Resolve API key and base URL for embedding requests. Priority: 1. ``OPENROUTER_API_KEY`` → OpenRouter base URL 2. ``OPENAI_API_KEY``

Start desktop control server Args: host: Listening address (127.0.0.1 for local, 0.0.0.0 for all interfaces) port: Liste

Persist all records (bulk).

4-level degrading search for a line pattern inside *lines*. Returns the 0-based index of the first matching position, or -1.

Dynamically switch debug level: 0 = WARNING, 1 = INFO, 2 = DEBUG.

Start screen recording. Args: auto_cleanup: If True, automatically end previous recording if one is in progress

Update agent status

Write (or overwrite) the ``.skill_id`` sidecar in *skill_dir*. Called by ``SkillEvolver`` after FIX / DERIVED / CAPTURED to stamp the new ``s

Add one LLM call's usage to a TokenStats object (caller holds lock).

Create ClientSession (or set up JSON-RPC client) and log success.

Hook called after disconnection. Override in subclasses if needed.

Apply SEARCH/REPLACE blocks to a file on disk.

If prompt is provided, ask user for confirmation, otherwise default to deny

Ask the user whether to install.

Rank candidates using BM25.

Build text for embedding, consistent with MCP search_skills.

Build BaseTool instances from cached tool metadata.

Call LLM with backoff retry on rate limit errors Timeout and retry strategy: - Single call timeout: self.timeout (default 120

Capture screenshot and notify callback.

Check if the command is available. Args: command: The command to check (e.g. "npx", "uvx") Returns:

Lazy wrapper — avoids importing skill_engine at module load time.

Clear screen

Collect all files in skill directory (skip .skill_id sidecar).

Return set of task_ids already completed (for resume).

Compose a content-addressed cache key. The cache stores embeddings under ``"{encoded_skill_id}:{sha256(text)[:16]}"`` so that any cha

Best-effort correction of LLM-hallucinated skill IDs. LLMs frequently garble the hex suffix of generated IDs (e.g. swap ``cb`` → ``bc``). Fo

Create Anthropic client (only supports anthropic provider).

Download a cloud skill and register it locally.

Download a single file from URL to dest path, with retries. Strategy order: 1. ``curl`` — bypasses Python SSL stack entirely (most robust)

Drop older content-addressed entries for a skill. Caller must hold ``self._cache_lock``.

Import livebench modules if available.

Extract a JSON object from LLM response text. Handles markdown code fences and bare JSON.

Find parquet file — handles both file path and directory.

Generate actionable recommendations based on quality data.

Get the cache file path for the current model.

Get structured cache key (backend, server, tool_name) from tool.

Generate platform-specific conda activation prefix.

Get embedding from structured cache.

Generate install command based on command type. Args: command: The command to execute (e.g. "npx", "uvx", "uv")

Merge OpenClaw top-level env vars with skill-level env overrides.

Get or create aiohttp session.

Get or create aiohttp session.

Lazy-init shared ``SkillRanker`` (with persistent embedding cache).

Get path for a test file

Infer backend from tool name when tool_results lack backend.

Insert an execution_analyses row + its skill_judgments. Called within a transaction holding ``self._mu``. Returns: int:

Terminate a timed-out subprocess and its descendants when possible.

Ask LLM to confirm whether a rule-based evolution candidate truly needs evolution. Returns True if LLM agrees, False otherwise.

Load from GDPVal parquet (has full prompts + reference files).

Load and parse the OpenClaw config file. Returns None on failure.

Create a OpenSpaceConfig for one worker.

Create a tuned SQLite connection. Write connection: ``check_same_thread=False`` for cross-thread usage via ``asyncio.to_thread()``.

Strip tzinfo so naive/aware datetimes can be compared safely.

Normalize message history only when a provider requires it.

Called by _TokenLoggerCallback after every successful LLM call.

Try to parse *data* into *model_cls* (a subclass of BaseModel). If `model_cls` is None or not a subclass of BaseModel, return the o

Parse prompt data into Prompt objects.