# Troubleshooting

> Fixes for the common problems: model servers that cannot be reached, 401s and 402s from cloud providers, models that will not call tools, skills that are not used, and too many approval prompts.

GBS SE Agent · Guides · Source: https://kit.gramproindia.com/se-agent/troubleshooting · Code: https://github.com/gramproindia/agent

For anything not covered here, **GBS Agent: Show Logs** records every request,
tool call, retry and error, with timings.

## Models and endpoints

- **"Could not reach Ollama"**: start Ollama (`ollama serve`) or check
  `gbsAgent.ollama.baseUrl`.
- **Model not found (HTTP 404)**: run `ollama pull <model>`, or pick a model
  from the model chip.
- **Hugging Face returns 401**: the token lacks the **Make calls to Inference
  Providers** permission, or it was saved for a different endpoint. Run
  **GBS Agent: Set API Key** again after changing `gbsAgent.openai.baseUrl`.
- **Hugging Face returns 402, or mentions credits**: the monthly credits are
  spent. Add a payment method, wait for the reset, or switch back to a local
  model.
- **A Hugging Face model id returns 404**: no provider is serving it right now.
  Check the model's page on the Hub, or pin a provider with a suffix such as
  `:groq`.
- **The model list is empty for OpenAI-compatible**: the endpoint could not be
  listed, usually because no key is saved yet or the base URL is missing its
  `/v1`. Type the model id meanwhile.
- **"Server overload" or 429 that keeps coming back**: the agent already retries
  three times. A free tier's per-minute cap may be too low for agent work; see
  [How free tiers run out](https://kit.gramproindia.com/se-agent/cloud-models#how-free-tiers-run-out).
- **Asked to send a saved API key to a local server**: choose **Cancel**. Local
  servers do not need the key.

## The agent's behaviour

- **It does not use tools, or stops early**: switch to a model trained for tool
  calling (see [Which models work](https://kit.gramproindia.com/se-agent/local-models#which-models-work)).
  For llama.cpp add `--jinja`; for vLLM add
  `--enable-auto-tool-choice --tool-call-parser <parser>`.
- **Slow, or out of memory**: use a smaller model, or set
  `gbsAgent.contextWindow` to `16384`.
- **It stops with a step-limit message on a big task**: raise
  `gbsAgent.maxSteps`, or split the task.
- **The context meter is high**: run `/compact`, or `/new` if you are switching
  tasks.
- **Too many approval prompts**: use **Always allow** for commands you repeat,
  and keep the mode on **Auto-edit**. Sensitive files always ask.

## Skills and MCP servers

- **A skill is not being used**: check `/skills`. It may be invalid or not
  allowed for this workspace (**GBS Agent: Reload Skills** asks again).
  Otherwise make the `description` name the words you actually use, or add an
  `autoAttach` glob.
- **Skill edits are not picked up**: skills are read once per request. Run
  **GBS Agent: Reload Skills**.
- **An MCP server failed to start**: see
  [MCP servers → Troubleshooting](https://kit.gramproindia.com/se-agent/mcp-servers#troubleshooting).

## The workspace

- **The robot icon does nothing, or the panel says the agent is disabled**: the
  folder is in Restricted Mode. Trust it from the notification or with
  **Workspaces: Manage Workspace Trust**.
- **Undo is not offered**: the window was reloaded since the turn. Originals
  are kept in memory only; use the diff view or git.