r"""Ensure exactly one \end{document} at the very end, with \bibliographystyle and \bibliography just before it. LLMs sometimes emit

Scan all generated files for cross-file import mismatches and fix them via LLM.

Ask LLM to fix cross-file import mismatches via search-replace patches.

Fix LaTeX compilation errors using a 2-level strategy. Level 1: Deterministic fixes (no LLM) — via shared latex.fixer module. Level 2

Fix LaTeX compilation errors using a 2-level strategy. Level 1: Deterministic fixes (no LLM) -- via shared latex_fixer module. Level

r"""Inject \small, \tabcolsep, and @{} into tables that lack them.

Try to auto-format generated code with black. Silently skips on failure.

Generate a single figure via the configured AI image model. Flow: generate prompt, call the configured image API with retries, and if all fai

Generate and run a smoke test that imports every generated module.

Have LLM generate plotting code, then execute it to create the chart. Retries up to MAX_CODE_CHART_RETRIES times, feeding error messages

Generate a lightweight conda environment file from the code plan.

Phase 2: Generate a single file via Codex.

Generate using HuggingFace Transformers.

Generate a result figure with the image model when chart code fails. The prompt writer is explicitly constrained to use only numbers already

Generate images via Gemini native API.

Generate images via OpenAI /v1/images/generations (DALL-E).

Generate Method through compact LLM calls instead of one oversized prompt.

Generate matplotlib plotting code for a specific figure.

Phase 1: Generate the project plan JSON via Codex.

Generate requirements.txt from code plan.

Generate a SLURM batch script for training.

Generate using vLLM HTTP API.

Initial results dict. Override to add pipeline_mode, etc.

Create a Jinja2 environment for LaTeX rendering.

Return the ordered list of stages to execute.

Return section-specific revision guidance for top-tier venue standards.

Search GitHub repos via API.

Clone a GitHub dataset repo, extract real download URLs, and fetch data.

Handle returncode == -1 (timeout/killed) by checking for partial artifacts. Returns a quick-eval result dict if recovery succeeded, or None t

Check if any extracted metric name matches a common metric substring.

Search ModelScope for a matching model (async, non-blocking). Uses ModelScope API to find if the model exists, rather than relying on

Shared initialisation called from FeishuBot.__init__.

Inject missing must-cite papers into the Related Work section. Adds a brief sentence for each must-cite paper that was identified but

Place method schematic after the Method text introduces the pipeline.

Install PyTorch with CUDA via conda into a named env. Uses ``conda install pytorch torchvision torchaudio pytorch-cuda=XX.X -c pytorc

Discover environments and prompt the user to pick one. Returns the selected python path, or None if user skips / no envs. Also saves

Return a match object if *url* points to a GitHub repo (not a raw file).

Get JSON parse error message for diagnostics.

Keep one method schematic and reserve remaining slots for result charts.

r"""Keep main-text experiment figures from overwhelming prose. Extra result-like figures are removed from the main text after the evi

Send image to LLM; get back trim decision + code. Uses figure_code stage (vision-capable model like Claude Sonnet), NOT figure_gen (c

Ask LLM to attribute the results and recommend next action.

Ask LLM to read the README and extract the actual download command.

Send trimmed image to LLM for visual verification. Uses figure_code stage (vision-capable model like Claude Sonnet), NOT figure_gen (

从环境变量或 config.json 加载飞书凭证。

Best-effort reload of local round artifacts from disk.

Log a citation quality report.

Heuristic: does the error look like an infrastructure problem?

Fuzzy-match baseline names from blueprint against paper titles. Returns a mapping: baseline_name → cite_key. This prevents the LLM fr

Merge section reviews from multiple reviewers. Strategy: per-section min score (strictest reviewer wins), union all issues/suggestion

Diagnose why a revision degraded quality, then retry with extra constraints. This is the runtime prompt self-optimization loop: 1. LL

Return a stable method name from the schema's structured method dict.

Return a metric value from a normalized result entry.

Run parallel reviews from multiple model personas, merge results. Falls back to single-model review if all reviewers fail.

Repair legacy manifests in-memory before validation.

Convert alternative metrics.json formats to the expected schema. Handles the common case where generated code writes: {"variants":

Normalize machine-generated artifacts into stable paper-facing fields. This function never fabricates missing values. Published baselines fro

Apply local-only endpoint routing from ~/.nanoresearch/private_endpoints.json. This file is intentionally outside the repository and is never