Run the full figure generation pipeline. Context keys: experiment_results (dict): Parsed results.json condition_summa

Write full session snapshot to stage directory.

Record effectiveness of all active skills for a completed stage.

Resolve the stage to resume from using checkpoint metadata.

Resume a project from its last checkpoint.

Run a multi-file experiment project on the remote host.

Run a multi-file experiment project in the sandbox. Copies all ``.py`` files from *project_dir* into the sandbox work directory and e

Search Google Scholar for papers matching query.

Run the full search + extraction pipeline. Parameters ---------- topic: Research topic string. search_que

Run multiple queries and return deduplicated union. Stops early when *max_time_sec* is exceeded or two consecutive queries return zero result

Return True if approaching 80% of time budget.

Raised-cosine window to reduce hard on/off transitions.

Serialize for caching.

Serialize to a JSON-safe dict.

Serialize to a JSON-safe dict.

Convert to researchclaw.literature.models.Paper.

Serialize to YAML string.

Analyze a repository by reading key files.

Append one line to the streaming text log (tail -f friendly).

Add default figures to meet minimum count.

Build the adapter registry with lazy imports for domain adapters.

Build the aider CLI command for a single invocation.

Build list of directories allowed for read operations.

Convert tool specs to OpenAI-compatible tool definitions.

Build the assistant message to append to conversation history.

Build mapping from author-year patterns to cite_keys. Returns dict like: "Raissi et al., 2019" → "raissi2019physicsinformed

Extract meaningful search queries from a long topic string. Handles mixed Chinese-English topics by translating Chinese domain terms and extr

Build the opencode.json configuration.

Build a role prompt for a Lab mode agent. Uses predefined prompts for known angles, otherwise generates a reasonable prompt from the angle name.

Build tool descriptions for system prompt (text-based mode). Used when the proxy doesn't reliably translate structured tool calls (e.

Build tool descriptions for system prompt (text-based mode). Used when the proxy doesn't reliably translate structured tool calls (e.

Call the LLM API with tool definitions (OpenAI-compatible).

Call with exponential backoff retry.

P7: Check if ablation results are within *threshold* of baseline. Returns a list of warning strings for ineffective ablations.

Use LLM to assess relevance of each citation to the research topic. Returns a dict mapping cite_key → relevance score (0.0–1.0). Processes ci

Quality gate before forced PROCEED. Returns (ok, message). ok=False means experiment results have critical quality issues and the forced-PROC

Check for classes with identical method bodies.

Quick connectivity check for literature API endpoints (parallel).

Determine if the smoke test passed. Success criteria (in order): 1. LLM final text explicitly declares the test passed → True

Run `python3 -c 'import main'` in the workspace; return error or None.

Classify user chat intent: LLM first, keyword fallback on failure.

Fast keyword-based fallback for intent classification.

Collect actual experiment parameters and results for peer review.

Aggregate experiment metrics from runs/ directory across prior stages. Returns a dict with ``runs``, ``metrics_summary``, ``best_run``, ``lat

Collect JSON context from a directory, with size limits. Large fields like ``stderr`` and ``stdout`` are stripped to avoid exceeding LLM toke

Read result.json from the done task directory and clean up. The result.json schema (written by the Colab worker): {"returncode": in

Collect table lines (header + separator + body rows).

Collect code and config files from workspace.

Copy fixed .py files back to the original experiment directory.