Outcome of executing a single stage.

Central registry for pipeline prompts with optional YAML overrides.

Verification result for one BibTeX entry.

Collection of prompt blocks for a specific pipeline stage. Each field is a string block that gets injected into the prompt template. Empty st

Minimal mock LLM client conforming to _LLMClientLike.

A single experimental condition (method, configuration, etc.).

Fake LLM that returns preconfigured responses.

Configuration for Docker-based experiment sandbox.

Execute experiment code inside a Docker container. Same public API as :class:`ExperimentSandbox` so the pipeline can use either backend trans

Complete description of a research domain's experiment conventions. Loaded from YAML files in ``researchclaw/domains/profiles/``.

A single validation finding.

Multi-phase code generation agent. Parameters ---------- llm : LLMClient The language model client to use for code generation.

Configuration for the advanced multi-phase code generation agent.

A paper author.

Configuration for the LLM client.

Parsed response from the LLM API.

A single lesson extracted from a pipeline run.

A single web search result.

Aggregated validation result for a code snippet.

Result of crawling a single URL.

Stateless OpenAI-compatible chat completion client.

JSONL-backed store for pipeline lessons.

Fake LLM client that returns configurable responses.

Extracted patterns from reference code.

Configuration for Google Drive-based async Colab execution.

Manages OpenCode CLI invocations for beast mode code generation.

A single paper from Semantic Scholar, arXiv, or similar sources. Fields are designed to hold the union of metadata available from both Semant

Aggregate report for all citations in a paper.

Unified experiment results container. Works for all domains — ML scalar metrics, physics convergence data, economics regression tables, etc.

A single knowledge-base entry to be written.

Extracted content from a PDF file.

Disk-based cache for code search results. Cache structure:: code_search_cache/ {domain_id}/ {topic_hash}.json

Combined result from all web search sources.

Minimal mock LLM client.

Reviews rendered charts for accuracy and quality.

Web search client backed by the Exa AI Search API. Uses the ``exa-py`` SDK. When no ``EXA_API_KEY`` is set the client is inert and ``search(

Analyzes experiment data and generates a figure plan.

One record of a skill's effectiveness in a pipeline stage.

Execute experiment code on a remote machine via SSH. Same public API as :class:`ExperimentSandbox` and :class:`DockerSandbox` so the pipeline

Searches local knowledge base and HuggingFace Hub for benchmarks.

Response from a web search query.

Detected hardware capabilities of the local machine.

Domain-agnostic experiment plan. This can represent ML train-eval, physics convergence studies, economics regression tables, and any other pa

Validates generated code artifacts for syntax and API correctness.

LaTeX template specification for one conference.

Immutable experiment infrastructure for time and metric management.

Decides what figures are needed and which backend generates them. This agent analyzes the paper context (topic, draft, experiment data) and p

Generic adapter for domains without a specialized adapter. Uses the DomainProfile's guidance fields (loaded from YAML) to construct prompt bl

Result from an OpenCode invocation.

Executes plotting scripts and verifies output files. Supports two execution modes: 1. **Docker sandbox** (preferred): Runs scripts inside a

One candidate solution in the search tree.

Output from a single agent step.

Generates visualization code (Python or LaTeX) for each planned figure. Supports two output formats: - ``"python"`` (default): Matplotlib/S

Complete result from a code search operation.

Fake sandbox for testing.

Determines figure placement and generates paper references.

Specification of a metric to evaluate.

Summary of a GitHub repository.

Filters and ranks datasets/baselines based on constraints.

JSONL-backed store for skill effectiveness records.

Generates data loading, baseline, and download code.

Configuration for the advanced code generation agent. All phases are independently toggleable. The default profile enables Phases 1 (bluepri

Result returned by a strategy's generate() call. Analogous to claw-code's TurnSummary returned by ConversationRuntime.run_turn().

Evaluation protocol for the experiment.

Configuration for the FigureAgent multi-agent system.

GitHub REST API client with rate limiting and caching. Uses ``GITHUB_TOKEN`` env var for authentication (strongly recommended). Falls back to

Google Scholar search client using the ``scholarly`` library. Parameters ---------- inter_request_delay: Seconds between requests

Manages a MetaClaw session spanning an AutoResearchClaw pipeline run.

Generates conceptual/architectural figures using Gemini image generation. This agent uses the Gemini API (Nano Banana) to create publication-qual

PDF text extraction using PyMuPDF. Parameters ---------- max_pages: Maximum pages to extract (0 = all). extract_sections:

Fully rendered prompt ready for ``llm.chat()``.

A paper result from Google Scholar.

Web page → Markdown crawler powered by Crawl4AI. Parameters ---------- timeout: Request timeout in seconds. max_content_lengt

Orchestrates all web search and content extraction capabilities. Parameters ---------- tavily_api_key: Tavily API key (optional,

General-purpose web search client. Uses Tavily (installed) as primary engine. Falls back to DuckDuckGo HTML scraping only if no Tavily API ke

Configuration for ACP agent connection.

Generic agentic turn loop: LLM iteratively calls tools. Parameters ---------- verification_hooks : list[VerificationHook] Called

Execute experiments asynchronously via Google Drive + Colab worker. Same public API as ExperimentSandbox/DockerSandbox/SshRemoteSandbox.

ML adapter: returns empty blocks so the existing prompts.py behavior is used verbatim. This is the zero-regression guarantee.