Generate LaTeX TikZ/PGFPlots code for a figure. This produces code that compiles directly in a LaTeX document that includes ``\\usepa

Generate a plotting script using LLM.

Ask LLM to review generated code for common pitfalls.

Ask LLM to suggest benchmarks and baselines when APIs unavailable.

Load and cache the seminal papers list.

Load results.json from the most recent experiment run stage.

Load pending human feedback from ``run_dir/human_feedback.jsonl``. Reads all feedback entries, formats them chronologically, and marks them a

Load the local benchmark knowledge base.

Load experiment plan summary from prior stages.

Load a single YAML profile into a DomainProfile.

Generate bullet-point text of approximately *word_count* words.

Generate related-work style prose with comparative language.

Create HTTP request handler for file downloads.

Generate results prose with statistical measures.

Flatten a chat-messages list into a single text prompt. Preserves role labels so the agent can distinguish context.

MetaClaw bridge: post-pipeline hook. 1. Convert high-severity lessons into MetaClaw skills. 2. Record skill effectiveness feedback. 3. Si

Create a mock GitHub client.

Map upstream domain profiles to benchmark-knowledge domains.

Map common parameter name aliases to canonical names.

Map common parameter name aliases to canonical names.

Handle S8 completion — wait for all agents, then pick the best hypothesis and create only ONE downstream task to avoid duplicate experiments.

Handle idea factory run completion: extract hypotheses, push to idea pool + L2 queue.

Handle idea factory S7-only completion: enter discussion flow.

Parse blueprint YAML into a structured dict.

Parse CSV results (one row per condition/seed/metric).

Extract PROCEED/PIVOT/REFINE from decision text. Looks for the first standalone keyword on its own line after a ``## Decision`` heading. Fal

Parse Python traceback to find failing file and line. Returns (filename, line_number, error_message) or None.

Parse a single Python file and extract its API surface.

Parse structured JSON results.

Best-effort JSON extraction from LLM response. BUG-17: Always returns ``dict | None`` — never a bare string or list, which would caus

Extract the Justification/Rationale section from LLM decision text.

Parse pdflatex log output for errors and warnings.

Convert a single OpenAlex work JSON to a ``Paper``.

Extract text content and tool_use calls from API response. Falls back to parsing JSON tool calls from text content when the structure

For passthrough layers (e.g. coding): read existing artifacts, mark done immediately.

Generate a deep implementation blueprint. Returns (raw_yaml_str, parsed_blueprint_dict_or_None).

Generate code in single shot, then iteratively fix via sandbox.

Generate files one-by-one following blueprint dependency order.

Explore multiple candidate solutions via tree search.

Reviewer agent examines code; coder fixes critical issues.

Check if a discussion subprocess has finished and handle completion.

POST *url* with exponential back-off retries.

Prepare a workspace for Aider-based sanity fix. Copies experiment .py files, reuses GUIDANCE.md and EXPERIMENT_PLAN.yaml from the Sta

Create a prior run that timed out with no metrics.

Create workspace with experiment code + symlinks to data.

Create a clean workspace with symlinks to data directories.

Clean up common LLM artifacts before parsing. 1. Strip outer fenced code blocks (e.g. triple-backtick markdown) that LLMs around the entir

Lightweight pointer to workspace instruction files. CODEGEN.md is generated directly in the workspace by _prepare_workspace. We only add a br

Sort baselines: knowledge_base first, fewer deps preferred.

Read ``figure_manifest.json`` from stage-16/charts/ (or stage-14). Returns ``{filename: metadata_dict}`` for each chart entry.

Read how many PIVOT/REFINE decisions have been made so far.

Append a decision event to the history log.

Remove \\cite{key} and [key] references for specified citation keys.

Render a fenced code block as a LaTeX environment. IMP-28: Detects pseudocode blocks (language hint 'algorithm' / 'pseudocode', or 3+ algorit

Render a single figure script.

Ask LLM to fix critical validation issues.

GET *url* with exponential back-off retries.

GET *url* with exponential back-off retries.

Resolve to coding model if configured.

Get the LLM config (may have coding_model override already applied).

Resolve LLM-selected names back to full benchmark/baseline dicts.

Resolve a user-provided reference string to a candidate dict. Accepts arXiv IDs, arXiv URLs, local PDF paths, or plain titles. Attempts arXiv

Restart a project from scratch: stop agents, clear checkpoint, re-submit.

Multi-dimensional LLM review of compiled paper (AI-Scientist style). Scores the paper on 7 academic review dimensions (1-10 each), identifies

Review a single rendered figure on three dimensions.