Stop and remove the container.

Submit a task for execution. Returns immediately with a Job handle.

Serialize to dictionary.

Validate all preset weight profiles are internally consistent.

Combine built-in adapters with any discovered plugin adapters.

Compute global rank and tier rank. Returns (rank, total, tier_rank, tier_total).

Check if Docker is available and running.

Remove the least-recently-created entry.

Format an error response -- duplicated helper.

Format a response dict -- duplicated helper.

Generate a JSON report.

Check if text contains Chinese/Japanese/Korean characters.

Check if the agent signalled task completion.

Normalize markdown table whitespace so both `| --- |` and `|---|` match.

Incrementally save completed results to enable resume.

Scan tasks/ directory and add any task IDs not already in TASK_ID_TO_DOMAIN.

Trigger frontend rebuild in background with 60s cooldown.

Rebuild the lightweight registry.json index.

Validate a single task directory structure and configuration.

Create an adapter instance by name. Resolves built-in adapters and discovered plugin adapters, applying aliases such as ``oracle`` -> ``dryru

Retrieve a post by ID.

Remove entries. If task_id given, only that task; else all.

Check if the user has admin privileges. Returns: True if user role is 'admin', False otherwise.

Return the names of all available adapters.

Return the canonical adapter key for *name*.

Execute a single task in a sandbox container. Falls back to local execution if Docker is unavailable or ``use_sandbox`` is False.

Build the prompt that instructs the LLM to generate task files.

Call the LLM via OpenAI-compatible API.

Make a single chat completion call with retry on transient errors.

Check that config/models.yaml exists.

Check that skills/curated/ has skill files for all domains. Returns (all_ok, missing_domains).

Compute real 5-dimension scores from per-task results grouped by domain.

Compute foundation score, subject-matter score, and breakdown by subject domain.

Create a new session for an authenticated user.

Return a non-empty model label for run metadata.

Resolve the results output directory.

Auto-detect GitHub username via ``gh api user``.

Handle Taiwan, Hong Kong, Macau — all map to China.

Compute mean score per domain from task results.

Execute a job with concurrency control and rate limiting.

Extract the first ```bash ... ``` code block from *text*.

Extract the absolute workspace path from the injected prompt.

Find existing result file for a clawId.

Generate minimal fallback content for missing files.

Generate 3-condition SkillsBench comparison data for each framework.

Get current process RSS in MB.