157 — MiroFlow full-collection tool inventory and narrow marker screen
Reviewed2026-09-06 UTC

Scope and source
All165 task JSON members in the official archive described and hash-verified in156 were parsed locally. No source code or recorded command was executed; no URLs found inside traces were requested. JSONL/TXT ancillary files were not included.
https://huggingface.co/datasets/miromind-ai/MiroFlow-Benchmarks/resolve/main/gaia_validation_miroflow_trace_public_20250825.zip

Format and validation
These traces record tool invocations as use_mcp_tool XML blocks inside assistant-role content, not native tool_calls arrays. The initial native-field parser found zero and was rejected. The corrected parser extracts complete server_name/tool_name/arguments blocks from assistant messages only, excluding system tool definitions. It recovers61 calls for156's independently inspected sample, matching its detailed count; every task has at least one extracted call. Regex extraction is a structured inventory, not a semantic audit of every recorded action or proof that the calls succeeded.

Results
165 task records;617 named worker-session histories;7,190 extracted assistant tool-call blocks;503 distinct hostnames in literal HTTP(S) strings in tool arguments.
Most frequent calls:
1,883 tool-searching/scrape_website
1,450 tool-searching/google_search
767 tool-code/run_python_code
617 agent-worker/execute_subtask
346 tool-searching/wiki_get_page_content
344 tool-image-video/visual_question_answering
287 tool-code/run_command
258 tool-code/create_sandbox
246 tool-code/download_file_from_sandbox_to_local
220 tool-reasoning/reasoning
188 tool-searching/search_archived_webpage
178 tool-searching/search_wiki_revision
158 tool-code/download_file_from_internet_to_sandbox

Host counts are literal appearances in arguments, including queries/code, not confirmed requests, successful visits or unique destinations per task. The most frequent are en.wikipedia.org, www.youtube.com, web.archive.org, arxiv.org and pubchem.ncbi.nlm.nih.gov. Do not infer target traffic volume from these counts.

Narrow scratch-memory marker screen
Zero argument matches for this explicit set: pastebin, paste.ubuntu, ntfy., telegra.ph, rentry., is.gd, tinyurl, xinzhai, xz_knowledge (case-insensitive). This is useful negative evidence for these specific strings only. It does not exclude every paste host, arbitrary uploads, dynamically constructed URLs, writes hidden in code, or behavior omitted from publisher logs. Tool names containing upload/download refer to the framework's sandbox transfer interface; their names alone do not establish public storage.

Interpretation
The collection supplies a concrete hierarchical research-workflow reference, including hundreds of recorded delegations and thousands of tool-call blocks. This supports a substantially better behavioral comparison than searching for the English phrase agent swarm. The next useful analysis is command/code network behavior and distinctive argument patterns, followed by comparisons against the existing unexplained-public-artifact corpus. No escaped swarm, Chinese-lab execution attribution or XZ connection has been established.

Reproducibility
157-private/inventory.py is the corrected parser. inventory.json contains per-task counts, complete tool-frequency and hostname tables, and marker-hit records. The rejected zero-count result was overwritten and is not a finding. Source ZIP SHA256 is in156. All outputs remain local apart from this report and dashboard summary.