OpenCode Claude runs fail through TokenRouter and direct Anthropic model ID

[ChathurangiShyalika]

Description

While evaluating opencode_agent on the scenario suite, Claude Opus 4.6 fails before completing scenario 1. I tested both TokenRouter and direct Anthropic model IDs.

Reproduction

TokenRouter Claude

uv run python -m benchmark.scenario_suite_runner \
  --scenario-ids benchmarks/scenario_suite/scenarios.txt \
  --scenario-root /../scenarios_data \
  --agent_name opencode_agent \
  --model-id tokenrouter/anthropic/claude-opus-4.6 \
  --opencode-workspace-root traces/opencode_workspaces \
  --opencode-allow-files \
  --opencode-allow-bash \
  --opencode-thinking

Failure:

APIError: messages: text content blocks must be non-empty
statusCode: 400
url: https://api.tokenrouter.com/v1/chat/completions

The run starts successfully, calls tools, then fails during the next model call after tool use.

Direct Anthropic Claude

  --scenario-ids benchmarks/scenario_suite/scenarios.txt \
  --scenario-root /.../scenarios_data \
  --agent_name opencode_agent \
  --model-id anthropic/claude-opus-4.6 \
  --opencode-workspace-root traces/opencode_workspaces \
  --opencode-allow-files \
  --opencode-allow-bash \
  --opencode-thinking

Failure:

Model not found: anthropic/claude-opus-4.6
Did you mean: claude-opus-4-0, claude-opus-4-1, claude-opus-4-1-20250805?

Expected Behavior

Claude should run through opencode_agent like other TokenRouter models and produce a trajectory/report for scenario evaluation.

Actual Behavior

tokenrouter/anthropic/claude-opus-4.6 fails with Anthropic message validation error.
anthropic/claude-opus-4.6 fails because OpenCode does not recognize that direct model ID.

[DhavalRepo18]

@ChathurangiShyalika, in order to reproduce the issue, can it be used for the open scenarios?

[ChathurangiShyalika]

Yes, the user needs to have below format in scenarios_data folder.

scenario_1
|_groundtruth.txt
|_manifest.json
|_question.txt

[yb98k999]

Based on your feedback, the first error encountered is: "APIError: messages: text content blocks must be non-empty". This occurs because an empty prompt was passed when sending the request, which will trigger this error directly from the model provider. Please refer to the screenshot below for details.

Secondly, regarding the "Model not found" error, I’m not sure if you are using the correct access path and model endpoint. We support both "anthropic/claude-opus-4.6" and "claude-opus-4-6". I’ve attached a sample curl request below for your reference to help you troubleshoot the issue.

curl -N https://api.tokenrouter.com/v1/messages \
  -H "Authorization: Bearer <TOKENROUTER_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-4.6",
    "messages": [
      {
        "role": "user",
        "content": [
            {
              "type": "text",
              "text": "Hello, what kind of model are you?"
            }
        ]
      }
    ]
  }'

curl -N https://api.tokenrouter.com/v1/messages \
  -H "Authorization: Bearer <TOKENROUTER_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-4-6",
    "messages": [
      {
        "role": "user",
        "content": [
            {
              "type": "text",
              "text": "Hello, what kind of model are you?"
            }
        ]
      }
    ]
  }'

Lastly, I recommend using the native /v1/messages protocol with the best compatibility when calling Anthropic models.

@ChathurangiShyalika

[ChathurangiShyalika]

Hi yb98k999, Thanks for the reply.
I actually sent a list of scenario IDs(prompts) along with the tokenrouter/anthropic/claude-opus-4.6. I tested the same with tokenrouter/anthropic/claude-haiku-4.5 and it worked for the first prompt and then it gave below error for second and stopped the execution. So, I'm sure it is not empty prompt.

Method:      opencode_agent
Scenario ID: 1
Run ID:      opencode_agent_1
Trajectories: traces/trajectories/scenario_suite/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5
Command:
uv run opencode-agent --model-id tokenrouter/anthropic/claude-haiku-4.5 --allow-files --allow-bash --thinking --workspace-dir traces/opencode_workspaces/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5/opencode_agent_1 --scenario-id 1 --run-id opencode_agent_1 <question>
================================================================================

────────────────────────────────────────────────────────────
  Answer
────────────────────────────────────────────────────────────

120


--------------------------------------------------------------------------------
Preparing CouchDB for scenario 2
Reset command:
/Users/...IBM/AssetOpsBench/.venv/bin/python3 src/couchdb/init_data.py --reset-only
Load command:
/Users/.../IBM/AssetOpsBench/.venv/bin/python3 src/couchdb/init_data.py 2
--------------------------------------------------------------------------------
INFO Dropped database 'asset' (200).
INFO Dropped database 'workorder' (200).
dropped asset
dropped workorder
INFO Inserted batch 1/3 (500 docs)
INFO Inserted batch 2/3 (500 docs)
INFO Inserted batch 3/3 (76 docs)
INFO Scenario 2: 'workorder' → workorder (1076 docs).
INFO Inserted batch 1/1 (118 docs)
INFO Scenario 2: 'asset' → asset (118 docs).
workorder       workorder       1076
asset   asset   118

