Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Qwen3:14b fails with CLI transcript compaction failed: Already compacted; identical MCP tool succeeds with Llama3.1:8b

Environment

• OpenClaw: 2026.6.10 (aa69b12)
• Ubuntu Linux
• Ollama (local)
• MCP transport: stdio
• Python MCP server
• No cloud/OpenAI providers configured

Summary

I have isolated what appears to be a model-specific issue in the OpenClaw agent runtime.

The same OpenClaw agent, same MCP server, same tool, same prompt, and same environment produce different results depending only on the selected Ollama model.

• ✅ ollama/llama3.1:8b works end-to-end.
• ❌ ollama/qwen3:14b fails before MCP tool execution with:

CLI transcript compaction failed for ollama/qwen3:14b:
Already compacted

This appears to occur inside the transcript compaction subsystem before the MCP tool is invoked.

Verified Working Components

The following components have been verified independently:

• Jarvis CLI command entrypoint
• Python MCP server
• MCP stdio transport
• OpenClaw MCP registration
• MCP tool discovery
• MCP probe

The MCP server is registered as:

openclaw mcp add jarvis \
  --command /home/phil-koserowski/projects/jarvis-briefings/.venv/bin/python \
  --arg -m \
  --arg interfaces.mcp.server \
  --cwd /home/phil-koserowski/projects/jarvis-briefings

Verification:

openclaw mcp probe jarvis --json

Output:

{
  "servers": {
    "jarvis": {
      "tools": 1,
      "diagnostics": []
    }
  },
  "tools": [
    "jarvis__expand"
  ]
}

The MCP layer appears to be functioning correctly.

Reproduction

Agent prompt:

Use the jarvis__expand tool with id 1 and return the result.

Case 1

Model:

ollama/qwen3:14b

Result:

GatewayClientRequestError:

CLI transcript compaction failed for ollama/qwen3:14b:

Already compacted

Gateway log immediately before failure:

embedded_run_compaction_start
reason: "threshold"

embedded_run_compaction_end
completed: true

late compaction count reconcile failed:
TypeError: reconcile is not a function

The MCP tool is never executed.

Case 2

Model:

ollama/llama3.1:8b

No other changes.

Result:

• Agent successfully invokes jarvis__expand
• MCP server executes normally
• Tool output is returned to the user

The gateway still logs:

late compaction count reconcile failed:
TypeError: reconcile is not a function

However, the request completes successfully instead of aborting.

Expected Behavior

Both models should successfully invoke the MCP tool.

A fresh one-turn session should not terminate with:

Already compacted

before tool execution.

Actual Behavior

Only ollama/qwen3:14b aborts the agent turn during transcript compaction.

ollama/llama3.1:8b successfully completes the identical workflow.

Why I Believe This Is an Agent Runtime Issue

The only variable changed between successful and failing runs was the configured model.

Everything else remained identical:

• OpenClaw version
• Gateway
• Agent
• MCP server
• Tool
• Prompt
• Session creation
• Local environment

Because the MCP tool executes successfully with one model but never executes with the other, the issue appears to occur within the OpenClaw agent runtime before MCP tool invocation.

Questions

1. Is this a known compatibility issue with ollama/qwen3:14b?
2. Is the late compaction count reconcile failed: TypeError: reconcile is not a function message related to the failure?
3. Is there additional logging that would help identify why transcript compaction aborts only for this model?
4. Is there a recommended configuration change or fix that preserves local Ollama usage while avoiding this failure?

I would be happy to provide additional logs or test patches if helpful.

Steps to reproduce

Qwen3:14b fails with CLI transcript compaction failed: Already compacted; identical MCP tool succeeds with Llama3.1:8b

Environment

• OpenClaw: 2026.6.10 (aa69b12)
• Ubuntu Linux
• Ollama (local)
• MCP transport: stdio
• Python MCP server
• No cloud/OpenAI providers configured

Summary

I have isolated what appears to be a model-specific issue in the OpenClaw agent runtime.

The same OpenClaw agent, same MCP server, same tool, same prompt, and same environment produce different results depending only on the selected Ollama model.

• ✅ ollama/llama3.1:8b works end-to-end.
• ❌ ollama/qwen3:14b fails before MCP tool execution with:

CLI transcript compaction failed for ollama/qwen3:14b:
Already compacted

This appears to occur inside the transcript compaction subsystem before the MCP tool is invoked.

Verified Working Components

The following components have been verified independently:

• Jarvis CLI command entrypoint
• Python MCP server
• MCP stdio transport
• OpenClaw MCP registration
• MCP tool discovery
• MCP probe

The MCP server is registered as:

openclaw mcp add jarvis \
  --command /home/phil-koserowski/projects/jarvis-briefings/.venv/bin/python \
  --arg -m \
  --arg interfaces.mcp.server \
  --cwd /home/phil-koserowski/projects/jarvis-briefings

Verification:

openclaw mcp probe jarvis --json

Output:

{
  "servers": {
    "jarvis": {
      "tools": 1,
      "diagnostics": []
    }
  },
  "tools": [
    "jarvis__expand"
  ]
}

The MCP layer appears to be functioning correctly.

