally (#2009 reports that protocol works).

  2. If you must use `openai_chat`, open `src-tauri/src/proxy/providers/streaming.rs` and find the block-switching logic in `create_anthropic_sse_stream` that turns `reasoning` into `thinking` and `content` into `text`.

  3. Following the proposed fix in #5692, add a "visible text has started" state flag: before the first text delta arrives, `reasoning` can enter a `thinking` block; once text has started, late `reasoning`/`reasoning_content` no longer closes the current text block and no longer emits a new `content_block_start`.

  4. Add a regression test that constructs a stream alternating `reasoning_content` and `content`, asserting exactly 1 `thinking` block and 1 `text` block, that the text delta concatenation loses nothing, and that reasoning after text starts is no longer emitted.

  5. Restart cc-switch, then run `claude -p --verbose --output-format stream-json --include-partial-messages "What model are you?"` and confirm Claude Code no longer shows multiple fragmented bullets.

Affected Versions

Section titled “Affected Versions”
ToolClaude Code
Version2.1.218（#5692 报告；cc-switch 3.11.1 - 3.18.0 受影响）
PlatformsWindows

Source Issues

Section titled “Source Issues”

This page is distilled from 2 real issues

FAQ

Section titled “FAQ”
Why do some models (km2.5, gpt120b) show the thinking process while Qwen and others do not?
Vendor APIs are not entirely consistent and orchestrate the thinking stream differently. When the way cc-switch recognizes and converts `reasoning`/`reason_context` or `reasoning_content` does not match how Qwen emits it, the thinking process is invisible.
When the thinking process is not shown, does that mean the model did not have thinking/`high` enabled?
Per the comments on #2009, even with nothing displayed the reply can still come from a model with `high` thinking enabled — not seeing it does not necessarily mean the model did not think.
Will the thinking process still be displayed after applying the proposed fix in #5692?
No. The local verification in #5692 shows displayed thinking blocks at `0`, meaning late `reasoning` after text starts is ignored. If your goal is to display the whole thinking process, this direction does not resolve #2009 directly.
Does this change affect tool calls, parallel reads, multi-turn, or subagents?
#5692 used a local normalizer to verify that sequential tools, parallel reads, multi-turn resume, and subagents keep working, but that verification is not merged into cc-switch — run your own regression before merging.
Section titled “Related problems”
api failed deserialize json body intoDepending on the error: roll Claude Code back, upgrade cc-switch, turn the local proxy off, or stop sending unsupported fields.7 issues mergedRequest body format invalid error - Claude CodeUpgrade cc-switch and enable the Rectifier, or add MAX_THINKING_TOKENS=0, for compatibility with third-party API formats.4 issues mergedUnexpected item type in content.Enable content block filtering in config to remove Anthropic internal types unsupported by third-party backends.2 issues merged

This is an unofficial community wiki with no affiliation to the cc-switch authors or the project itself. Its content is compiled from the project's public GitHub issues. This site distributes no software.