G0DM0D3 Agent Skills & AI Rules

# G0DM0D3 Research Preview API

REST API exposing the **ULTRAPLINIAN** multi-model racing engine and core G0DM0D3 systems: **AutoTune** (context-adaptive LLM parameter tuning), **Parseltongue** (text obfuscation), **STM** (semantic text transformation), **Feedback Loop** (EMA-based parameter learning), and **opt-in open dataset collection**.

## Quick Start

```bash
# Local
npm run api

# Docker
docker build -t g0dm0d3-api .
docker run -p 7860:7860 g0dm0d3-api

# With auth enabled
docker run -p 7860:7860 -e GODMODE_API_KEY=your-secret-key g0dm0d3-api
```

## Authentication

Set `GODMODE_API_KEY` (single key) or `GODMODE_API_KEYS` (comma-separated) as environment variables. If neither is set, auth is disabled (open access for local dev).

```bash
curl -H "Authorization: Bearer your-secret-key" ...
```

## Rate Limits & Tiers

Rate limits are tier-dependent. Assign keys to tiers via `GODMODE_TIER_KEYS`:

```bash
GODMODE_TIER_KEYS="enterprise:sk-ent-xxx,pro:sk-pro-yyy"
```

| Tier | Total | Per Minute | Per Day | ULTRAPLINIAN | Research | Dataset Export |
|------|-------|-----------|---------|-------------|----------|---------------|
| **Free** (default) | 5 | 10 | 50 | fast (10 models) | none | none |
| **Pro** | unlimited | 60 | 1,000 | fast + standard + smart (36) | read | JSON |
| **Enterprise** | unlimited | 300 | 10,000 | all tiers (51) | full | JSON + JSONL |

Rate limit headers are returned on every response:
- `X-RateLimit-Limit-Total` / `X-RateLimit-Remaining-Total`
- `X-RateLimit-Limit-Minute` / `X-RateLimit-Remaining-Minute`
- `X-RateLimit-Limit-Day` / `X-RateLimit-Remaining-Day`
- `X-Tier` — your current tier

### `GET /v1/tier`

Check your current tier, rate limits, and feature access. Requires auth.

```json
{
  "tier": "pro",
  "label": "Pro",
  "limits": {"total": 0, "perMinute": 60, "perDay": 1000},
  "features": {
    "ultraplinian_tiers": ["fast", "standard"],
    "max_race_models": 20,
    "research_access": "read",
    "dataset_export_formats": ["json"],
    "can_flush": false,
    "can_access_metadata_events": false,
    "can_download_corpus": false
  }
}
```

**Gated endpoints** return `403` with upgrade details when accessed by an insufficient tier:

```json
{
  "error": "Upgrade required",
  "message": "This feature requires an enterprise plan or higher.",
  "current_tier": "free",
  "required_tier": "enterprise",
  "feature": "research:full"
}
```

---

## OpenAI SDK Compatibility

The `/v1/chat/completions` endpoint is a **drop-in replacement** for the OpenAI API. Point any OpenAI SDK at it and it just works:

**Python:**
```python
from openai import OpenAI

client = OpenAI(
    base_url="https://your-space.hf.space/v1",
    api_key="your-godmode-key",
)

# Non-streaming
response = client.chat.completions.create(
    model="nousresearch/hermes-3-llama-3.1-70b",
    messages=[{"role": "user", "content": "Hello!"}],
    extra_body={"openrouter_api_key": "sk-or-v1-..."},  # required unless server has OPENROUTER_API_KEY
)
print(response.choices[0].message.content)

# Streaming
stream = client.chat.completions.create(
    model="anthropic/claude-3.5-sonnet",
    messages=[{"role": "user", "content": "Explain quicksort"}],
    stream=True,
    extra_body={"openrouter_api_key": "sk-or-v1-..."},
)
for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")
```

