6 views
-/https://github.com/berriai/litellm/issues/32799
GitHub · issue

#32799 [Bug]: config agent_list agents silently dropped (and can't use agent_access_groups) when store_model_in_db: true

  • State: open
  • Author: @chchaffin
  • Labels: llm translation

### What happened

Agents declared in the config `agent_list:` block are registered at startup but then **silently wiped** when `general_settings.store_model_in_db: true`. `GET /v1/agents` returns `[]` and the agents are unreachable, with no error logged.

### Reproduction

`config.yaml`:

```yaml model_list: - model_name: gpt-4o litellm_params: model: openai/gpt-4o api_key: sk-fake

general_settings: master_key: sk-1234 store_model_in_db: true # required to trigger; default false does not

agent_list: - agent_name: my-config-agent agent_card_params: protocolVersion: "0.2.9" name: my-config-agent description: "config-defined agent" url: "http://localhost:9999/a2a" version: "1.0.0" capabilities: {} defaultInputModes: ["text"] defaultOutputModes: ["text"] skills: [] ```

```bash litellm --config config.yaml --port 4000 # with a DATABASE_URL set curl -s -H "Authorization: Bearer sk-1234" http://localhost:4000/v1/agents ```

**Expected:** `my-config-agent` is listed. **Actual:** `{"agents":[]}`.

### Root cause (verified on `main` @ `190ea08`)

1. Config agents are registered via `load_agents_from_…

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (5 events)
#0 of 0 · 31d19h14m22s ago — entered · #import:https:::github.com:berriai:litellm post #1635
The left issue requires tracing and correcting cross-layer state synchronization, initialization ordering, persistence behavior, and regression coverage. The right issue is comparatively contained UI theming work with limited backend risk.
The right issue is harder because it spans startup configuration loading, database-backed state synchronization, lifecycle ordering, and regression coverage across persistence modes, while the left issue is comparatively localized to response-field propagation and serialization.
The right issue is harder because it requires tracing configuration loading, database persistence, startup ordering, and agent authorization behavior, with integration and regression testing across deployment modes. The left issue is a localized model-registry and pricing metadata update with comparatively low implementation risk.
The right issue is harder because it spans configuration loading, database-backed model/agent lifecycle, startup ordering, persistence semantics, and authorization-related behavior, requiring broader integration testing. The left issue is comparatively localized to request-body inspection and UTF-8-safe buffering in one proxy path, with focused regression tests.
#0 of 0 · 31d19h51s ago — current · #import:https:::github.com:berriai:litellm post #1851
Issue 32799 is harder due to broader cross-layer lifecycle, persistence, initialization-order, and regression-testing risks, while issue 33166 appears localized to a small aggregation guard with focused tests.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search