Get cached result if it exists and is not expired.

Search the web for a query.

Append a timestamped log entry and auto-persist for live debugging.

Parse experiment results from a run directory. Tries formats in order: JSON → CSV → stdout regex.

Log to both Python logger and the internal validation log.

Return a fully rendered prompt for *stage* with variables filled. If *evolution_overlay* is provided, it is appended to the user prompt

Match text against keyword rules. Returns domain_id or None.

Parse JSON from text, handling noisy ACP output. Tries multiple strategies: direct parse, markdown fence extraction, balanced brace matching

Send a chat request with optional retry on timeout/transient errors. Parameters ---------- retries: Number of extra attempts afte

Get a specific domain profile by ID.

Render a reusable prompt block.

Send a chat completion request with retry and fallback. Args: messages: List of {role, content} dicts. model: Overrid

Convert inline Markdown formatting to LaTeX. Preserves: - Inline math ``\\(...\\)`` and ``$...$`` - ``\\cite{...}`` references - Disp

Append a single lesson to the store.

Replace ``{var_name}`` placeholders with *variables* values. Only bare ``{word_chars}`` tokens are substituted — JSON schema examples like ``

Return True if *name* looks like a metric name rather than a log line. Used to filter stdout lines when parsing ``name: value`` metric output.

Load evolution lessons + MetaClaw skills for prompt injection. Combines intra-run lessons (from current run's evolution dir) with cross-run a

Return total number of stored lessons.

Get the appropriate PromptAdapter for a given domain. Lookup order: 1. Exact domain_id match 2. Prefix match (e.g., "ml_" for all ML doma

Split Markdown into a flat list of sections by heading.

Run a single Python code string inside a container.

Close the acpx session.

Execute all enabled phases and return generated files.

Compute the next state given current stage, status, and event. Raises ValueError on unsupported transitions.

Look up a conference template by name. Raises ``KeyError`` if *name* is not in the registry. Accepts both full names (``"neurips_2024"``) and

Convert a Markdown paper to a complete LaTeX document. Parameters ---------- paper_md: Full paper in Markdown with embedded LaTeX

Quick connectivity check - one minimal chat call. Returns (success, message). Distinguishes: 401 (bad key), 403 (model forbidden),

Validate *entry_point* syntax (no filesystem access needed). Returns an error message if invalid, ``None`` if valid. Call this **before** cop

Scan runs/projects/ and return status info for all projects.

Store a result in the cache.

Return a rendered sub-prompt (e.g. code_repair).

Track cumulative LLM usage.

Generate flowing prose text of approximately *word_count* words.

Return the appropriate sandbox backend for *config.mode*. - ``"sandbox"`` → :class:`ExperimentSandbox` (subprocess) - ``"docker"`` → :class:

Detect the research domain from topic and context. Three-level detection: 1. Keyword matching (fast, deterministic) 2. LLM classification

Scan text for template/placeholder patterns. Returns list of TemplateMatch objects for each detected pattern.

Extract lessons from a list of StageResult objects. Detects: - Failed stages → error lesson - Blocked stages → pipeline lesson - Deci

Score the complexity of an experiment to determine if beast mode is warranted. Returns a ComplexityScore with score in [0.0, 1.0].

Execute a complete code search for a research topic. Flow: 1. Check cache 2. Generate search queries 3. Search GitHub

Word-overlap Jaccard-ish similarity between two titles. Returns 0.0–1.0. Uses max(len) as denominator so short titles don't inflate the scor

Extract paper title from markdown text for LaTeX generation. Prioritises H1 headings that appear *before* the abstract section and look like

Reset agent to idle state, clearing all project-related fields.

Deterministic cache key from query parameters.

Generate a Markdown report from a pipeline run directory. Args: run_dir: Path to the run artifacts directory (e.g., artifacts/rc-xxx/)

Search multiple academic sources and return deduplicated results. Parameters ---------- query: Free-text search query. limit:

Convert to flat metric dict for backward compatibility. The existing pipeline expects dict[str, float] from parse_metrics(). This met

Make an LLM call and track count.

Escape curly braces in user-provided strings for str.format().

Extract meaningful keywords from text (lowercased, 3+ chars, no stops).

Build a fake Exa SDK result object.

Parse LLM response into decision list.

Try to extract JSON from text using three strategies. Always returns a ``dict`` or ``None`` — lists and other JSON primitives are dis

Check whether the proposed research has significant overlap with existing work. Parameters ---------- topic: Research topic strin

Factory: return the right LLM client based on ``config.llm.provider``. - ``"acp"`` → :class:`ACPClient` (spawns an ACP-compatible agent) - ``

Review all rendered figures. Context keys: rendered (list[dict]): From Renderer — each has 'figure_id', 'success'

Write search results to cache.

Remove all reasoning artifacts from LLM output. Handles XML <think> tags, bracket [thinking] blocks, insight decorators, plan markers, and ac