Run multiple search queries with cross-query deduplication.

Search OpenAlex for papers matching *query*. Parameters ---------- query: Free-text search query. limit: Maximum numb

Return True if approaching 80% of time budget.

Common setup when assigning a task to an agent.

IMP-32: Generate a descriptive caption from table header columns.

Build a CodeMem-style compressed summary via AST analysis.

Build a structured manifest of all figures.

Build a human-readable status summary for all running/recent projects.

Record a 429 response. Returns True if breaker is now OPEN.

Record a successful request.

Return (available, missing) sets, or None if endpoint unreachable.

If S12 sanity_report.json shows fail, pause the project and return notification messages.

Collect new/modified Python files from the workspace.

Collect variables assigned only inside if/elif branches.

Collect consecutive list items matching *pattern*.

Collect raw experiment metric lines from stdout for paper writing. Returns a tuple of (formatted block, has_parsed_metrics). ``has_parsed_met

Reproduce IMP-16 logic from executor.py lines 3367-3397. Returns (ci_low, ci_high, used_fallback).

Use Crawl4AI for high-quality extraction.

Detect the current git branch name, or None if not in a repo.

Read checkpoint and return (target_layer, next_stage) or None if no checkpoint.

Find or create the named acpx session.

Run a node's code in sandbox and update its score.

Run exec-fix loop if sandbox is available.

Core execution: single container, three-phase via entrypoint.sh.

Core execution flow for remote experiments. Steps: 1. Create a unique temporary directory on the remote host 2. Upload ex

S14 EXPERIMENT_RUN — delegates to the refactored experiment_run package. The full implementation uses a claw-code agentic turn loop to execute

Execute script in a local subprocess (no sandbox).

Parse ``<!-- FIGURE_PROMPT ... -->`` blocks from a paper draft. Returns a list of dicts, each with keys: figure_id, figure_type, section,

Extract key phrases from a research topic.

Extract markdown from a Crawl4AI result object.

Extract the agent's actual response from acpx output. Strips acpx metadata lines ([client], [acpx], [tool], [done]) and their continu

Extract 1-3 word search keywords from a topic string.

Generate a basic plan without LLM (used as fallback).

Find small, core .py files in codebases/ to pass as --read to aider. These give aider full visibility into key functions (e.g. mrsa_forward,

Find PDF URLs from search results.

Like ``_read_prior_artifact`` but returns the *Path* instead of content.

Fix a runtime error using targeted or full-file repair. E-05: Parse the error traceback to identify the failing file and line, then s

Format lessons into a text block for the LLM prompt.

Return a text-only tool transcript for Claude-style loops.

Return a text-only tool transcript for Claude-style loops.

Single code generation call with architecture spec injected.

Generate figure descriptions for injection into paper writing prompt.

Generate markdown image references for paper embedding.

Generate search queries from a topic string.

Generate requirements.txt content for additional packages.

Retrieve benchmarks and baselines from local knowledge base.

Run AST-based hard validation and repair critical issues. Critical issues trigger targeted file regeneration. Non-critical issues ar

Copy the immutable experiment harness into the target directory.

Launch L1 agent to produce ideas. s7_only=True runs only S7 (for discussion mode).

Launch S8 (HYPOTHESIS_GEN) for a single agent after discussion.

Use LLM to classify a research topic into a domain. Synchronous — ``llm.chat()`` is a blocking call.

Build _Section objects from (heading, word_count, use_bullets) specs.

Generate YAML frontmatter block.

Generate outputs from multiple debate perspectives. Each role has its own system/user prompt. Outputs are saved to *perspectives_dir* and ret

Return the next figure number for the current thread.

Collect all final user-facing deliverables into a single ``deliverables/`` folder. Returns the deliverables directory path, or None if nothing wa

Parse alignment indicators from separator line.

Parse stdout using the existing regex-based parser. Delegates to ``sandbox.parse_metrics`` for backward compatibility.

Interactively prompt the user to install OpenCode. Returns True if OpenCode is now available (already installed or just installed successfull

Return the input queue name that feeds into target_layer.

Enforce rate limiting between search requests.

Extract quality score from the most recent quality_report.json.

Render extracted FIGURE_PROMPT entries via NanoBananaAgent. Uses the OpenAI-compatible proxy (same base_url/api_key as the LLM) to call an im

Render a Markdown table as a LaTeX tabular inside a table environment. IMP-23: Auto-wraps in ``\\resizebox`` when columns > 5 or any cell tex

Reset per-render figure and table counters for the current thread.

Best-effort name resolution for a Call node's func.