Compress old events into a structured summary. Args: hard: if True, keep fewer tail events (context overflow recovery). R

Create a conda environment when requested and missing.

Create a zip snapshot. Returns the snapshot id.

Download a PDF from URL and extract its full text. Args: pdf_url: URL to the PDF file. max_pages: Maximum pages to process.

Force state to a specific stage, bypassing transition validation. BUG-21 fix: used during checkpoint resume where we know the stage w

Render a LaTeX document from a Jinja2 template. Args: template_name: Template file name (e.g. "paper.tex.j2"). data: Template var

Call the LLM with an image attachment (vision).

Return the specialized system prompt for reviewing a section.

Inspect local project manifests in the same priority order as local execution.

Whether the unified pipeline should prefer SLURM/cluster execution.

Return the ordered list of stages that do actual work.

Return the ordered working stages for the selected pipeline mode.

Rollback to the most recent snapshot. Returns snapshot id or None.

Find a single paper by title match. Uses filter-based search which is more precise than full-text search and does NOT count against the 1,000

Return a full cost summary across all stages.

Serialise all tools to the OpenAI ``tools`` parameter format. Each entry has the shape:: {"type": "function", "function": {"name

Serialize for inclusion in writing output metadata.

Analyze cloned repos to understand their structure and key components.

Use LLM to interpret and summarize experiment results.

Analyze training_log for convergence, overfitting, stability.

r"""Prefer near-text, compact float defaults before PDF compilation. This is a precompile layout guard, not a post-PDF repair. It makes

Apply a single patch to a file. Returns True if successful.

Approximate two-tailed p-value. Uses normal approximation for df > 30. For df <= 30, uses a conservative lookup table with df=10 critical val

Validate that paper-facing result artifacts satisfy the blueprint contract.

Backfill downloadable dataset entries directly from the blueprint.

Try to repair environment creation after venv failure. Strategies (tried in order): 1. If conda is available, create a fresh conda en

Format baseline->cite key mapping block for experiments context.

Flatten an experiment blueprint dict into a set of search tokens.

Build a deterministic LaTeX ablation table from structured data.

Create one agent per pipeline stage.

Build mapping from BibTeX cite key -> paper dict (with title/abstract). Matches BibTeX entries in the paper to papers from ideation b

Build a chart-specific prompt from the chart type and research context.

Build a mapping: paper_index → cite_key (authorYear format).

Compress round history to ~1 line per round for LLM context.

Extract shared primitives once; return a dict consumed by section builders. This is called ONCE in run(), and the resulting dict is passed to

Build the system prompt for diagnosis (extracted for readability).

Format extracted evidence for inclusion in the planning prompt.

Build evidence context block for writing prompts.

Compact execution context used for iterative repair.

Pre-build LaTeX figure/table blocks to embed inline.

Build full context string (legacy, used as fallback). Prefer _build_section_context().

Walk *root* and index every directory containing a SKILL.md.

Build a deterministic LaTeX main-results table from structured data. Returns empty string if data is insufficient.

Convert pipeline data into (baselines, proposed, metrics) for comparison_matrix. Returns None if insufficient data.

Build chat messages for the RAM model.

Build a grounding context block for revision prompts. Tells the revision LLM which numbers are real and must be preserved.

Build a tailored context string for a specific section.

Register all paper editing tools.

Build a ToolRegistry with search tools for writing. Returns None if no tools could be registered (missing deps).

Scan LaTeX text for common AI-writing artifacts. Returns a list of ConsistencyIssue for each detected problem.

Check citation coverage: total count and must-cite enforcement.

Return an error message if command tries to use a disallowed conda env. Returns None if the command is safe. ``allowed_envs=None`` disables i

Post-generation consistency check across all sections.

Scan all generated files for cross-file import mismatches via AST. Checks two patterns: 1. `from X import Y` where Y doesn't exist in

Evaluate heuristic termination conditions. Returns {"should_continue": bool, "reason": str | None}.

Classify how complete the experiment results are.

Classify error as ('data_missing', path) or ('code_bug', '').

Clean LaTeX equation text for SymPy parsing.

Delete stale generated figure images before a fresh FIGURE_GEN run.

Remove BibTeX entries that are not cited anywhere in the LaTeX source. This prevents the .bib file from accumulating unused entries c

Shallow-clone a GitHub dataset repo. Returns True on success.