For DEBUG=False, collect and serve static files.

python manage.py collectstatic

Also confirm your web server serves Django admin static files.

If the browser still shows old styles, hard refresh the page. Admin assets often use cache-busting query strings such as guided-v8.

Confirm the model admin uses the ai_hub styled change form or mixins.

Expected behavior:

• fieldsets are grouped,
• form rows have subtle panels,
• fields include help text,
• text fields include placeholders,
• JSON fields include JSON examples,
• errors are visible and readable.

If a new model admin is added, connect it to the shared form style.

Check:

• provider is active,
• model is active,
• api_key_env_var exists in the process environment,
• base_url is correct for local providers,
• local provider is reachable from the Django process.

For Ollama, confirm:

GET <base_url>/api/tags

returns installed models.

The control center may warn when a configured model is not reported by the provider.

Check:

• model name spelling,
• provider prefix,
• local model installation,
• active/inactive state.

Examples:

ollama/qwen3:8b
qwen3:8b
gpt-4.1-mini

The exact accepted form depends on the provider adapter.

The training provider is a deterministic stub. Its model name must be exactly training or start with training/ (for example training/assistant). Saving a training-provider model with any other name fails with:

Training-provider models must be named 'training' or start with 'training/'

This is intentional: any other name would not be routed to the stub and would fail at runtime against the real client. Rename the model to follow the convention.

The GAME workspace highlights agents that are already used in GAME sessions or look GAME-ready.

To make an agent easy to identify as GAME-ready, include goal-loop keys in the input contract:

{
  "required": ["goal", "iteration", "memory", "game_response_contract"]
}

Also mention goal, action, memory, observation, completion, and final answer in the prompt.

If strict response contracts are enabled, the agent must return valid JSON with:

• action
• message
• complete
• final_answer

Example:

{
  "action": "finish",
  "message": "The goal is complete.",
  "complete": true,
  "final_answer": "Final answer."
}

If this fails often, reduce prompt complexity and keep the first goal smaller.

A GAME service can raise:

GAME feature 'AI_HUB_GAME_..._ENABLED' is disabled. Set ... =True in Django settings to enable it.

This means the matching feature flag is off (the reusable default is fail-closed). Enable the flag in settings or the environment — see 03_CONFIGURATION.md. Remember flags gate the service layer only; the admin add/change forms write directly to the model.

A goal stays in running if its session never reached a terminal state (an interrupted run, or stub sessions left by tests). Cancel orphaned goals — those with no active session — with:

python manage.py cleanup_orphaned_goals --dry-run
python manage.py cleanup_orphaned_goals

A goal with an active session (pending/running/waiting_async) is never touched.

Check:

• it has at least one step,
• step order is continuous,
• each step agent is active,
• each agent has input and output contracts,
• fallback agents are active when required.

Common message:

final_output must be valid JSON

Possible causes:

• final model output was truncated,
• final model included commentary or markdown,
• final model generated malformed JSON,
• max_tokens_default was too low,
• final step is trying to combine too much data.

Recommended fixes:

1. Inspect the final step in the execution timeline.
2. Check final_context.final_output.
3. Check final_context.final_output_parse_error.
4. Increase max tokens for the final model if truncation is likely.
5. Make the final step produce a smaller payload.
6. Move domain-specific recovery into the host adapter if intermediate drafts are valid.

The runtime protects against running the same session twice.

Create a new session if you want to rerun the same goal or source payload.

WAITING_ASYNC means the session intentionally stopped for continuation.

Use a worker or continuation process if your runtime mode expects asynchronous completion.

Check:

• tool is active,
• tool is attached to the agent,
• runtime supports the tool kind,
• model supports tool usage if the tool requires model-native tool calls,
• tool config is allowlisted and valid.

Warnings are intended to surface configuration drift.

Common causes:

• provider cannot be reached,
• configured model is not reported by provider,
• agent has missing or inactive dependencies,
• pipeline contains inactive or misconfigured steps.