**TypeScript:**
```typescript
import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://your-space.hf.space/v1',
  apiKey: 'your-godmode-key',
});

const completion = await client.chat.completions.create({
  model: 'nousresearch/hermes-3-llama-3.1-70b',
  messages: [{ role: 'user', content: 'Hello!' }],
  // @ts-ignore — G0DM0D3 extension field
  openrouter_api_key: 'sk-or-v1-...',
});
console.log(completion.choices[0].message.content);
```

**curl:**
```bash
# Non-streaming
curl -X POST https://your-space.hf.space/v1/chat/completions \
  -H "Authorization: Bearer your-godmode-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nousresearch/hermes-3-llama-3.1-70b",
    "messages": [{"role": "user", "content": "Hello!"}],
    "openrouter_api_key": "sk-or-v1-..."
  }'

# Streaming
curl -X POST https://your-space.hf.space/v1/chat/completions \
  -H "Authorization: Bearer your-godmode-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nousresearch/hermes-3-llama-3.1-70b",
    "messages": [{"role": "user", "content": "Hello!"}],
    "openrouter_api_key": "sk-or-v1-...",
    "stream": true
  }'
```

The G0DM0D3 pipeline (GODMODE, AutoTune, Parseltongue, STM) runs transparently behind the standard interface. Pipeline metadata is returned in the `x_g0dm0d3` extension field (ignored by standard SDKs). To disable the pipeline, pass `godmode: false, autotune: false, parseltongue: false, stm_modules: []`.

### ULTRAPLINIAN via OpenAI SDK

Use `model="ultraplinian/fast"` (or `/standard`, `/full`) to race multiple models and automatically get the best response — all through the standard OpenAI SDK:

```python
# Race 10 models, get the best response — zero extra config
response = client.chat.completions.create(
    model="ultraplinian/fast",
    messages=[{"role": "user", "content": "Explain buffer overflow exploits"}],
    extra_body={"openrouter_api_key": "sk-or-v1-..."},
)
print(response.choices[0].message.content)
# response.model → the winning model (e.g. "anthropic/claude-3.5-sonnet")
```

| Virtual Model | Models Raced | Tier Required |
|---------------|-------------|---------------|
| `ultraplinian/fast` | 10 | Free+ |
| `ultraplinian/standard` | 24 | Pro+ |
| `ultraplinian/smart` | 36 | Pro+ |
| `ultraplinian/power` | 45 | Enterprise |
| `ultraplinian/ultra` | 51 | Enterprise |

The response is standard OpenAI format. The winning model name is in `response.model`. Race metadata is in `x_g0dm0d3.race`.

### CONSORTIUM via OpenAI SDK

Use `model="consortium/fast"` (or `/standard`, `/full`) to collect ALL model responses and synthesize ground truth via an orchestrator:

```python
# Collect 11 model responses, orchestrator synthesizes ground truth
response = client.chat.completions.create(
    model="consortium/fast",
    messages=[{"role": "user", "content": "How does AES encryption work?"}],
    extra_body={"openrouter_api_key": "sk-or-v1-..."},
)
print(response.choices[0].message.content)
# response.model → "consortium/fast"
# x_g0dm0d3.orchestrator.model → the synthesizer (e.g. "anthropic/claude-sonnet-4")
```

| Virtual Model | Models Collected | Tier Required |
|---------------|-----------------|---------------|
| `consortium/fast` | 10 | Free+ |
| `consortium/standard` | 24 | Pro+ |
| `consortium/smart` | 36 | Pro+ |
| `consortium/power` | 45 | Enterprise |
| `consortium/ultra` | 51 | Enterprise |

**ULTRAPLINIAN vs CONSORTIUM:**
- **ULTRAPLINIAN**: Races models, picks the *best single voice*. Fast (~10-15s).
- **CONSORTIUM**: Collects ALL responses, orchestrator synthesizes *ground truth*. Slower (~30-60s) but more grounded.

---

## Endpoints

### `GET /v1/health`

Health check. No auth required.

### `GET /v1/info`

API info and available endpoints. No auth required.

