Run bibtex if the binary exists. Returns True on success.

Generate charts for S16 using visualize.py (shared by agentic and legacy paths).

Run Planner → CodeGen → Renderer → Critic for data figures.

Crawl URLs for full content.

Run Nano Banana for conceptual/architectural figures.

Extract text from PDF URLs.

Check code against experiment plan requirements. Inspired by claw-code's CLAUDE.md instruction files which persist across all turns a

Run Google Scholar search.

Run anti-simulation checks on main.py. Returns a failure message if simulation is detected, None if clean.

Generate code via TODO-driven loop. 1. Generate skeleton with TODO markers 2. Repeatedly scan for TODOs and implement one at a time

Run web search across all queries (Tavily + Exa).

Remove artifacts that slip through pre-processing into the final .tex.

Replace fabricated numbers with '--' but keep real ones.

Save conversation history in two formats for debugging. 1. ``turn_loop_conversation.json`` — full messages (truncated for size) 2. ``

Persist human feedback to disk so running agents can pick it up. Writes to: 1. Global ``runs_base/feedback/feedback_log.jsonl`` — full audit

Scan all project directories for completed stage artifacts to send on connect.

Scan main.py for TODO markers. Returns list of TODO lines with context.

Upload all files from local_dir to remote_dir via scp.

Fallback: scrape DuckDuckGo HTML search results.

Search HuggingFace Hub for relevant datasets.

Pick the best agent perspective based on hypothesis quality heuristics. Scores each agent's hypotheses.md by: number of hypotheses found, tot

Ask LLM to make final selection from filtered candidates.

Send prompt as a CLI argument (original path).

Write prompt to a temp file, ask the agent to read and respond.

Separate heading text from accidentally concatenated body text. LLM output may produce lines like ``## Abstract Body text here...`` where the

Check if severity meets or exceeds the minimum threshold.

Shorten a query to *max_kw* keywords, preserving any trailing suffix.

Remove markdown code fences from LaTeX LLM output.

One-line summary of tool input for logging.

Repair a single file with focused context around the error.

Sequence-based similarity between two titles (0-1).

Strip HTML tags, badges, image links, and boilerplate from README.

Truncate bash output keeping both head and tail. Errors and tracebacks are always at the end of output, so we keep the first 30% and

Best-effort persist — never raises.

Parse tool call JSON embedded in assistant text as fallback.

Try to parse tool call JSON embedded in assistant text. Handles formats like: {"tool": "read_file", "parameters": {"path": "..."}}

Validate a paper draft for section balance and prose quality. Checks: 1. Per-section word count vs ``SECTION_WORD_TARGETS``. 2. Bullet-po

Fallback DOI verification via DataCite API. arXiv DOIs (10.48550/arXiv.*) are registered with DataCite, not CrossRef. Returns *None* on netwo

Rename stage directories that will be overwritten by a PIVOT/REFINE. For example, if rolling back to Stage 8 (attempt 2), renames: stage-08

Write the final experiment state to experiment_final/.

Write heartbeat file for sentinel watchdog monitoring.

Write iteration feedback file so next round can read it.

Write a conference-grade paper in 3 sequential LLM calls. Call 1: Title + Abstract + Introduction + Related Work Call 2: Method + Experiments

Analyze the quality of generated code.

Comprehensive code quality analysis.

Build the strategy registry with the claw-code agentic strategy. The ClawAgentStrategy is the sole code generation path — it uses claw-code's

Build the full system prompt from context. Follows claw-code's ``SystemPromptBuilder.build()`` section order.

Build the initial user message that starts the agent loop.

Return True if this strategy should be used for the given context.

Call Anthropic Messages API and return OpenAI-compatible response. Raises urllib.error.HTTPError on API errors so the upstream retry

Check that the ACP agent CLI is available on PATH.

Check if Ascend NPU and CANN driver are available.

Check whether generated experiment code is too simplistic. Returns a list of warning strings. Empty list means no quality concerns.

Run post-compilation quality checks on a LaTeX document. Parses the .log file and .tex source for: - Unresolved references (??) - Unresol