Let me thoroughly analyze the AI/generation code in /opt/bagger-api/. I'll start by exploring the directory structure and then dive into the relevant files.

Let me read all the relevant files in parallel.

Now let me also check the .env file referenced in the code to understand the Gemini configuration.

Now I have a thorough understanding of the entire codebase. Let me analyze the code in detail.

## 1. Model Configuration

In `game_engine.py:32-36`, the `gemini_model()` function:
```python
def gemini_model():
    model = os.environ.get("GEMINI_TEXT_MODEL") or os.environ.get("GEMINI_MODEL") or "gemini-2.5-flash"
    if "live" in model:
        return "gemini-2.5-flash"
    return model
```

The .env file has `GEMINI_MODEL=gemini-3.1-flash-live-preview`.

So `gemini_model()` will:
1. Check `GEMINI_TEXT_MODEL` env var → not set
2. Check `GEMINI_MODEL` env var → `"gemini-3.1-flash-live-preview"`
3. Check if "live" is in the model → YES, "live" is in "gemini-3.1-flash-live-preview"
4. Return `"gemini-2.5-flash"` (hardcoded fallback)

So the **actual model used** is `gemini-2.5-flash`, regardless of what the .env says, because the `if "live" in model` check catches "gemini-3.1-flash-live-preview" and replaces it with "gemini-2.5-flash".

## 2. How replies are generated (full flow)

### Flow: user interaction → reply generation

1. **`POST /api/interact`** in `app.py:97-98` calls `resolve_interaction(payload)`

2. **`resolve_interaction(payload)`** in `game_engine.py:644-705`:
   - Normalizes state (line 645)
   - Extracts intent (lines 646-651)
   - Applies schedule effects if type is "schedule" (line 661)
   - Applies gift effects if type is "gift" (line 662)
   - Selects scene (line 663) — picks a node from `ROUTES` based on state
   - **Composes prompt** (line 664) — `compose_prompt(state, intent, node, event, gift, schedule_effects)`
   - **Calls Gemini** (line 667): `call_gemini(prompt)` wrapped in `validate_ai()`
   - **On failure** (lines 668-670): falls back to `mock_gemini(prompt)` wrapped in `validate_ai()`
   - **Applies results** (line 671): `apply_scene_result(...)`
   - Returns the response dict (lines 689-705)

3. **`compose_prompt()`** in `game_engine.py:371-417` builds a long prompt instructing the model to act as a VN scene author. It includes:
   - Scene/world context
   - Bagger character description
   - Style bible rules
   - Game state, location, relationship data
   - The player's free text message
   - Gift/schedule effects
   - Instructions to return **valid JSON** with specific fields: `reply`, `emotionalRead`, `bondDelta`, `trustDelta`, `warmthDelta`, `depthDelta`, `courageDelta`, `suggestedFlags`, `routePressure`, `nextTone`, `memory`, `visual`

4. **`call_gemini(prompt)`** in `game_engine.py:420-444`:
   - **If `BAGGER_MOCK_GEMINI == "1"`**: calls `mock_gemini(prompt)` → static text
   - **If no API key**: calls `mock_gemini(prompt)` → static text
   - **Otherwise**: Makes a **non-streaming** HTTP POST to the Gemini API
     - URL: `https://generativelanguage.googleapis.com/v1beta/models/{gemini_model()}:generateContent?key={api_key}`
     - Temperature: 0.86
     - maxOutputTokens: 1100
     - `thinkingConfig: {"thinkingBudget": 0}` (no thinking/reasoning)
     - `responseMimeType: "application/json"` (forces structured JSON output)
   - Parses the JSON response back

5. **`validate_ai(data, node)`** in `game_engine.py:533-551` sanitizes the AI output:
   - If `data.get("reply")` is empty/falsy → **falls back to `mock_gemini("")["reply"]`** (static text!)
   - Clamps all numeric fields to valid ranges
   - Validates `nextTone` and `visual` against allowed values, defaulting to "listening"
   - Cleans/filters other fields