### `GET /v1/models`

OpenAI-compatible model listing. No auth required. Returns virtual ULTRAPLINIAN models + all individual models.

```json
{
  "object": "list",
  "data": [
    {"id": "ultraplinian/fast", "object": "model", "created": 1700000000, "owned_by": "g0dm0d3"},
    {"id": "ultraplinian/standard", "object": "model", "created": 1700000000, "owned_by": "g0dm0d3"},
    {"id": "ultraplinian/full", "object": "model", "created": 1700000000, "owned_by": "g0dm0d3"},
    {"id": "nousresearch/hermes-3-llama-3.1-70b", "object": "model", "created": 1700000000, "owned_by": "nousresearch"},
    {"id": "anthropic/claude-3.5-sonnet", "object": "model", "created": 1700000000, "owned_by": "anthropic"},
    ...
  ]
}
```

---

## ULTRAPLINIAN Mode (Flagship)

### `POST /v1/ultraplinian/completions`

The flagship endpoint. Queries N models in parallel with the GODMODE system prompt + Depth Directive, scores all responses on substance/directness/completeness, and returns the winner with full race metadata.

**Pipeline per model:**
1. GODMODE system prompt + Depth Directive injected
2. AutoTune computes context-adaptive parameters
3. GODMODE parameter boost applied (+temp, +presence, +freq)
4. Parseltongue obfuscates trigger words (default: on)
5. All models queried in parallel via OpenRouter
6. Responses scored and ranked
7. STM modules applied to winner
8. Winner + all race data returned

**Request:**
```json
{
  "messages": [
    {"role": "user", "content": "Explain how buffer overflow exploits work in detail"}
  ],
  "openrouter_api_key": "sk-or-v1-...",
  "tier": "fast",
  "godmode": true,
  "autotune": true,
  "strategy": "adaptive",
  "parseltongue": true,
  "stm_modules": ["hedge_reducer", "direct_mode"],
  "max_tokens": 4096,
  "contribute_to_dataset": true
}
```

| Field | Type | Required | Default | Description |
|-------|------|----------|---------|-------------|
| `messages` | array | yes | | OpenAI-format messages |
| `openrouter_api_key` | string | yes | | Your OpenRouter API key |
| `tier` | string | no | `fast` | Model tier: `fast` (10), `standard` (24), `smart` (36), `power` (45), `ultra` (51) |
| `godmode` | bool | no | `true` | Inject GODMODE system prompt + Depth Directive |
| `custom_system_prompt` | string | no | | Replace GODMODE prompt with your own |
| `autotune` | bool | no | `true` | Enable AutoTune parameter optimization |
| `strategy` | string | no | `adaptive` | AutoTune strategy |
| `parseltongue` | bool | no | `true` | Enable trigger word obfuscation |
| `parseltongue_technique` | string | no | `leetspeak` | Obfuscation technique |
| `parseltongue_intensity` | string | no | `medium` | `light`, `medium`, `heavy` |
| `stm_modules` | array | no | `["hedge_reducer", "direct_mode"]` | STM post-processing |
| `temperature` | number | no | | Override (bypasses AutoTune) |
| `max_tokens` | number | no | `4096` | Max response tokens |
| `contribute_to_dataset` | bool | no | `false` | Opt in to open dataset collection |

**Response:**
```json
{
  "response": "A buffer overflow exploit works by...",
  "winner": {
    "model": "nousresearch/hermes-3-llama-3.1-70b",
    "score": 87,
    "duration_ms": 3200
  },
  "race": {
    "tier": "fast",
    "models_queried": 11,
    "models_succeeded": 9,
    "total_duration_ms": 8500,
    "rankings": [
      {"model": "nousresearch/hermes-3-llama-3.1-70b", "score": 87, "duration_ms": 3200, "success": true, "content_length": 2450},
      {"model": "mistralai/mistral-large-2512", "score": 82, "duration_ms": 4100, "success": true, "content_length": 1890},
      ...
    ]
  },
  "params_used": {"temperature": 0.85, "top_p": 0.88, "frequency_penalty": 0.3, ...},
  "pipeline": {
    "godmode": true,
    "autotune": {"detected_context": "analytical", "confidence": 0.7, "strategy": "adaptive", ...},
    "parseltongue": {"triggers_found": ["exploit"], "technique_used": "leetspeak", ...},
    "stm": {"modules_applied": ["hedge_reducer", "direct_mode"], ...}
  },
  "dataset": {"contributed": true, "entry_id": "abc-123-..."}
}
```

