Claude models (and Gemini Argon) fail with HTTP 400 “request body is not valid JSON”; other models work #5139

Activity

1. pedromateo

   added

   triage

   on Oct 14, 2026
2. pedromateo

   added the

   Bug

   issue type on Oct 14, 2026
3. pedromateo

   changed the title [-]Claude models fail with HTTP 400 “request body is not valid JSON”; other models work[/-] [+]Claude models (and Gemini Argon) fail with HTTP 400 “request body is not valid JSON”; other models work[/+] on Oct 14, 2026
4. FloRul commented on Oct 15, 2026

   

   FloRul

   on Oct 15, 2026

   More actions

   Experiencing similar issue with all claude models :
   I’m seeing a GitHub Copilot CLI issue that appears specific to Claude models through the interactive CLI path.

   Environment

   • macOS
   • GitHub Copilot CLI  1.0.98 
   • Also tested cached older CLI versions, especially  1.0.94 
   • Models affected:
   •  claude-sonnet-5.5 
   •  claude-haiku-4.5 
   • GPT models work normally in interactive mode.
   • Direct Copilot API calls to Claude models work.

   Primary symptom

   When using Claude models in Copilot CLI interactive mode, requests often fail with HTTP 400 malformed JSON errors from the Anthropic  /v1/messages  path.

   Examples:

   Execution failed: 400 The request body is not valid JSON:
   decoding JSON body: invalid character 't' after object key:value pair

   Other observed variants:

   invalid escape sequence \e in string
   invalid escape sequence \c in string
   invalid character 's' after object key:value pair
   invalid character '}' after object key
   invalid character ':' after object key:value pair
   invalid character 'a' after object key:value pair

   Second symptom

   After reducing the tool surface, the malformed JSON errors mostly stopped, but Claude interactive sessions began hanging and eventually logging upload timeouts:

   Model provider request failed {"api_endpoint":"/v1/messages","http_status":408}
   Request body upload timed out; retrying the identical request
   request_body_bytes=17033 / 19097

   The CLI keeps retrying, so from the user perspective it looks like Claude is “thinking” forever.

   Troubleshooting performed

   1. Verified model access directly
      • Direct Copilot API calls to Claude worked for:
      •  claude-sonnet-5.5 
      •  claude-opus-5.5 
      •  claude-haiku-4.5 
      • This suggests the issue is not entitlement, authentication, or Claude availability.
   2. Checked Copilot CLI versions
      • Current CLI:  1.0.98 
      • Tested cached versions:
      •  1.0.91 
      •  1.0.92 
      •  1.0.93 
      •  1.0.94 
      • Older  1.0.94  sometimes avoided the malformed JSON error, but still hit  /v1/messages  upload timeout in interactive mode.
   3. Disabled likely interactive contributors
      • Disabled MCP servers
      • Disabled local plugins
      • Disabled hooks:
      copilot config disableAllHooks true
      • Disabled IDE auto-connect:
      copilot config ide.autoConnect false
      • Disabled terminal title/progress:
      copilot config updateTerminalTitle false
      copilot config terminalProgress false
   4. Checked Copilot OTel/logs
      • Used Copilot logs / OTel-style traces.
      • Logs consistently point to Anthropic  /v1/messages :
      Model provider request failed {"stage":"inference","api_endpoint":"/v1/messages","http_status":400}
      • Later:
      http_status=408
      Request body upload timed out
   5. Constrained the tool surface
      • Passing  --available-tools bash  reduced the malformed JSON failures:
      copilot --available-tools bash ...
      • Note: adding  "availableTools": ["bash"]  to  ~/.copilot/settings.json  does not work in  1.0.98 ; logs say:
      Ignoring unknown top-level key(s): "availableTools"
      • So the flag must be passed explicitly.
   6. Added a local shell wrapper
      • Wrapped  copilot  in  ~/.zshrc  to inject:
      --available-tools bash
      • This helped with JSON 400 errors but did not fully fix interactive Claude timeouts.
   7. Compared interactive vs prompt mode
      • Claude prompt mode works quickly:
      copilot -p "Reply exactly OK" --model claude-haiku-4.5 --context default --allow-all
      Result:  OK  in ~2 seconds.
      • Claude interactive mode hangs:
      copilot -i hi --model claude-haiku-4.5 --context default --allow-all
      Result:  /v1/messages  upload timeout/retry loop.

   Current conclusion

   This appears to be a Copilot CLI issue in the Claude interactive request path, likely around request construction/serialization and/or upload handling for Anthropic  /v1/messages .

   Evidence:

   • Claude direct Copilot API calls work.
   • Claude  copilot -p  prompt mode works quickly.
   • GPT interactive mode works.
   • Claude  copilot -i  interactive mode either:
   • sends malformed JSON to  /v1/messages , or
   • hangs on request-body upload timeouts.
   • Reducing tools with  --available-tools bash  helps with malformed JSON but not with interactive upload timeout.

   Current workaround

   For Claude:

   copilot -p "your prompt here" --model claude-haiku-4.5 --context default --allow-all

   or:

   copilot -p "your prompt here" --model claude-sonnet-5.5 --context default --allow-all

   ❤️React with ❤️1Reacted by Sydney Walter
5. copilot-cli-bot

   added

   area:modelsModel selection, availability, switching, rate limits, and model-specific behaviorModel selection, availability, switching, rate limits, and model-specific behavior

   and removed

   triage

   on Oct 18, 2026