[BUG] CLI crashes with "CLI output was not valid JSON" when Bedrock /count-tokens returns 403
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
When using Claude Code with CLAUDE_CODE_USE_BEDROCK=1 (both standalone CLI and via the Agent SDK @anthropic-ai/claude-agent-sdk@0.2.32), the CLI process crashes when processing large context (e.g., large HTML pages as tool results).
### Root cause
The CLI internally calls /model/{modelId}/count-tokens on Bedrock for context window management. This endpoint is not supported on Bedrock and returns a 403. The CLI does not handle this failure gracefully — it crashes mid-stdout-write, producing a truncated JSON line.
When used via the Agent SDK, the SDK parent process reads this truncated line and throws:
Error: CLI output was not valid JSON. This may indicate an error during startup.
Output: {"type":"user","message":{"role":"user","content":[{"tool_use_id":"toolu_bdrk_01ANFzhQWMo5XE2V5QmKCaqH","type":"tool_result","content":" 200→ \n...
When used as a standalone CLI, the session crashes with no recovery.
### Two sub-issues
- Wrong model ID format for Bedrock: The CLI calls
/model/anthropic.claude-opus-4-6-v1/count-tokens— using the Anthropic-native model ID (no region prefix, no version suffix) instead of the Bedrock format (e.g.,us.anthropic.claude-opus-4-6-v1:0). This is inconsistent with how
/invoke and /invoke-with-response-stream paths are constructed.
- Unhandled 403 crashes the process: When count-tokens returns 403, the CLI process dies mid-write to stdout. There is no fallback or graceful degradation.
What Should Happen?
The CLI should either:
- Skip count-tokens when running on Bedrock (if the endpoint isn't supported)
- Fall back to a local token estimation when the API call fails
- At minimum, handle the 403 without crashing the process
Error Messages/Logs
"http_response", "host": "bedrock-runtime.us-east-1.amazonaws.com", "path": "/model/anthropic.claude-opus-4-6-v1/count-tokens", "method": "POST", "status_code": "403", "duration_seconds": 0.66, "response_size_bytes": 274
Claude Code agent stopped error=Error: CLI output was not valid JSON. This may indicate an error during startup.
Steps to Reproduce
- Claude Code CLI (tested via
@anthropic-ai/claude-agent-sdk@0.2.32) CLAUDE_CODE_USE_BEDROCK=1with bearer token auth- Send a message that causes the agent to read a large HTML file (100KB+) as a tool result
- The CLI calls count-tokens, gets 403, and crashes
Claude Model
Opus
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
2.1.32 (Claude Code)
Platform
AWS Bedrock
Operating System
macOS
Terminal/Shell
Terminal.app (macOS)
Additional Information
### Workaround
Intercept /model/{modelId}/count-tokens requests at the proxy layer and return a synthetic {"input_tokens": <estimate>} response before the request reaches Bedrock.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