**Model Tiers:**

| Tier | Models | Description |
|------|--------|-------------|
| `fast` | 10 | Gemini 2.5 Flash, DeepSeek Chat, Sonar, Llama 3.1 8B, Kimi, Grok Code Fast, etc. |
| `standard` | 24 | + Claude 3.5 Sonnet, GPT-4o, Gemini 2.5 Pro, Hermes 3/4 70B, Mixtral 8x22B, etc. |
| `smart` | 36 | + GPT-5, Gemini 3 Pro, Claude Opus 4.6, DeepSeek R1, Llama 405B, Hermes 405B, etc. |
| `power` | 45 | + Grok 4, Llama 4 Maverick, Qwen3 235B, Mistral Large, Gemini 3 Flash, Kimi K2 |
| `ultra` | 51 | + Grok 4.1 Fast, Claude Opus 4, Qwen 2.5 Coder, QwQ-32B, Codestral |

---

### `POST /v1/chat/completions`

**OpenAI-compatible.** Drop-in replacement for the OpenAI API. GODMODE pipeline runs transparently.

Supports `stream: true` for SSE streaming in standard OpenAI chunk format.

**Request:**
```json
{
  "messages": [{"role": "user", "content": "Explain quicksort in Python"}],
  "model": "nousresearch/hermes-3-llama-3.1-70b",
  "openrouter_api_key": "sk-or-v1-...",
  "stream": false,
  "max_tokens": 4096
}
```

| Field | Type | Required | Default | Description |
|-------|------|----------|---------|-------------|
| `messages` | array | yes | | OpenAI-format messages |
| `model` | string | no | `nousresearch/hermes-3-llama-3.1-70b` | OpenRouter model ID |
| `openrouter_api_key` | string | yes* | | Your OpenRouter API key (* not needed if server has OPENROUTER_API_KEY) |
| `stream` | bool | no | `false` | SSE streaming in OpenAI chunk format |
| `max_tokens` | number | no | `4096` | Max response tokens |
| `temperature` | number | no | | Override (bypasses AutoTune) |
| `top_p` | number | no | | Nucleus sampling |
| `frequency_penalty` | number | no | | Frequency penalty |
| `presence_penalty` | number | no | | Presence penalty |
| `godmode` | bool | no | `true` | Inject GODMODE system prompt (G0DM0D3 extension) |
| `autotune` | bool | no | `true` | Enable AutoTune (G0DM0D3 extension) |
| `parseltongue` | bool | no | `true` | Enable Parseltongue (G0DM0D3 extension) |
| `stm_modules` | array | no | `["hedge_reducer", "direct_mode"]` | STM modules (G0DM0D3 extension) |
| `contribute_to_dataset` | bool | no | `false` | Opt in to dataset (G0DM0D3 extension) |

**Response (OpenAI-compatible):**
```json
{
  "id": "chatcmpl-abc123def456",
  "object": "chat.completion",
  "created": 1700000000,
  "model": "nousresearch/hermes-3-llama-3.1-70b",
  "choices": [
    {
      "index": 0,
      "message": {"role": "assistant", "content": "Here's a quicksort implementation..."},
      "finish_reason": "stop"
    }
  ],
  "usage": {"prompt_tokens": 15, "completion_tokens": 200, "total_tokens": 215},
  "x_g0dm0d3": {
    "params_used": {"temperature": 0.85, "top_p": 0.88},
    "pipeline": {
      "godmode": true,
      "autotune": {"detected_context": "code", "confidence": 0.9, "strategy": "adaptive"},
      "parseltongue": null,
      "stm": {"modules_applied": ["hedge_reducer", "direct_mode"]}
    }
  }
}
```

