Retrieve conversation data by key. Args: key: Storage key (hash or custom key) Returns: Conversatio

Context manager for LMDB transactions. Args: write: Whether the transaction should be writable.

Generate a hash for a list of messages and model name, tied to a specific client_id.

Generate a HMAC-SHA256 token for a filename using the API key.

Build conversation payload and optionally compact oversized histories.

Calculate prompt, completion, total and reasoning tokens consistently.

Persist an image provided by gemini_webapi and return base64 plus dimensions, filename, and hash.

Perform safe semantic normalization for hashing using helper utilities.

Unified logic to save conversation history to LMDB.

Return a copy of messages enriched with tool instructions when needed.

Post-process Gemini output to extract tool calls and prepare clean text for display and storage. Returns: (thoughts, visible_text, storage_ou

Remove a specific storage key from the index for a given hash.

Return a healthy client by id or using round-robin.

Close the LMDB environment.

Delete conversation model by key.

Remove system hints, ChatML tags, and technical protocol markers from text.

Remove CommonMark backslash escapes from LLM-generated text.

Decode index value, handling both legacy single-string and new list-of-strings formats.

Internal find implementation based on a message list. Args: model: Model name messages: Message list to hash

Generate a stable, canonical hash for a single message.

Determines if the current state allows yielding text to the stream.

Surround content with ChatML role tags.

Estimate the number of tokens heuristically based on character count.

Check if a key exists in the store.

Extract tool calls and return cleaned text.

Initialize all clients in the pool.

Clean all messages of internal markers, hints and normalize tool calls.

Ensure the last user message carries the tool wrap hint.

Translate OpenAI-style response_format into internal instructions.

Unified factory for building ResponseCreateResponse objects.

Decode a value from LMDB to Python data.

Compute effective request size guardrail from config values.

Make sure the client is running, attempting a restart if needed.

Find an existing chat session matching the longest suitable history prefix.

Retrieve a Model instance by name.

Extract tool metadata and return text stripped of technical markers. Arguments are parsed into JSON and assigned deterministic call IDs.

Send text to Gemini, splitting or converting to attachment if too long.

Add a storage key to the index for a given hash, avoiding duplicates.

Search conversation data by message list. Tries raw matching, then sanitized matching, and finally fuzzy matching. Args:

Release remaining buffer content and perform final cleanup at stream end.

Returns a persistent directory for storing images.

Safely normalize LLM-generated text for both display and hashing. Includes: HTML unescaping, NFC normalization, and line ending standardizati

Process a Message into Gemini API format using the PascalCase technical protocol. Extracts text, handles files, and appends ToolCalls

Strip tool call blocks from text for display.

Download content from a URL and save to a temporary file.

Concatenate text and tool arguments from a message for token estimation.

Create a compact summary message for older turns to reduce oversized replay payloads.

Construct explicit guidance so Gemini emits images when requested.

Generate a system prompt describing available tools and the PascalCase protocol.

Keep recent turns verbatim and compact older turns into one summary message.

Unified factory for building Chat Completion response objects.

Create a real-time streaming response. Reconciles manual delta accumulation with the model's final authoritative state.

Create a real-time streaming response for the Responses API. Ensures final accumulated text and thoughts are synchronized and follow the form

Return all records from the database.

Return records for the provided keys.

Create database directory if it doesn't exist.

Lowercase and remove all whitespace and punctuation. Used as a fallback for complex/malformed contents matching.

Return a list of available models based on configuration strategy.

Initialize LMDB environment.

Normalize instructions payload into Message objects.

Override base_clients with env_overrides, return the new clients list.

Override base_models with env_overrides using standard update (replace whole fields).

Parse duration in the format '14d' or '24h'.

Convert Responses API input items into internal Message objects and normalized input.

Periodically enforce LMDB retention policy until the stop_event is set.

Setup standard logging library interceptor to redirect to loguru

Check if the record is older than the threshold.

Remove one layer of outermost Markdown code fences, supporting nested blocks by detecting variable fence lengths.

Delete conversations older than the given retention period.

Delete images in IMAGE_STORE_DIR older than retention_days.

Detect image extension from magic bytes.

Print selected or all key-value pairs from the LMDB database.

Extract and remove all Gemini clients related environment variables, return a mapping from index to field dict.

Extract and remove all Gemini models related environment variables, supporting nested fields.

Return image dimensions (width, height) if PNG or JPEG headers are present.

Initialize the configuration. Returns: Config: Configuration object

Remove records older than the specified duration.

Decode base64 file data and save to a temporary file.

Setup loguru logging configuration to unify all project logging output Args: level: Log level diagnose: Whether to enable di

Get database statistics.

Return running status for each client.

Store a conversation model in LMDB. Args: conv: Conversation model to store custom_key: Optional custom key,

Remove ChatML role blocks (<|im_start|>role...<|im_end|>). Role 'tool' blocks are removed entirely; others have markers stripped but content

Verify the provided token against the filename.