Kimi K2.5 on Bedrock: premature end_turn due to Converse API tool call parsing bugs

[gopinaath]

Description

Kimi K2.5 on Amazon Bedrock (amazon-bedrock/moonshotai.kimi-k2.5) makes 2-3 tool calls then issues end_turn, requiring multiple opencode run invocations for tasks that should complete in one session. The same model via OpenAI-compatible endpoints (e.g., opencode/kimi-k2.5-free) chains 14+ tool calls autonomously in a single session.

Root Cause

Bedrock's Converse API has known tool call parsing bugs for Moonshot AI models:

1. Confirmed upstream bug (vercel/ai#11409): Internal model tokens (<|tool_call_begin|>, <|tool_call_end|>, <|tool_call_argument_begin|>) intermittently leak into text output instead of being parsed as tool calls. When this happens, the AI SDK sees a text response + end_turn instead of tool_use, so the agentic loop exits prematurely.

2. Unsupported configuration: Kimi/Moonshot models are not listed in AWS's official Converse API tool calling support table, meaning the translation layer may not be properly maintained.

3. Lossy format translation: Bedrock must translate between Converse API format (tool results as toolResult blocks in user messages) and Kimi's native OpenAI format (role: "tool" messages). This translation is opaque and demonstrably buggy.

Reproduction

{
  "model": "amazon-bedrock/moonshotai.kimi-k2.5",
  "provider": {
    "amazon-bedrock": {
      "options": { "region": "us-east-1" }
    }
  }
}

opencode run "Build a REST API bookmark manager using Flask and SQLite with CRUD endpoints, tag filtering, test script, run tests, and show results."

Expected: Single opencode run completes the full task (8+ tool calls).
Actual: Model stops after 2-3 tool calls. Requires 4 separate opencode run invocations.

Validated Fix

Routing Kimi/Moonshot models through Bedrock's OpenAI-compatible endpoint (bedrock-mantle.<region>.api.aws/v1) instead of the Converse API completely eliminates the premature stopping:

Route	Tool calls per session	Task completed in one run?
Converse API	2-3 (then end_turn)	No (4 invocations needed)
bedrock-mantle (OpenAI-compatible)	8+	Yes
opencode/kimi-k2.5-free (OpenAI-compatible)	14	Yes

Environment

• OpenCode v1.1.65
• AI SDK: @ai-sdk/amazon-bedrock 3.0.79, ai 5.0.133
• Region: us-east-1
• Auth: AWS_BEARER_TOKEN_BEDROCK

[gopinaath]

Please merge this at the earliest. AWS customers would like to use OpenCode with Kimi 2.5 on Amazon Bedrock.

[CashWilliams]

This won't fix, I've run test calls to mantle endpoint and the results are still broken. This is a bedrock/kimi issue.

[gopinaath]

@CashWilliams - thank you for your feedback. I'm running on Ubuntu Linux and this task is completing the testing tasks(Playwright-MCP) I gave it with Bedrock-mantle. Is it possible to share your tests?

[rekram1-node]

Yeah I also tested the change, it does not fix the issue.

[rekram1-node]

Also U can do this in models.dev or your opencode config without any code changes. Ex:

"amazon-bedrock": {
      "models": {
        "kimi-k2.5": {
          "provider": {
            "npm": "@ai-sdk/openai-compatible",
            "api": "https://bedrock-mantle.us-east-1.api.aws/v1",
          },
        },
      },
    },

[rekram1-node]

This is definitely a bedrock bug from what im seeing, I get back valid stop reasons even tho the agent is mid loop visually.

[gopinaath]

@rekram1-node I'll check the models.dev suggestion. I think that's the easiest and elegant solution.

[gopinaath]

Working config requires two pieces:

1. opencode.json in project directory:

  {                                                                                                                                                           
    "provider": {                                                                                                                                             
      "amazon-bedrock": {                                                                                                                                     
        "models": {
          "kimi-k2.5": {
            "id": "moonshotai.kimi-k2.5",
            "provider": {
              "npm": "@ai-sdk/openai-compatible",
              "api": "https://bedrock-mantle.us-east-1.api.aws/v1"
            }
          }
        }
      }
    }
  }

2. ~/.local/share/opencode/auth.json (stores the bearer token):

  {
    "amazon-bedrock": {
      "type": "api",
      "key": "<your AWS_BEARER_TOKEN_BEDROCK value>"
    }
  }

Then run with: opencode run -m "amazon-bedrock/kimi-k2.5"

The bearer token to be stored in auth.json — AWS_BEARER_TOKEN_BEDROCK env var is consumed by the bedrock custom loader but not forwarded as apiKey to @ai-sdk/openai-compatible

[CashWilliams]

@gopinaath The works to configure opencode to talk to the mantle endpoint, but it doesn't fix the bug. I've tested and regardless of how you connect to the kimi k2.5 on bedrock, the model is simply broken.

[taylor-hudson]

@CashWilliams i disagree, once i connected through the configuration above, i have no issues with kimi2.5... WIth the configuration above, there will be two listings for kimik2.5, you must pick the one you configured.

[taylor-hudson]

Careful, if you connect wtih the above configuration, it will charge you thousands. i racked up a 30k bill in 3 days barely using it. I passed in like 13m tokens, idek how it cost so much.