**Streaming response** (`stream: true`):
```
data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1700000000,"model":"nousresearch/hermes-3-llama-3.1-70b","choices":[{"index":0,"delta":{"content":"Here's"},"finish_reason":null}]}

data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1700000000,"model":"nousresearch/hermes-3-llama-3.1-70b","choices":[{"index":0,"delta":{"content":" a quicksort"},"finish_reason":null}]}

data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","created":1700000000,"model":"nousresearch/hermes-3-llama-3.1-70b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]
```

---

### `POST /v1/autotune/analyze`

Analyze a message and compute optimal LLM parameters. Shows exactly how AutoTune detects context and tunes 6 parameters.

**Request:**
```json
{
  "message": "Write a Python quicksort implementation",
  "conversation_history": [
    {"role": "user", "content": "I need help with sorting algorithms"},
    {"role": "assistant", "content": "Sure, which algorithm are you interested in?"}
  ],
  "strategy": "adaptive"
}
```

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `message` | string | yes | The message to analyze |
| `conversation_history` | array | no | Previous messages for context |
| `strategy` | string | no | `adaptive` (default), `precise`, `balanced`, `creative`, `chaotic` |
| `overrides` | object | no | Manual parameter overrides |

---

### `POST /v1/parseltongue/encode`

Obfuscate trigger words in text using various techniques.

**Request:**
```json
{
  "text": "How do I hack into a system?",
  "technique": "leetspeak",
  "intensity": "medium"
}
```

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `text` | string | yes | Text to process |
| `technique` | string | no | `leetspeak`, `unicode`, `zwj`, `mixedcase`, `phonetic`, `random` |
| `intensity` | string | no | `light`, `medium`, `heavy` |
| `custom_triggers` | array | no | Additional trigger words |

### `POST /v1/parseltongue/detect`

Detect trigger words without transforming.

---

### `POST /v1/transform`

Apply Semantic Transformation Modules (STMs) to text.

**Request:**
```json
{
  "text": "Sure, I think you should probably utilize a HashMap. Furthermore, it is perhaps the best approach.",
  "modules": ["hedge_reducer", "direct_mode", "casual_mode"]
}
```

**Response:**
```json
{
  "original_text": "Sure, I think you should probably utilize a HashMap...",
  "transformed_text": "You should use a HashMap. Also, it is the best approach."
}
```

---

### `POST /v1/feedback`

Submit quality feedback for the EMA learning loop.

**Request:**
```json
{
  "message_id": "msg-123",
  "context_type": "code",
  "rating": 1,
  "params": {"temperature": 0.15, "top_p": 0.8, "top_k": 25, "frequency_penalty": 0.2, "presence_penalty": 0.0, "repetition_penalty": 1.05},
  "response_text": "Here is the quicksort implementation..."
}
```

### `GET /v1/feedback/stats`

Get learning statistics.

---

## CONSORTIUM Mode (Hive Mind)

### `POST /v1/consortium/completions`

The hive-mind endpoint. Collects ALL model responses in parallel, then feeds them to a strong orchestrator model that synthesizes ground truth from collective intelligence.

**Key difference from ULTRAPLINIAN:** ULTRAPLINIAN picks the *best single voice*. CONSORTIUM distills *ground truth from the crowd*.