6. **`mock_gemini(prompt)`** in `game_engine.py:457-478` — the static fallback:
   - Always returns the same paragraph regardless of context (except minor variations based on matching "freund" or "work" in the prompt)
   - The reply text is always: *"Der Motor wurde leiser, als haette die Szene ploetzlich weniger Platz fuer Ausreden. Meine Schaufel blieb ueber dem Boden stehen. Wenn du morgen wiederkommst, werde ich so tun, als haette ich nicht darauf gewartet. Aber die Lampe in meiner Kabine weiss es besser."*
   - emotionalRead, bondDelta etc. vary slightly based on detecting "friend" or "work" in the player's message

7. **`call_gemini_for_ending(state, route, ending)`** in `game_engine.py:520-530`:
   - If `BAGGER_MOCK_GEMINI == "1"`: returns `FALLBACK_ENDINGS[kind]` (static pre-written text)
   - Calls `call_gemini(prompt)` for the ending
   - If that fails: falls back to `FALLBACK_ENDINGS[kind]`
   - `FALLBACK_ENDINGS` has pre-written text for each ending type (lines 510-517)

## 3. Where static/fallback text is

- **`mock_gemini()`** at `game_engine.py:457-478` — the main fallback for `reply`, `emotionalRead`, `memory`, `visual` fields
- **Line 539**: `validate_ai()` falls back to `mock_gemini("")["reply"]` if AI reply is empty
- **Lines 668-670**: `resolve_interaction()` catches errors and falls back to `mock_gemini(prompt)`
- **`FALLBACK_ENDINGS`** at `game_engine.py:510-517` — pre-written ending prose
- **Lines 524-525, 530**: `call_gemini_for_ending()` falls back to FALLBACK_ENDINGS

## 4. Streaming vs non-streaming

**Non-streaming only.** The API call at line 441 uses `urllib.request.urlopen()` which is a synchronous HTTP request to `generateContent` (not `streamGenerateContent`). There is no streaming anywhere in the codebase.

## 5. Key observations about fields

- **`reply`**: Generated by AI in `call_gemini()`, validated by `validate_ai()`, fallback to `mock_gemini()["reply"]`
- **`emotionalRead`**: From AI, or set to `"fallback after {error}"` in the error handler (line 670)
- **`visual`**: From AI, validated to be one of the allowed set, defaults to `"listening"`
- **`memory`**: From AI, or a fallback generated string in `apply_scene_result()` line 606
- **`endingProse`**: From `call_gemini_for_ending()`, which uses `FALLBACK_ENDINGS` (static)

## 6. Caching / pregeneration

There is **no caching** of dialogue anywhere. Every `resolve_interaction` call hits the Gemini API anew. The `.bak` files suggest manual backups, not caching.

## 7. game_data.py pre-written content

`game_data.py` has pre-written dialogue in:
- **`CHOICE_SETS`** (lines 306-362): These are the player's choice labels and messages, not the AI bagger replies
- **No pre-written bagger dialogue** — all bagger speech comes from the AI or the `mock_gemini` fallback

## 8. What would need to change to use gemini-3.1-flash-lite

The .env has `GEMINI_MODEL=gemini-3.1-flash-live-preview` but the code replaces "live" models with "gemini-2.5-flash".

To use `gemini-3.1-flash-lite` (or the correct model ID `gemini-3.1-flash-lite-preview` or whatever the actual ID is):
1. **Remove/modify the "live" check** in `gemini_model()` at `game_engine.py:34-35`
2. **Set either `GEMINI_TEXT_MODEL` or `GEMINI_MODEL`** to the desired model ID (e.g., `gemini-3.1-flash-lite-preview`)
3. ✅ The API key is already set in the .env file

The current behavior is that `gemini-3.1-flash-live-preview` gets downgraded to `gemini-2.5-flash` because of the "live" substring check.