# Bug Report: Cerebras Provider Streaming Response Format Corruption

[nichm]

Summary

When using Cerebras provider (both provider_type = "cerebras" and provider_type = "openai") with streaming enabled, CCM fails to properly convert OpenAI SSE format to Anthropic SSE format, resulting in malformed Server-Sent Events that Claude Code cannot parse.

Environment

• CCM Version: 0.6.3
• Provider: Cerebras (via direct API and Cloudflare Gateway)
• Client: Claude Code v2.0.50
• OS: macOS (Darwin 25.1.0)

Bug Details

Expected Behavior

When streaming is enabled, CCM should convert Cerebras/OpenAI SSE chunks to Anthropic SSE format:

data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"}}

data: {"type":"message_stop"}

Actual Behavior

CCM outputs malformed SSE with double data: prefixes and unconverted OpenAI format:

data: data: {"id":"chatcmpl-...","choices":[{"delta":{"reasoning":"Hello"},"index":0}],"created":1764019479,"model":"zai-glm-4.6","object":"chat.completion.chunk"}
data: 
data: 

data: data: {"id":"chatcmpl-...","choices":[{"delta":{"reasoning":" world"},"index":0}],"created":1764019479,"model":"zai-glm-4.6","object":"chat.completion.chunk"}
data: 
data: 

Issues:

1. ❌ Double data: prefix: data: data: {...} instead of data: {...}
2. ❌ OpenAI format unchanged: Still using {"choices":[{"delta":...}]} instead of Anthropic format
3. ❌ Extra blank lines: Adds data: \ndata: \n between chunks
4. ✅ Non-streaming works: When stream: false, responses work correctly

Impact

• Severity: Critical
• Affects: All Cerebras users with streaming enabled
• Workaround: Disable streaming (stream: false), but this breaks real-time response experience in Claude Code

Reproduction Steps

Configuration

[[models]]
name = "zai-glm-4.6"

[[models.mappings]]
actual_model = "zai-glm-4.6"
priority = 1
provider = "cerebras"

[[providers]]
name = "cerebras"
provider_type = "cerebras"  # Also fails with "openai"
api_key = "csk-..."
models = []
enabled = true

[router]
default = "zai-glm-4.6"

Test Command

curl -N -X POST "http://127.0.0.1:13456/v1/messages" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: test" \
  -d '{
    "model": "zai-glm-4.6",
    "max_tokens": 100,
    "stream": true,
    "messages": [{"role": "user", "content": "say hello"}]
  }'

Result

data: data: {"id":"chatcmpl-...","choices":[{"delta":{"reasoning":"..."},"index":0}],...}
data: 
data: 

(continues with malformed format)

Expected vs Actual Comparison

Aspect	Expected (Anthropic SSE)	Actual (CCM Output)
SSE Prefix	data: {...}	data: data: {...} ❌
Format	{"type":"content_block_delta","delta":{"type":"text_delta","text":"..."}}	{"choices":[{"delta":{"reasoning":"..."}}]} ❌
Blank Lines	Single \n\n between events	data: \ndata: \n after each event ❌
Completeness	Full message with proper stop event	Stream appears complete but unparseable ❌

Additional Context

What Works

• ✅ Non-streaming mode (stream: false) - Returns complete, properly formatted Anthropic response

What Fails

• ❌ Cerebras with provider_type = "cerebras" - Malformed SSE
• ❌ Cerebras with provider_type = "openai" - Same malformed SSE
• ❌ Both direct Cerebras API and Cloudflare Gateway - Same issue

Root Cause Hypothesis

The streaming conversion logic for Cerebras provider appears to:

1. Not properly strip/replace the OpenAI data: prefix before adding Anthropic data: prefix
2. Not convert OpenAI delta format ({"choices":[{"delta":...}]}) to Anthropic format ({"type":"content_block_delta",...})
3. Add extra data: lines creating malformed SSE

Suggested Fix

The Cerebras streaming handler should:

1. Parse incoming OpenAI SSE chunks (strip data: prefix)
2. Convert OpenAI delta format to Anthropic delta format
3. Re-emit as proper Anthropic SSE (with single data: prefix)

Similar to how other OpenAI-compatible providers work (which stream correctly).

Related

• Cerebras is listed as "OpenAI-compatible" in README
• README shows ✅ streaming support for Cerebras
• This appears to be a format conversion bug, not a Cerebras API issue

Logs

Debug logs available on request. Shows successful connection to Cerebras but malformed SSE output.

[elidickinson]

Oh, oops! I didn't see this issue and then spent like an hour rediscovering it. Should be fixed in #14