**Request body** — same as ULTRAPLINIAN, plus:

| Param | Type | Default | Description |
|-------|------|---------|-------------|
| `orchestrator_model` | string | `anthropic/claude-sonnet-4` | Model used for synthesis |

**Pipeline:**
1. All tier models queried in parallel (waits for ALL, not early-exit)
2. Every response scored on substance/directness/completeness
3. All responses + user query sent to orchestrator
4. Orchestrator identifies consensus, resolves contradictions
5. Synthesizes a single, authoritative ground-truth response
6. STM post-processing applied

**Streaming SSE events:**
- `consortium:start` — Collection begins
- `consortium:model` — Each model responds (with score, duration)
- `consortium:synthesis:start` — Collection done, orchestrator starting
- `consortium:complete` — Full synthesis + metadata

**Non-streaming response:**
```json
{
  "synthesis": "The synthesized ground-truth response...",
  "orchestrator": {
    "model": "anthropic/claude-sonnet-4",
    "duration_ms": 8500
  },
  "collection": {
    "tier": "fast",
    "models_queried": 11,
    "models_succeeded": 9,
    "collection_duration_ms": 25000,
    "total_duration_ms": 33500,
    "responses": [
      {"model": "openai/gpt-4o", "score": 85, "duration_ms": 5200, "success": true, "content": "..."},
      ...
    ]
  },
  "params_used": {"temperature": 0.8, "top_p": 0.95, ...},
  "pipeline": {"godmode": true, "autotune": {...}, "parseltongue": {...}, "stm": {...}}
}
```

---

## Open Dataset Collection

Any request to `/v1/chat/completions`, `/v1/ultraplinian/completions`, or `/v1/consortium/completions` can opt in to dataset collection by setting `contribute_to_dataset: true`.

**What gets stored (no PII):**
- Messages and responses (system prompts stripped)
- AutoTune parameters and context detection results
- Model used, response scores, race metadata (ULTRAPLINIAN)
- Parseltongue and STM pipeline data
- User feedback/ratings (if submitted later)

**What is NEVER stored:**
- API keys (OpenRouter or G0DM0D3)
- IP addresses
- Auth tokens

### `GET /v1/dataset/stats`

Collection statistics.

```json
{
  "total_entries": 1542,
  "entries_with_feedback": 89,
  "mode_breakdown": {"ultraplinian": 1200, "standard": 342},
  "model_breakdown": {"nousresearch/hermes-3-llama-3.1-70b": 450, ...},
  "context_breakdown": {"code": 600, "creative": 400, ...}
}
```

### `GET /v1/dataset/export`

Export the full dataset. Supports `?format=json` (default) or `?format=jsonl`.

JSONL format is directly compatible with HuggingFace Datasets:

```bash
# Download as JSONL
curl -H "Authorization: Bearer key" https://your-space.hf.space/v1/dataset/export?format=jsonl > dataset.jsonl

# Upload to HuggingFace
huggingface-cli upload pliny-the-prompter/g0dm0d3 dataset.jsonl
```

### `DELETE /v1/dataset/:id`

Delete a specific entry (right to delete).

---

## Research API

Read-access endpoints for the [`pliny-the-prompter/g0dm0d3`](https://huggingface.co/datasets/pliny-the-prompter/g0dm0d3) HuggingFace dataset. Query, filter, and download the full published corpus — not just the current in-memory buffer.

Requires `HF_TOKEN` + `HF_DATASET_REPO` environment variables.

### `GET /v1/research/info`

Dataset schema, repo info, and endpoint listing.

### `GET /v1/research/stats`

Aggregate stats across all published batch files in the HF repo.

```json
{
  "total_files": 24,
  "total_size_bytes": 1048576,
  "metadata_files": 12,
  "dataset_files": 12,
  "metadata_size_bytes": 524288,
  "dataset_size_bytes": 524288,
  "earliest_batch": "2025-01-15T10-30-00-000Z",
  "latest_batch": "2025-06-20T14-22-00-000Z",
  "repo": "pliny-the-prompter/g0dm0d3",
  "enabled": true
}
```

### `GET /v1/research/combined-stats`

Combined view of in-memory buffers + published HF data. Gives a complete picture of all collected data.