Reproduction

Agent prompt:

Use the jarvis__expand tool with id 1 and return the result.

Case 1

Model:

ollama/qwen3:14b

Result:

GatewayClientRequestError:

CLI transcript compaction failed for ollama/qwen3:14b:

Already compacted

Gateway log immediately before failure:

embedded_run_compaction_start
reason: "threshold"

embedded_run_compaction_end
completed: true

late compaction count reconcile failed:
TypeError: reconcile is not a function

The MCP tool is never executed.

Case 2

Model:

ollama/llama3.1:8b

No other changes.

Result:

• Agent successfully invokes jarvis__expand
• MCP server executes normally
• Tool output is returned to the user

The gateway still logs:

late compaction count reconcile failed:
TypeError: reconcile is not a function

However, the request completes successfully instead of aborting.

Expected Behavior

Both models should successfully invoke the MCP tool.

A fresh one-turn session should not terminate with:

Already compacted

before tool execution.

Actual Behavior

Only ollama/qwen3:14b aborts the agent turn during transcript compaction.

ollama/llama3.1:8b successfully completes the identical workflow.

Why I Believe This Is an Agent Runtime Issue

The only variable changed between successful and failing runs was the configured model.

Everything else remained identical:

• OpenClaw version
• Gateway
• Agent
• MCP server
• Tool
• Prompt
• Session creation
• Local environment

Because the MCP tool executes successfully with one model but never executes with the other, the issue appears to occur within the OpenClaw agent runtime before MCP tool invocation.

Questions

1. Is this a known compatibility issue with ollama/qwen3:14b?
2. Is the late compaction count reconcile failed: TypeError: reconcile is not a function message related to the failure?
3. Is there additional logging that would help identify why transcript compaction aborts only for this model?
4. Is there a recommended configuration change or fix that preserves local Ollama usage while avoiding this failure?

I would be happy to provide additional logs or test patches if helpful.

Expected behavior

Qwen3:14b fails with CLI transcript compaction failed: Already compacted; identical MCP tool succeeds with Llama3.1:8b

Environment

• OpenClaw: 2026.6.10 (aa69b12)
• Ubuntu Linux
• Ollama (local)
• MCP transport: stdio
• Python MCP server
• No cloud/OpenAI providers configured

Summary

I have isolated what appears to be a model-specific issue in the OpenClaw agent runtime.

The same OpenClaw agent, same MCP server, same tool, same prompt, and same environment produce different results depending only on the selected Ollama model.

• ✅ ollama/llama3.1:8b works end-to-end.
• ❌ ollama/qwen3:14b fails before MCP tool execution with:

CLI transcript compaction failed for ollama/qwen3:14b:
Already compacted

This appears to occur inside the transcript compaction subsystem before the MCP tool is invoked.

Verified Working Components

The following components have been verified independently:

• Jarvis CLI command entrypoint
• Python MCP server
• MCP stdio transport
• OpenClaw MCP registration
• MCP tool discovery
• MCP probe

The MCP server is registered as:

openclaw mcp add jarvis \
  --command /home/phil-koserowski/projects/jarvis-briefings/.venv/bin/python \
  --arg -m \
  --arg interfaces.mcp.server \
  --cwd /home/phil-koserowski/projects/jarvis-briefings

Verification:

openclaw mcp probe jarvis --json

Output:

{
  "servers": {
    "jarvis": {
      "tools": 1,
      "diagnostics": []
    }
  },
  "tools": [
    "jarvis__expand"
  ]
}

The MCP layer appears to be functioning correctly.

Reproduction

Agent prompt:

Use the jarvis__expand tool with id 1 and return the result.

Case 1

Model:

ollama/qwen3:14b

Result:

GatewayClientRequestError:

CLI transcript compaction failed for ollama/qwen3:14b:

Already compacted

Gateway log immediately before failure:

embedded_run_compaction_start
reason: "threshold"

embedded_run_compaction_end
completed: true

late compaction count reconcile failed:
TypeError: reconcile is not a function

The MCP tool is never executed.

Case 2

Model:

ollama/llama3.1:8b

No other changes.

Result:

• Agent successfully invokes jarvis__expand
• MCP server executes normally
• Tool output is returned to the user

The gateway still logs:

late compaction count reconcile failed:
TypeError: reconcile is not a function

However, the request completes successfully instead of aborting.

Expected Behavior

Both models should successfully invoke the MCP tool.

A fresh one-turn session should not terminate with:

Already compacted

before tool execution.