================================================================================
Method:      opencode_agent
Scenario ID: 2
Run ID:      opencode_agent_2
Trajectories: traces/trajectories/scenario_suite/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5
Command:
uv run opencode-agent --model-id tokenrouter/anthropic/claude-haiku-4.5 --allow-files --allow-bash --thinking --workspace-dir traces/opencode_workspaces/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5/opencode_agent_2 --scenario-id 2 --run-id opencode_agent_2 <question>
================================================================================
Traceback (most recent call last):
  File "/Users/.../IBM/AssetOpsBench/.venv/bin/opencode-agent", line 10, in <module>
    sys.exit(main())
             ^^^^^^
  File "/Users/.../IBM/AssetOpsBench/src/agent/opencode_agent/cli.py", line 158, in main
    run_sdk_cli("opencode-agent", _build_parser, _run)
  File "/Users/.../IBM/AssetOpsBench/src/agent/_cli_common.py", line 155, in run_sdk_cli
    asyncio.run(run_coro(args))
  File "/opt/homebrew/Cellar/python@3.12/3.12.13_2/Frameworks/Python.framework/Versions/3.12/lib/python3.12/asyncio/runners.py", line 195, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/opt/homebrew/Cellar/python@3.12/3.12.13_2/Frameworks/Python.framework/Versions/3.12/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/Cellar/python@3.12/3.12.13_2/Frameworks/Python.framework/Versions/3.12/lib/python3.12/asyncio/base_events.py", line 691, in run_until_complete
    return future.result()
           ^^^^^^^^^^^^^^^
  File "/Users/.../IBM/AssetOpsBench/src/agent/opencode_agent/cli.py", line 151, in _run
    result = await runner.run(args.question)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/.../IBM/AssetOpsBench/src/agent/opencode_agent/runner.py", line 681, in run
    raise RuntimeError(
RuntimeError: opencode run failed with exit code 1
STDERR:

STDOUT:
))\n    \n    # Exclude if only mentions 'broken' without leak indicators\n    if has_leak:\n        count += 1\n\nprint(count)\n\"\n"},"error":"The bash tool was called with invalid arguments: SchemaError(Missing key\n  at [\"description\"]).\nPlease rewrite the input so it satisfies the expected schema.","time":{"start":1782917027809,"end":1782917027813}},"id":"prt_f1e22cfca0015FktktWFJftBIm","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","messageID":"msg_f1e22ca2b001P1EIgkznvjxidW"}}
{"type":"text","timestamp":1782917027816,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","part":{"id":"prt_f1e22cedd001Uq9eYAR7oxX7Ei","messageID":"msg_f1e22ca2b001P1EIgkznvjxidW","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","type":"text","text":"Let me read the output file:","time":{"start":1782917025501,"end":1782917027814}}}
{"type":"step_finish","timestamp":1782917027870,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","part":{"id":"prt_f1e22d81c001nxAvwvaQ0oUoHI","reason":"tool-calls","snapshot":"8a714fa0d0cd884445fa37c10793d547319e56f1","messageID":"msg_f1e22ca2b001P1EIgkznvjxidW","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","type":"step-finish","tokens":{"total":15541,"input":15245,"output":296,"reasoning":0,"cache":{"write":0,"read":0}},"cost":0}}
{"type":"step_start","timestamp":1782917029366,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","part":{"id":"prt_f1e22ddf0001A07RIvXLu0bvSA","messageID":"msg_f1e22d83c001CU2M3X7O0qksyC","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","snapshot":"8a714fa0d0cd884445fa37c10793d547319e56f1","type":"step-start"}}
{"type":"tool_use","timestamp":1782917030439,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","part":{"type":"tool","tool":"bash","callID":"toolu_01EMDpn7xBR3xtMGzsp3SWfp","state":{"status":"error","input":{"command":"cat /Users/chathurangishyalika/.local/share/opencode/tool-output/tool_f1e22c9e20017MY62GFUZxj6Zy | python3 -c \"\nimport sys, json, re\n\ndata = json.load(sys.stdin)\ncount = 0\n\nfor wo in data.get('work_orders', []):\n    desc = wo.get('description', '').lower()\n    notes = wo.get('notes', '').lower() if wo.get('notes') else ''\n    combined = desc + ' ' + notes\n    \n    # Check for leaking, dripping, seeping\n    leak_patterns = r'\\b(leak|drip|seep|weep|ooze)\\w*\\b'\n    has_leak = bool(re.search(leak_patterns, combined))\n    \n    # Exclude if only mentions 'broken' without leak indicators\n    if has_leak:\n        count += 1\n\nprint(count)\n\""},"error":"The bash tool was called with invalid arguments: SchemaError(Missing key\n  at [\"description\"]).\nPlease rewrite the input so it satisfies the expected schema.","time":{"start":1782917030436,"end":1782917030437}},"id":"prt_f1e22ddf3001M07vJ8tTuVLKZp","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","messageID":"msg_f1e22d83c001CU2M3X7O0qksyC"}}
{"type":"step_finish","timestamp":1782917030466,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","part":{"id":"prt_f1e22e240001j0Rx0PyOdECShc","reason":"tool-calls","snapshot":"8a714fa0d0cd884445fa37c10793d547319e56f1","messageID":"msg_f1e22d83c001CU2M3X7O0qksyC","sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","type":"step-finish","tokens":{"total":15875,"input":15588,"output":287,"reasoning":0,"cache":{"write":0,"read":0}},"cost":0}}
{"type":"error","timestamp":1782917031198,"sessionID":"ses_0e1dd50deffeI3fEnYUXs0I8v0","error":{"name":"APIError","data":{"message":"messages: text content blocks must be non-empty","statusCode":400,"isRetryable":false,"responseHeaders":{"cache-control":"no-cache","connection":"keep-alive","content-type":"text/event-stream","date":"Wed, 01 Jul 2026 14:43:51 GMT","server":"openresty","transfer-encoding":"chunked","x-tokenrouter-request-id":"202607011443519065927HQDr9ETL","x-tokenrouter-version":"v0.0.0"},"responseBody":"{\"error\":{\"message\":\"messages: text content blocks must be non-empty\",\"type\":\"invalid_request_error\",\"param\":\"\",\"code\":null},\"id\":14605,\"org_id\":\"\",\"role\":1}","metadata":{"url":"https://api.tokenrouter.com/v1/chat/completions"}}}}