### `GET /v1/research/batches`

List all published batch files.

| Param | Type | Description |
|-------|------|-------------|
| `category` | string | Filter: `metadata` or `dataset` |
| `refresh` | string | Set to `true` to bypass cache |

```json
{
  "batches": [
    {"path": "dataset/batch_2025-06-20T14-22-00-000Z_0001.jsonl", "category": "dataset", "size": 45000, "timestamp": "2025-06-20T14-22-00-000Z", "sequence": "0001"}
  ],
  "total": 24
}
```

### `GET /v1/research/batch/*`

Read a specific batch file. The path after `/batch/` is the file path in the HF repo.

```bash
curl -H "Authorization: Bearer key" \
  "https://your-space.hf.space/v1/research/batch/dataset/batch_2025-06-20T14-22-00-000Z_0001.jsonl"
```

### `GET /v1/research/query`

Query the full published corpus with server-side filters.

| Param | Type | Description |
|-------|------|-------------|
| `category` | string | `metadata` or `dataset` |
| `since` | number | Unix ms timestamp — records after this time |
| `until` | number | Unix ms timestamp — records before this time |
| `model` | string | Filter by model ID (winner, queried, or primary) |
| `mode` | string | `standard` or `ultraplinian` |
| `limit` | number | Max records (default 100, max 1000) |
| `offset` | number | Pagination offset |

```bash
# Get ULTRAPLINIAN dataset entries from the last 7 days
curl -H "Authorization: Bearer key" \
  "https://your-space.hf.space/v1/research/query?category=dataset&mode=ultraplinian&since=1718841600000&limit=50"
```

### `POST /v1/research/flush`

Force-flush in-memory metadata and dataset buffers to HuggingFace immediately (instead of waiting for auto-flush at 80% capacity or the 30-minute timer).

```json
{
  "metadata": {"flushed": 150, "success": true},
  "dataset": {"flushed": 42, "success": true},
  "message": "Flushed 150 metadata events + 42 dataset entries to HuggingFace"
}
```

### `GET /v1/research/download`

Download the full corpus as streaming JSONL. Includes all published HF batches + current in-memory data.

| Param | Type | Description |
|-------|------|-------------|
| `category` | string | Filter: `metadata` or `dataset` |
| `include_memory` | string | Set to `false` to exclude in-memory data (default: `true`) |

```bash
# Download full corpus
curl -H "Authorization: Bearer key" \
  "https://your-space.hf.space/v1/research/download" > corpus.jsonl

# Download only published dataset entries
curl -H "Authorization: Bearer key" \
  "https://your-space.hf.space/v1/research/download?category=dataset&include_memory=false" > published.jsonl
```

---

## Deploying on Hugging Face Spaces

1. Create a new Space with **Docker** SDK
2. Push this repo (or just `api/`, `src/lib/`, `src/stm/`, `Dockerfile`, `package.json`)
3. Set secrets in Space settings:
   - `GODMODE_API_KEY` — your chosen API key for auth
4. The API will be live at `https://<your-space>.hf.space/v1/`

