Gemini provider: tool calls with errors enter infinite retry loop

[imtiazqa]

When using the Gemini LLM provider, tools that return an error
(isError: true) enter an infinite retry loop instead of returning a
response to the user. The request must be manually cancelled. The same
tools work correctly with the Anthropic provider.

Configuration

llm:
  enabled: true
  provider: gemini
  model: gemini-2.5-flash
  gemini_api_key_file: "/tmp/.gemini-api-key"
  per_attempt_timeout: 60

Tools Tested

Tool	Result
list_database_connections	Pass
get_schema_info	Pass
read_resource	Infinite loop
count_rows	Infinite loop
query_database	Infinite loop

Steps to Reproduce

1. Configure the server with provider: gemini, model: gemini-2.5-flash
2. Open the web client and authenticate
3. Ask: "Read the system info resource"
   → read_resource loops indefinitely
4. Ask: "How many rows are in the docs table?"
   → count_rows loops indefinitely
5. Ask: "Query the primary database: SELECT version()"
   → query_database loops indefinitely

Actual Result

Tools enter an infinite retry loop (40+ calls observed) until the
user manually cancels. Example:

read_resource ❌ → read_resource ❌ → read_resource ❌ → ...

Expected Result

All tools return a result and the LLM responds to the user .

Possible Root Cause

The pgedge-go-llm-lib proxy converts MCP tool results into Gemini's
functionResponse format. When a tool returns isError: true, the proxy
likely passes it without Gemini-specific error formatting. Gemini does
not recognize isError: true as a terminal failure and treats it as an
unresolved function call to retry. Anthropic natively understands
isError: true as part of the MCP spec.

Gemini expects errors formatted as:

{
"functionResponse": {
"name": "tool_name",
"response": {
"error": "error message here"
}
}
}