Round excessively precise metric values (>4 decimal places). Uses significant-figure-aware rounding so small values like learning rates (e.g.

Write files to a temp directory and run in sandbox.

Persist project metadata so it can be recovered on restart.

Save a versioned snapshot of .py files after each write/edit. Snapshots go to workspace/.snapshots/ for post-hoc debugging — you can

Search using Tavily API (installed SDK).

Extract a human-readable signature from a function AST node.

Turn arbitrary text into a filesystem-safe slug.

Send a streaming request and reassemble into a standard response dict.

Remove markdown code fences if present.

Submit an experiment task and wait for the Colab worker to complete it. Protocol: 1. Copy experiment files to ``pending/<task_id>/`

Mark all stages from layer_range[0] to done_up_to as completed, emit events for new ones.

Synthesize multiple perspective outputs into a unified result.

Launch a discussion between two agents from different projects.

Launch the discussion runner when all agents in a group have completed S7.

Write a single skill to disk as a SKILL.md file.

Return cache statistics.

Check permission. Returns None if allowed, error string if denied. Write and read restrictions are relaxed — only dangerous bash comm

Detect common API misuse patterns. Catches: - np.erf() (should be scipy.special.erf) - nn.Linear/nn.Conv2d inside forward() (unregistered

Analyze class implementations across all experiment files. Detects: - Empty or trivial class inheritance (class B(A): pass) - Classes wit

Detect common variable scoping bugs in experiment code. Catches the pattern where a variable is defined inside an if-branch but used outside

Remove all cache files. Return count of files deleted.

Compile *tex_path* with pdflatex, auto-fixing common errors. Parameters ---------- tex_path: Path to the ``.tex`` file. Must be

Compute success rate per skill across all recorded runs. Returns: Dict mapping skill_name to {total, successes, success_rate}.

Synchronous crawl — tries Crawl4AI via asyncio.run, falls back to urllib.

Quick keyword-only detection that returns a domain_id string. Useful for lightweight checks where a full profile isn't needed.

Validate all generated code artifacts. Context keys: acquisition (dict): Output from AcquirerAgent

Generate a figure plan from experiment results. Context keys: experiment_results (dict): Parsed results.json / experiment_summary

Generate figure manifest and markdown references. Context keys: rendered (list[dict]): Successfully rendered figures with

Download a PDF from URL and extract text.

Extract code patterns from reference snippets. Parameters ---------- code_snippets : list[str] Code content from GitHub repos.

Write results.json with all reported metrics and partial results.

Write results.json with all reported metrics and partial results.

Format writing tips as a prompt-injectable string. Parameters ---------- categories: Subset of tip categories to include. If *Non

Deserialize from cache.

Generate search queries for GitHub code search. If no LLM is provided, generates queries from topic keywords and library names using heuristi

Return list of chart filenames from manifest.

Get the content of a file from a repository. Parameters ---------- repo_full_name : str Repository in "owner/repo

Return HTTP headers for MetaClaw proxy requests. Args: stage_name: Current pipeline stage (for logging/tracking). Return

Return deduplicated list of canonical template names.

Log an error with optional exception details.