Run a shell command WITHOUT proxy (for domestic sources like ModelScope).

Sanitize revised LaTeX using WritingAgent's sanitizer. Falls back to inline critical fixes if WritingAgent import fails.

Save PNG (if not already saved) + convert to PDF + register artifacts. Args: code_generated: True for matplotlib/code-generated c

r"""Place a figure block at the best position in the document. Strategy (matches top-venue conventions): 1. Near first \ref{fig:label

Remove LLM filler words/phrases at the start of generated section content.

Remove ``<think>…</think>`` blocks emitted by reasoning models.

Export a clean, self-contained output folder. Structure: {topic_slug}/ +-- paper.pdf +-- paper.tex

Return the specialized system prompt for a writing section.

Return True when a command token refers to a Python launcher.

Search OpenAlex for scholarly works. Uses the /works endpoint with search parameter. Args: query: Search query string. max_r

Search the web via DuckDuckGo. Args: query: Search query string. max_results: Maximum number of results to return. Returns:

Return a JSON-serializable snapshot for manifest storage. Strips all API keys (global and per-stage) to prevent accidental leaks.

Parse traceback, fix each affected file with a targeted LLM call. Surgical approach: for each file in the traceback, send ONLY that file

Build a brief context block informing the LLM about evidence completeness.

Quick structural checks for LaTeX source (no compilation). Returns a list of issue description strings. Used by the backpressure mec

Check if circuit breaker is open. Returns True if we should skip the call.

Record a response status for circuit breaker logic.

Parse metrics.json after a successful quick-eval run.

Build common query params (api_key).

Decode subprocess output bytes to str safely on Windows (GBK fallback).

Fallback figure plan driven by available evidence.

Extract meaningful keywords from text (first *max_lines* lines).

Return ``"conda"`` if conda is installed, else ``None``.

r"""Fix 'N components: (1)...(2)...' where N doesn't match actual count. Detects patterns like 'four key components' followed by enumerated i

Get or create an OpenAI client for the given endpoint + timeout.

Check if an exception is transient and worth retrying.

Coerce raw execution/analysis metrics into the main_results schema.

Normalize S2 paper data to a common format.

Normalize real result rows without inventing metrics.

Try to extract metrics from training log output.

Run automated consistency checks on the LaTeX source.

Run a shell command asynchronously with proxy environment.

Save iteration state checkpoint for crash recovery.

Scan a possibly truncated JSON fragment.

Choose the best available pip install strategy for a project.

Run full analysis: heuristics + LLM attribution.

Build a centralized execution/remediation policy for a project.

Compile a .tex file to PDF using tectonic (preferred) or pdflatex. Args: tex_path: Path to the .tex file. output_dir: Output dire

Scan the system for all available Python environments. Returns a list of dicts, each with: - name: human-readable label (e.g. "conda:

Run a ReAct loop: let the LLM call tools until it produces text.

Return list of ``{"id": ..., "time": ..., "path": ...}``.

Read all code files in the directory, return list of {path, content}.

Search arXiv for papers matching the query. Args: query: Search query string (supports arXiv query syntax). max_results: Maximum

Remove review-introduced quantitative claims unsupported by run artifacts.

Build a compact narrative summary from execution analysis artifacts.

Build context block from real experiment results for writing prompts.

Check that claims in the paper match the experiment blueprint. Detects: - Metrics mentioned in the paper but not defined in the bluep

Check that figure references match figure definitions.

Shorten an AI-image caption that is actually a generation prompt. If the caption is <= MAX_ACADEMIC_CAPTION_LEN characters it is returned as-

Count pages in a PDF file. Returns None if unavailable.

Remove duplicate consistency issues by (issue_type, description) key.

Fallback figure plan for survey papers.

Extract pip-installable dependencies from environment.yml.

Extract (heading, content, level) tuples from LaTeX source. Handles \\section{} (level=0), \\subsection{} (level=1), and \\subsubsect

Extract token usage dict from an OpenAI response object.

Return the first supported legacy experiment entry script.

Convert [{metric_name: str, value: float}, ...] to {name: value}. Filters out NaN/Inf values.

Generate an image with image2 first and audited compatible fallbacks. Returns (base64_png, used_config, metadata). Metadata is deliberately

Get the full content of a top-level section including its subsections. Merges subsection content back into the parent section so the reviewer

Lazy import OpenAlex search (returns None if module unavailable).

r"""Insert *figure_block* after the paragraph that first references *fig_key*. Searches for \ref{fig:KEY}, \autoref{fig:KEY}, \cref{fig:KEY}.