Copy results and modified files back.

Create a per-agent config file with a different primary_model.

Minimal data path pointers — detailed info is in CODEGEN.md.

IMP-30: Remove duplicate tables that share the same header row. LLMs sometimes repeat tables (e.g. same results table in Results and Discussi

Delete a project: stop any running processes, clean up state, remove files.

Detect Huawei Ascend NPU via npu-smi.

P10: Detect contradictions in experiment results before paper writing. Returns a list of advisory strings to inject into paper writing prompt.

Reconstruct Paper objects from cached dicts.

Quick hash of file mtimes to detect changes.

Pre-discover filesystem context before prompt building. Analogous to claw-code's ``ProjectContext.discover_with_git()`` which reads git statu

Return True if Docker daemon is reachable.

Auto-install missing dependencies from experiment code.

Expand search queries for broader literature coverage. Generates additional queries by extracting key phrases from the topic and creating foc

Extract rationale from the most recent decision_structured.json. Supports multiple field formats: - ``rationale`` or ``reason`` key (direct)

Extract full text from a local PDF for downstream stages (S9, S11). Uses the PDFExtractor utility which reads all pages. Falls back to the s

Parse LLM response containing multiple files with filename markers. Expected format:: ```filename:main.py import model .

Analyze the experiment plan and generate targeted technical hints. This replaces hardcoded advice: hints are only generated for technologies

Extract fine-grained lessons from experiment run artifacts.

Extract Python code from LLM response for a single file.

Extract meaningful keywords from the research topic + domain list. Returns lowercased keyword list (2+ chars, no stop words). Used by the dom

Fill bar comparison template with condition data.

Fill grouped bar template with multi-metric data.

Fill heatmap template — rows=conditions, cols=metrics.

Fill a template with actual data values.

Filter baselines by pip availability.

Find the acpx binary — check PATH, then OpenClaw's plugin directory.

Find the latest experiment/ directory from prior stages.

Close the stale session and reset so _ensure_session creates a new one.

Human-readable age string.

Ask LLM to generate baseline method instantiation code.

Ask LLM to generate a robust data loading function.

Generate a text-to-image prompt for a methodology framework diagram. Reads the paper's method section and produces a detailed prompt suitable

Generate image via the best available backend. Priority: OpenAI images API → OpenAI chat proxy → google-genai SDK → Gemini REST API.

Generate a complete LaTeX package (zip) ready for Overleaf upload. Uses LLM to convert markdown paper to LaTeX, then collects figures, refere

Generate a NeurIPS-style paper checklist appendix in markdown. This checklist is based on the NeurIPS 2025 submission requirements. It is app

Use LLM to generate a detailed figure plan.

Generate a plotting script for a single figure.

Generate image via an OpenAI-compatible chat completions proxy (fallback). The proxy (e.g. vectorengine.ai) forwards to a Gemini image model

Generate image via the OpenAI-compatible /images/generations endpoint. Uses the standard OpenAI images API which returns b64_json or url.

Generate image using google-genai SDK.

Lazily create a single sandbox instance for all validation runs.

Run AST-based checks and classify as CRITICAL or WARNING. Returns (critical_issues, warning_issues).

Lazily create a lightweight LLM client for intent classification.

Install OpenCode globally via npm. Returns True on success.

Run ``opencode run`` in the workspace. Returns (success, log, elapsed).

Check if ``npm`` is available on PATH.

Check if a blueprint has the minimum required structure.

Check if num matches any verified value within 1% relative tolerance.

Yield candidate paths for the experiment plan, best match first.

Reject ambiguous keyword matches that commonly leak across domains.

Quick list of workspace files (for trace logging).

Ask LLM to analyze paper and decide on figures.

Extract patterns using LLM analysis.

Generate queries using LLM.