Make a single API call.

Reproduce the IMP-14 logic from runner.py lines 505-532.

Save intermediate artifact to stage directory.

Order sections for figure numbering.

Execute a command on the remote host via ssh.

Write checkpoint atomically via temp file + rename to prevent corruption.

Generate requirements.txt in staging dir from auto-detected imports and explicit pip_pre_install, unless one already exists (LLM-generated).

Remove hallucinated citations from paper text. - HALLUCINATED citations: removed from text (recorded in verification report) - SUSPICIOUS/VER

Full quality assessment of a text document.

Return True if the ``opencode`` CLI is installed and callable.

Select optimal benchmarks and baselines from survey results. Context keys: topic (str): Research topic survey (dict):

Generate plotting scripts for all planned figures. Context keys: figures (list[dict]): Figure plan from Planner exper

Return a cleaned BibTeX string with only verified entries. Parameters ---------- bib_text: Original BibTeX string. report:

Check whether a stage requires human-in-the-loop approval.

Return the generic fallback profile.

Return paths to bundled ``.sty`` and ``.bst`` files for this template. Files are stored under ``researchclaw/templates/styles/<name>/``.

Return ASCII-folded last name for citation keys.

Report a metric value. Validates and prints in standard format. Non-finite values (NaN, Inf) are rejected and logged as warnings.

Return first existing config from search order, or explicit path if given.

Run a single Python code string on the remote host.

Search arXiv for papers matching *query*. Parameters ---------- query: Free-text search query. Supports arXiv field syntax

Search Semantic Scholar for papers matching *query*. Parameters ---------- query: Free-text search query. limit: Maxi

Convert all results to a single context string for LLM injection. The output is structured Markdown suitable for prompt injection.

Check *code* parses as valid Python.

Look up a paper by arXiv ID and compare titles. Returns *None* on network failure so that the caller can fall through to the next verificatio

Verify a DOI via CrossRef API, with DataCite fallback for arXiv DOIs. Returns *None* on network failure.

Check that imported modules are available or declared.

R6-4: Check if the current and previous REFINE cycles both produced empty metrics.

Convert a block of Markdown body text to LaTeX.

Detect research domain from topic string.

Detect macOS Apple Silicon GPU (MPS).

Extract abstract text from sections.

Extract a human-readable summary from an artifact file.

Extract meaningful keywords from text, removing stop words. Handles mixed Chinese-English text by also extracting English tokens and

Extract paper title from sections or raw markdown.

Format condition name for display: underscores → spaces, title case.

Format files for inclusion in a prompt.

Get specific guidelines for each figure type.

Infer generation backend from figure type.

Lower-case, strip punctuation, collapse whitespace.

Parse DuckDuckGo HTML results page.

Convert an Exa SDK response into a list of ``SearchResult``.

Convert a single Semantic Scholar JSON entry to a ``Paper``.

Parse ``| a | b | c |`` into ``['a', 'b', 'c']``.

Reconstruct abstract from OpenAlex inverted index format.

Resolve the acpx binary path (cached).

Resolve the model identifier for OpenCode CLI's ``-m`` flag. Azure OpenAI endpoints use the Responses API which many Azure deployments

Reproduce IMP-15 logic from runner.py lines 486-503.

Score a solution node based on execution results.

Send a prompt via acpx and return the response text. For large prompts that would exceed the OS argument-length limit (``E2BIG``), th

Create typical stage-22 and stage-23 output files for testing.

Build user@host string.

Remove markdown code fences from LLM output.

Register all identifiers of *p* in the lookup dicts at *idx*.

Check if the Google Drive mount is accessible.

Check if text passes strict quality gate. Returns (passed, message). Fails if template_ratio > threshold.

Clear cache. Returns number of entries removed.

Count past Stage 10 failures from stage directories and logs. Each stage directory is counted at most once, even if multiple failure indicato

Return the configured rollback target, or the previous stage.

Evaluate a stage output using majority-vote PRM scoring. Args: stage_num: Pipeline stage number (5, 9, 15, or 20). ou

Execute plotting scripts and verify outputs. Context keys: scripts (list[dict]): From CodeGen — each has 'figure_id',

Execute pipeline stages sequentially from `from_stage` to `to_stage` (inclusive).

Execute one pipeline stage, validate outputs, and apply gate logic.

Run OpenCode to generate experiment code. Returns an OpenCodeResult with success status and generated files.

Generate a weekly summary report from completed pipeline runs. Scans ``run_dirs`` for ``pipeline_summary.json`` files and aggregates statisti

Return the style preamble string for injection into chart scripts.

Load all lessons from disk.

Return the most relevant lessons for a stage, weighted by recency. Includes lessons that directly match the stage, plus high-severity

Resolve the effective API base URL for a provider.