hermes - 💡(How to fix) Fix [Bug]: API server returns output-truncation failure as successful assistant message [5 pull requests]

Bug Description

Issue

Hermes agentic loop can detect output truncation and attempt continuation. However, when truncation recovery ultimately fails, the API server currently flattens the failure into a normal /v1/chat/completions response.

Example assistant content:

Response truncated due to output length limit

The response is still returned as a successful chat completion, so API clients treat it as a valid assistant answer instead of a failed or partial run.

Why this matters

Integrations using /v1/chat/completions cannot distinguish a real assistant answer from a failed truncated run. This causes clients to display raw internal failure text to end users.

Suggested fix

In /v1/chat/completions, avoid mapping result["error"] to normal assistant content when result["failed"] or result["partial"] is true. Preserve the failure state in a structured way so clients can handle it correctly.

Steps to Reproduce

1. Start Hermes API server.
2. Send a /v1/chat/completions request with a prompt that is likely to exceed the model output limit, for example a very long generation or a complex agentic task requiring a large final answer.
3. Let the agent run until Hermes detects output truncation and continuation recovery fails.
4. Observe the API response.

Example observed assistant content:

Response truncated due to output length limit

or:

Response remained truncated after 3 continuation attempts

Expected Behavior

The API should preserve the internal failed/partial state instead of flattening it into normal assistant content.

For example, return one of:

• a structured error response;
• a Hermes-specific metadata field/header;
• partial=true, completed=false, and a machine-readable error_code;
• or a non-2xx response when the result is not a valid assistant answer.

Actual Behavior

The API returns a successful chat completion response. The truncation failure is placed in:

choices[0].message.content

and the client receives it as if it were a normal assistant answer.

Affected Component

Gateway (Telegram/Discord/Slack/WhatsApp)

Messaging Platform (if gateway-related)

No response

Debug Report

From source inspection, the agentic loop already detects truncation.

In `run_agent.py`, `AIAgent.run_conversation()` handles `finish_reason == "length"` and attempts continuation. If recovery still fails, it returns a structured result like:


{
  "final_response": None,
  "completed": False,
  "partial": True,
  "error": "Response truncated due to output length limit"
}


However, in `gateway/platforms/api_server.py`, the non-streaming `/v1/chat/completions` handler maps this into normal assistant content:


final_response = result.get("final_response", "")
if not final_response:
    final_response = result.get("error", "(No response generated)")


That loses the `partial/error/completed=false` semantics. Downstream clients cannot distinguish a real assistant answer from an agent failure, so they display raw internal failure text to end users.

Operating System

Ubuntu

Python Version

No response

Hermes Version

No response

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

No response

Proposed Fix (optional)

No response

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR