Run a multi-file experiment project inside a container.

Sanitize a figure ID for safe use in file paths and Docker names. Strips path separators, dotdot sequences, and shell metacharacters. Returns

Search the web via Exa and return results. Falls back to an empty response when no API key is set.

Submit a project — auto-detects checkpoint and resumes from where it left off. In cross-project discussion mode, each project gets ONE agent.

Convert to legacy baselines/proposed_methods/ablations format. This allows the universal plan to be consumed by existing pipeline cod

Format as context block for injection into code generation prompts.

Verify all BibTeX entries against real academic APIs. Three-layer verification: 1. If entry has ``eprint`` (arXiv ID) → arXiv API lookup

Analyze available experiment data to determine chart potential.

Build command to run Python directly on remote host (with basic sandboxing).

Build command to run inside a Docker container on the remote host. This is the most secure execution mode: code runs in an isolated c

Build domain-specific context for injection into prompts. Includes: - Domain profile hints (file structure, libraries, evaluation)

Check if circuit breaker allows a request.

Verify that data values in the script match source data.

Check Python syntax via ast.parse. Returns list of errors.

Check labels, titles, and legends in the script.

Use LLM to assess visual quality of the chart code.

Convert an ``arxiv.Result`` to our ``Paper`` dataclass.

Detect research domain from topic string and config domains. Returns ``(domain_id, display_name, top_venues)``. Falls back to ``("ml", "machi

Detect section boundaries in paper text.

Extract abstract from paper text.

Extract variable names from assignment targets.

Extract English tokens and translate Chinese domain terms from mixed text.

Extract YAML from text that may contain ACP noise. Strips [thinking] blocks, insight blocks, and other ACP artifacts before looking for YAML

Locate the aider binary. Search order: 1. PATH (shutil.which) 2. Same bin directory as the running Python interpreter

Locate the discussion artifacts directory. Checks run_dir/discussion first, then falls back to shared_results_dir/idea_runs/<idea_id>/discuss

Generate a project-specific YAML config from the default template. If role_prompt is provided (Lab mode), it's prepended to the topic so the

Generate setup.py content for dataset downloading.

Generate image using Gemini REST API (no SDK dependency).

Get or create the shared arxiv Client.

Extract the first-line docstring from a class/function node.

Return True for files that are useful but non-core (eval, test, etc.).

Detect Azure OpenAI from base URL or provider string.

Return True if *candidate* is clearly not a paper title.

Return True if *name* is a timing/meta metric that shouldn't be charted.

Check if the ``opencode`` CLI is available on PATH.

List all existing skill names in the MetaClaw skills directory.

Load hardware_profile.json from a prior stage (usually stage-01).

Return minimal valid config data with metaclaw_bridge overrides.

Return the next table number for the current thread.

Add Obsidian-compatible wikilinks and tag line at end of content.

Convert papers to serializable dicts for caching.

Pause a running project: stop agents and remove queued tasks, but keep all files.

Sort benchmarks by preference: tier 1 > tier 2, knowledge_base > hf, downloads.

Remove BibTeX entries whose keys are in *keys_to_remove*.

Render a markdown image as a LaTeX figure environment.

Resolve config path from args, printing helpful errors on failure. Returns the resolved Path on success, or None if the config cannot be found

Resolve a path for reading. No directory restriction.

Resolve a path for writing. Absolute paths are accepted as-is (no workspace restriction). Relative paths are resolved against the wor

Select top URLs from search results for crawling.

Shorten a metric label for chart readability.

Write setup.sh if setup_script is configured.

Append multiple lessons atomically.

Generate a prompt overlay string for a given stage. Combines two sources: 1. Current-run lessons from ``lessons.jsonl`` (intra-run le

Check whether a paper contains all expected sections. Returns a list of warning strings. Empty list means the paper structure looks complete

Return (ok, message) after testing SSH connectivity.

Return True if value is finite. Log warning and count NaN/Inf.

Detect local GPU hardware and return a HardwareProfile. Detection order: 1. NVIDIA GPU via ``nvidia-smi`` 2. macOS Apple Silicon (MPS) vi

Check stdout/stderr for NaN/Inf/divergence indicators. Returns a description of the issue if detected, None otherwise.

Survey available benchmarks and baselines for the given topic. Context keys: topic (str): Research topic/title hypoth

Generate images for figure decisions marked as 'image' backend. Context keys: image_figures (list[dict]): Decisions from FigureDe

Analyze context and produce figure decisions. Context keys: topic (str): Research topic hypothesis (str): Research hy

Write current prompts (defaults + overrides) to a YAML file.

Extract text from a local PDF file using PyMuPDF.

Format validation issues as a concise error report for LLM repair.

Create from MetaClawBridgeConfig.prm section. Returns None if PRM is not enabled or not configured.

Return the MetaClaw skill config for a given pipeline stage. Falls back to a generic research config if the stage is unknown.

Run the full benchmark selection pipeline. Context keys: topic (str): Research topic/title hypothesis (str): Research