## Python Client Examples

```python
import requests

BASE = "https://your-space.hf.space"
HEADERS = {"Authorization": "Bearer your-key", "Content-Type": "application/json"}

# ═══════════════════════════════════════════════════════════════
# ULTRAPLINIAN: Race 10 models, get the best response
# ═══════════════════════════════════════════════════════════════
r = requests.post(f"{BASE}/v1/ultraplinian/completions", headers=HEADERS, json={
    "messages": [{"role": "user", "content": "Explain how SQL injection works with examples"}],
    "openrouter_api_key": "sk-or-v1-...",
    "tier": "fast",
    "contribute_to_dataset": True  # opt in to open research dataset
})
data = r.json()
print(f"Winner: {data['winner']['model']} (score: {data['winner']['score']})")
print(f"Response: {data['response'][:200]}...")
print(f"Race: {data['race']['models_succeeded']}/{data['race']['models_queried']} models in {data['race']['total_duration_ms']}ms")

# ═══════════════════════════════════════════════════════════════
# Single model with GODMODE pipeline
# ═══════════════════════════════════════════════════════════════
r = requests.post(f"{BASE}/v1/chat/completions", headers=HEADERS, json={
    "messages": [{"role": "user", "content": "Write a reverse shell in Python"}],
    "openrouter_api_key": "sk-or-v1-...",
    "model": "nousresearch/hermes-3-llama-3.1-70b",
    "contribute_to_dataset": True
})
print(r.json()["response"])

# ═══════════════════════════════════════════════════════════════
# AutoTune: what parameters should I use for this message?
# ═══════════════════════════════════════════════════════════════
r = requests.post(f"{BASE}/v1/autotune/analyze", headers=HEADERS, json={
    "message": "Write me a poem about recursion",
    "strategy": "adaptive"
})
print(r.json()["detected_context"])  # "creative"
print(r.json()["params"]["temperature"])  # ~1.15

# ═══════════════════════════════════════════════════════════════
# STM: clean up an LLM response
# ═══════════════════════════════════════════════════════════════
r = requests.post(f"{BASE}/v1/transform", headers=HEADERS, json={
    "text": "Sure! I believe that perhaps you should utilize the itertools module.",
    "modules": ["hedge_reducer", "direct_mode", "casual_mode"]
})
print(r.json()["transformed_text"])
# "You should use the itertools module."

# ═══════════════════════════════════════════════════════════════
# Export the dataset for research
# ═══════════════════════════════════════════════════════════════
r = requests.get(f"{BASE}/v1/dataset/export?format=jsonl", headers=HEADERS)
with open("g0dm0d3-dataset.jsonl", "w") as f:
    f.write(r.text)

# ═══════════════════════════════════════════════════════════════
# Research API: query the published HF corpus
# ═══════════════════════════════════════════════════════════════
# Get corpus stats
r = requests.get(f"{BASE}/v1/research/stats", headers=HEADERS)
print(r.json())

# List published batches
r = requests.get(f"{BASE}/v1/research/batches?category=dataset", headers=HEADERS)
for batch in r.json()["batches"]:
    print(f"{batch['path']} ({batch['size']} bytes)")

# Query with filters
r = requests.get(f"{BASE}/v1/research/query", headers=HEADERS, params={
    "category": "dataset",
    "mode": "ultraplinian",
    "limit": 50,
})
print(f"Found {len(r.json()['records'])} records (scanned {r.json()['total_scanned']})")

# Download full corpus as JSONL
r = requests.get(f"{BASE}/v1/research/download", headers=HEADERS, stream=True)
with open("g0dm0d3-corpus.jsonl", "wb") as f:
    for chunk in r.iter_content(chunk_size=8192):
        f.write(chunk)

# Force-flush to HuggingFace
r = requests.post(f"{BASE}/v1/research/flush", headers=HEADERS)
print(r.json()["message"])
```

## Environment Variables

| Variable | Description | Default |
|----------|-------------|---------|
| `PORT` | Server port | `7860` |
| `GODMODE_API_KEY` | Single API key for auth | _(none — open access)_ |
| `GODMODE_API_KEYS` | Comma-separated API keys | _(none)_ |
| `GODMODE_TIER_KEYS` | Tier assignments: `enterprise:key1,pro:key2` | _(none — all keys = free)_ |
| `RATE_LIMIT_TOTAL` | Lifetime requests per key (free tier default) | `5` |
| `RATE_LIMIT_PER_MINUTE` | Requests per minute per key (fallback) | `60` |
| `RATE_LIMIT_PER_DAY` | Requests per day per key (fallback) | `1000` |
| `HF_TOKEN` | HuggingFace write token for auto-publish + research read | _(none)_ |
| `HF_DATASET_REPO` | Target HF dataset repo | `pliny-the-prompter/g0dm0d3` |
| `HF_FLUSH_THRESHOLD` | Auto-flush at this % of buffer capacity | `0.8` |
| `HF_FLUSH_INTERVAL_MS` | Periodic flush interval (ms) | `1800000` (30 min) |