Summary
With ollama provider, both qwen3.5:9b and qwen3.5:35b-a3b frequently produce reasoning/thinking text but fail to execute structured tool calls in ZeroClaw, causing dropped actions and user-facing incomplete replies.
This reproduces as:
Ollama returned empty content with only thinking ... Model may have stopped prematurely.Malformed <tool_call>: expected tool-call object in tag body (JSON/XML/GLM)- occasional provider-side parse errors:
expected element type <function> but have <parameter>
Affected component
provider / agent tool-call parsing (ollama path)
Severity
S1 - workflow blocked
Current behavior
Agent replies with planning text like "I need to use file_open..." but tool action is not actually dispatched/executed.
Expected behavior
If model indicates tool usage intent, ZC should receive/normalize valid tool-call payload and dispatch tool execution, without leaking internal malformed protocol states.
Environment
- ZeroClaw gateway:
127.0.0.1:42617
- Provider:
ollama
- Models tested:
qwen3.5:9b, qwen3.5:35b-a3b
- Date observed: 2026-03-08 through 2026-03-09
Evidence (local logs)
From /opt/homebrew/var/zeroclaw/logs/daemon.stdout.log:
2026-03-08T15:44:28Z qwen3.5:9b only-thinking + malformed <tool_call>2026-03-09T17:47:49Z qwen3.5:35b-a3b only-thinking2026-03-09T17:54:36Z qwen3.5:35b-a3b only-thinking2026-03-09T18:15:30Z ollama 500: expected element type <function> but have <parameter>
Additional validation
Direct calls to Ollama API (outside ZC) with the same task and a simple file_open tool schema returned valid structured tool_calls and done_reason: "stop" for both models. This suggests a ZC integration/parsing mismatch rather than a base model inability.
Reproduction steps
- Configure ZC with provider
ollama and model qwen3.5:35b-a3b (also seen on qwen3.5:9b).
- Ask agent to run a simple tool action, e.g. "open SOUL.md in the default editor".
- Observe intermittent no-op assistant response and daemon warnings listed above.
Notes
Potentially related historical issues (closed): #1109, #1442, #234, #418.
This appears to be a regression/new variant because the same failure pattern is still reproducible on current builds.
Summary
With
ollamaprovider, bothqwen3.5:9bandqwen3.5:35b-a3bfrequently produce reasoning/thinking text but fail to execute structured tool calls in ZeroClaw, causing dropped actions and user-facing incomplete replies.This reproduces as:
Ollama returned empty content with only thinking ... Model may have stopped prematurely.Malformed <tool_call>: expected tool-call object in tag body (JSON/XML/GLM)expected element type <function> but have <parameter>Affected component
provider / agent tool-call parsing (ollama path)
Severity
S1 - workflow blocked
Current behavior
Agent replies with planning text like "I need to use file_open..." but tool action is not actually dispatched/executed.
Expected behavior
If model indicates tool usage intent, ZC should receive/normalize valid tool-call payload and dispatch tool execution, without leaking internal malformed protocol states.
Environment
127.0.0.1:42617ollamaqwen3.5:9b,qwen3.5:35b-a3bEvidence (local logs)
From
/opt/homebrew/var/zeroclaw/logs/daemon.stdout.log:2026-03-08T15:44:28Zqwen3.5:9b only-thinking + malformed<tool_call>2026-03-09T17:47:49Zqwen3.5:35b-a3b only-thinking2026-03-09T17:54:36Zqwen3.5:35b-a3b only-thinking2026-03-09T18:15:30Zollama 500:expected element type <function> but have <parameter>Additional validation
Direct calls to Ollama API (outside ZC) with the same task and a simple
file_opentool schema returned valid structuredtool_callsanddone_reason: "stop"for both models. This suggests a ZC integration/parsing mismatch rather than a base model inability.Reproduction steps
ollamaand modelqwen3.5:35b-a3b(also seen onqwen3.5:9b).Notes
Potentially related historical issues (closed): #1109, #1442, #234, #418.
This appears to be a regression/new variant because the same failure pattern is still reproducible on current builds.