Actual Behavior

Only ollama/qwen3:14b aborts the agent turn during transcript compaction.

ollama/llama3.1:8b successfully completes the identical workflow.

Why I Believe This Is an Agent Runtime Issue

The only variable changed between successful and failing runs was the configured model.

Everything else remained identical:

• OpenClaw version
• Gateway
• Agent
• MCP server
• Tool
• Prompt
• Session creation
• Local environment

Because the MCP tool executes successfully with one model but never executes with the other, the issue appears to occur within the OpenClaw agent runtime before MCP tool invocation.

Questions

1. Is this a known compatibility issue with ollama/qwen3:14b?
2. Is the late compaction count reconcile failed: TypeError: reconcile is not a function message related to the failure?
3. Is there additional logging that would help identify why transcript compaction aborts only for this model?
4. Is there a recommended configuration change or fix that preserves local Ollama usage while avoiding this failure?

I would be happy to provide additional logs or test patches if helpful.

Actual behavior

Qwen3:14b fails with CLI transcript compaction failed: Already compacted; identical MCP tool succeeds with Llama3.1:8b

Environment

• OpenClaw: 2026.6.10 (aa69b12)
• Ubuntu Linux
• Ollama (local)
• MCP transport: stdio
• Python MCP server
• No cloud/OpenAI providers configured

Summary

I have isolated what appears to be a model-specific issue in the OpenClaw agent runtime.

The same OpenClaw agent, same MCP server, same tool, same prompt, and same environment produce different results depending only on the selected Ollama model.

• ✅ ollama/llama3.1:8b works end-to-end.
• ❌ ollama/qwen3:14b fails before MCP tool execution with:

CLI transcript compaction failed for ollama/qwen3:14b:
Already compacted

This appears to occur inside the transcript compaction subsystem before the MCP tool is invoked.

Verified Working Components

The following components have been verified independently:

• Jarvis CLI command entrypoint
• Python MCP server
• MCP stdio transport
• OpenClaw MCP registration
• MCP tool discovery
• MCP probe

The MCP server is registered as:

openclaw mcp add jarvis \
  --command /home/phil-koserowski/projects/jarvis-briefings/.venv/bin/python \
  --arg -m \
  --arg interfaces.mcp.server \
  --cwd /home/phil-koserowski/projects/jarvis-briefings

Verification:

openclaw mcp probe jarvis --json

Output:

{
  "servers": {
    "jarvis": {
      "tools": 1,
      "diagnostics": []
    }
  },
  "tools": [
    "jarvis__expand"
  ]
}

The MCP layer appears to be functioning correctly.

Reproduction

Agent prompt:

Use the jarvis__expand tool with id 1 and return the result.

Case 1

Model:

ollama/qwen3:14b

Result:

GatewayClientRequestError:

CLI transcript compaction failed for ollama/qwen3:14b:

Already compacted

Gateway log immediately before failure:

embedded_run_compaction_start
reason: "threshold"

embedded_run_compaction_end
completed: true

late compaction count reconcile failed:
TypeError: reconcile is not a function

The MCP tool is never executed.

Case 2

Model:

ollama/llama3.1:8b

No other changes.

Result:

• Agent successfully invokes jarvis__expand
• MCP server executes normally
• Tool output is returned to the user

The gateway still logs:

late compaction count reconcile failed:
TypeError: reconcile is not a function

However, the request completes successfully instead of aborting.

Expected Behavior

Both models should successfully invoke the MCP tool.

A fresh one-turn session should not terminate with:

Already compacted

before tool execution.

Actual Behavior

Only ollama/qwen3:14b aborts the agent turn during transcript compaction.

ollama/llama3.1:8b successfully completes the identical workflow.

Why I Believe This Is an Agent Runtime Issue

The only variable changed between successful and failing runs was the configured model.

Everything else remained identical:

• OpenClaw version
• Gateway
• Agent
• MCP server
• Tool
• Prompt
• Session creation
• Local environment

Because the MCP tool executes successfully with one model but never executes with the other, the issue appears to occur within the OpenClaw agent runtime before MCP tool invocation.

Questions

1. Is this a known compatibility issue with ollama/qwen3:14b?
2. Is the late compaction count reconcile failed: TypeError: reconcile is not a function message related to the failure?
3. Is there additional logging that would help identify why transcript compaction aborts only for this model?
4. Is there a recommended configuration change or fix that preserves local Ollama usage while avoiding this failure?

I would be happy to provide additional logs or test patches if helpful.

OpenClaw version

2026.6.10 (aa69b12)

Operating system

Ubuntu Linux

Install method

No response

Model

Qwen3:14b

Provider / routing chain

not sure

Additional provider/model setup details

see description

Logs

Screenshots, recordings, and evidence

Impact and severity

No response

Additional information

No response