[BUG] Unknown API Error: Please contact Roo Code support

[uaktags]

Problem (one or two sentences)

It seems regardless if I'm using RooCloud or OpenRouter and the many models provided from them, I have been getting hit with these "Empty assistance response" errors. Starteed happening about 2days ago and once it starts, it doesn't seem to recover

Context (who is affected and when)

Happens randomly during a session and will use up tokens as it gets stuck in "Empty assistance response" loops

Reproduction steps

1. Windows, VSCode

Not sure how to reproduce?

Expected result

Task should continue on or be able to recover

Actual result

Please contact Roo Code support

Variations tried (optional)

No response

App Version

3.36.7

API Provider (optional)

OpenRouter

Model Used (optional)

No response

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

The issue involves two related errors:

1. "Unknown API Error" - A catch-all error message for unrecognized HTTP status codes (defined in webview-ui/src/i18n/locales/en/chat.json)
2. "Empty assistant response" - Generated in src/core/task/Task.ts when the API returns no content

The root cause appears to be in the retry logic within Task.ts. When autoApprovalEnabled is true and an "Empty assistant response" error occurs, the system enters a retry loop using backoffAndAnnounce(). Without a maximum retry limit for this specific scenario, the loop can continue indefinitely, consuming tokens.

Proposed Solution

1. Add Maximum Retry Limits: Implement a maximum retry count (e.g., 3-5 attempts) specifically for "Empty assistant response" errors in the recursivelyMakeClineRequests() method to prevent infinite loops.

2. Add Circuit Breaker Logic: After reaching the retry limit, stop automatic retries and show a clear message prompting user intervention, even when autoApprovalEnabled is true.

3. Improve Error State Recovery: Ensure that after consecutive failures, the task transitions to a state that allows the user to either retry manually, modify their request, or abort - rather than being stuck in a broken state.

4. Better Error Categorization: Distinguish between transient errors (worth retrying) vs persistent errors (should stop retrying immediately) to avoid wasting tokens on unrecoverable situations.

Files to Modify

• src/core/task/Task.ts - Add retry limits and circuit breaker logic
• Potentially add tests to verify the retry behavior

This approach will:

• Prevent token consumption in infinite loops
• Allow recovery from error states
• Provide clearer guidance to users when errors occur

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[hannesrudolph]

Does it persist if you start a new chat @uaktags ? And if so how soon before you see that message?

[uaktags]

It's hit or miss. Sometimes that does fix it and I'll just continue on from the implementation plan from the last update. But most of the time it doesnt, instead leaving me to just walk away from it for about 15-20 mins. For models, I tend to use GLM 4.6 and GPT-5.1-Codex on open router. Grok-code-fast on RooCloud provider. I was using Roo up until 2-3days ago when this started happening , moved over to Open router thinking it was a provider issue, but then the issue started happening there. In both cases it's "empty assistant response".

…

On Mon, Dec 15, 2025, 3:57 PM Hannes Rudolph ***@***.***> wrote: *hannesrudolph* left a comment (RooCodeInc/Roo-Code#10106) <#10106 (comment)> Does it persist if you start a new chat @uaktags <https://github.com/uaktags> ? And if so how soon before you see that message? — Reply to this email directly, view it on GitHub <#10106 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAG2F5EWSZH44V2OIJKXDST4B4OCHAVCNFSM6AAAAACPDYQMDGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMNJXGU3TINRTGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[manyfranks]

I'm facing a similar issue. Primarily using Openrouter as API provider, with several flagship models. I've also tried connecting to Claude Code as the provider, with the exact same "API Error - Unknown API error. Please contact Roo Code support." message.

[antsbuildit]

I'm facing the same issue on and off. Experienced it with Anthropic (Opus 4-5), 16k thinking token allowance.

[kavehsfv]

@roomote @hannesrudolph
I have same issue with openrouter selecting Mimo-V2-Flash
Error: Unknown API error. Please contact Roo Code support.

The problem starts when the following message is in the Thinking bubble

<tool_call>
<function=read_file>
<parameter=path>Depth-Anything-3/src/depth_anything_3/api.py</parameter>
</function>
</tool_call>

Another problem in thinking bubble

<tool_call>
<function=read_file>
<parameter=path>Depth-Anything-3/src/depth_anything_3/model/da3.py</parameter>
</function>
</tool_call><tool_call>
<function=read_file>
<parameter=path>Depth-Anything-3/src/depth_anything_3/cfg.py</parameter>
</function>
</tool_call>

I have seen your comment and will investigate the issue with Mimo-V2-Flash model outputting tool calls in a non-standard format (using <tool_call><function=...> tags) in the thinking bubble. I will explore a potential fix and open a PR to share progress.

Opened a PR with my attempt at addressing this: #10136

The fix adds handling for the case where the model outputs reasoning content (like the malformed tool_call tags in the thinking bubble) but no actual text or tool call content. This mirrors an existing safeguard in the Gemini provider. Feedback welcome!