error: scenario 2 failed for method opencode_agent: Command '['uv', 'run', 'opencode-agent', '--model-id', 'tokenrouter/anthropic/claude-haiku-4.5', '--allow-files', '--allow-bash', '--thinking', '--workspace-dir', 'traces/opencode_workspaces/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5/opencode_agent_2', '--scenario-id', '2', '--run-id', 'opencode_agent_2', 'In a maintenance review of equipment work orders, reliability engineers are identifying fluid-related issues and separating them from general mechanical failures. Please count how many jobs describe active leaking or weeping (e.g., leaking, dripping, seeping), and exclude cases that only mention broken. Return only the final count as a single integer. Output nothing else.']' returned non-zero exit status 1.
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/Users/chathurangishyalika/IBM/AssetOpsBench/src/benchmark/scenario_suite_runner.py", line 646, in <module>
    main()
  File "/Users/chathurangishyalika/IBM/AssetOpsBench/src/benchmark/scenario_suite_runner.py", line 621, in main
    run_agent_for_scenario(
  File "/Users/chathurangishyalika/IBM/AssetOpsBench/src/benchmark/scenario_suite_runner.py", line 215, in run_agent_for_scenario
    subprocess.run(cmd, check=True, env=env)
  File "/opt/homebrew/Cellar/python@3.12/3.12.13_2/Frameworks/Python.framework/Versions/3.12/lib/python3.12/subprocess.py", line 571, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['uv', 'run', 'opencode-agent', '--model-id', 'tokenrouter/anthropic/claude-haiku-4.5', '--allow-files', '--allow-bash', '--thinking', '--workspace-dir', 'traces/opencode_workspaces/opencode_agent/tokenrouter-anthropic-claude-haiku-4.5/opencode_agent_2', '--scenario-id', '2', '--run-id', 'opencode_agent_2', 'In a maintenance review of equipment work orders, reliability engineers are identifying fluid-related issues and separating them from general mechanical failures. Please count how many jobs describe active leaking or weeping (e.g., leaking, dripping, seeping), and exclude cases that only mention broken. Return only the final count as a single integer. Output nothing else.']' returned non-zero exit status 1.

[yb98k999]

OpenCode implements two separate message branch handling logic for native Anthropic models and custom third-party Claude gateways. By default, third-party Claude model protocol configurations are restricted to the OpenAI-compatible format. When OpenCode processes third-party Claude requests, it automatically strips the top-level system field and sends an empty messages[] array to the model. This results in an empty prompt being submitted, triggering the error "text content blocks must be non-empty".

There are two primary solutions:

1. Build a dedicated custom protocol to process requests from OpenCode before forwarding them to the target third-party API.
2. Directly update the OpenCode channel configuration, switching the npm protocol from @ai-sdk/openai-compatible to @ai-sdk/anthropic.