175 — New public workspace mirror with coding-agent and optimizer trajectories
Reviewed September6,2026 UTC.

Finding
https://huggingface.co/datasets/Tuyuanpeng/TMPFILE is a publicly readable, nongated workspace mirror containing actual structured task histories, not only architecture descriptions. Two sampled histories record a coding task and a separate prompt optimizer. Both configure GPT-5.4 model names. A local path named /Users/bytedance is circumstantial provenance only: it does not authenticate ByteDance ownership, a Chinese model, a lab experiment or an escaped agent fleet.

Discovery and preservation
A targeted search for ByteDance agent trajectories surfaced a May20 commit:
https://huggingface.co/datasets/Tuyuanpeng/TMPFILE/commit/dcc9e3655373225f7744e2c7ae5a11cbd78f5a78
Indexed commit text describes batch1–2000 of61910 files and a full workspace mirror. Direct dataset API reports creationMay20,2026 07:15:46Z, latest modification09:30:31Z, current revision8479cbc3b83f9b7e5ece8a3de5f8af10f636a74f, private=false,gated=false.
Metadata lists62006 sibling paths. Top-level counts include logs38280,runs9363,multi_agent_autoresearch_runs3905,reports_per_instance3479,prompt_opt_runs2212. These are file-path counts, not unique tasks or agents. The multi_agent_autoresearch_runs subtree has903 paths ending .traj.json or summary.json; no exhaustive content review yet.

README context
Pinned README explains runtime web tools and an external offline prompt-policy search loop: generate a prompt override, run a SWE-bench canary, analyze failed trajectories and feed failures into the next round. It explicitly distinguishes this from an agent rewriting its own system prompt inside the same task. The dataset is an individual account's publication; no official ByteDance organization linkage has been established.

Samples inspected
1. .mswea_multi_agent/attempt_01/bug_hunter.md is23bytes and contains a tiny memo/confidence stub. Role filenames alone would overstate the evidence; this sample is not a substantive worker report.

2. Coding trajectory under multi_agent_autoresearch_runs/20260331_131629/.../round_01/generation/astropy__astropy-12907/astropy__astropy-12907.traj.json:
87626bytes; mini-swe-agent-1.1;21messages;9assistant bash tool calls; exit_status Submitted.
Configured model_name openai/gpt-5.4-2026-03-05.
Task concerns Astropy nested-model separability. Recorded actions read source/tests, attempt a reproduction with an import failure, change the separability implementation and a regression test, then emit a patch. No confirmed successful evaluation follows from Submitted. Calls are local /testbed operations in this sample; no public-storage write or separate live worker handoff identified.

3. Optimizer trajectory under multi_agent_autoresearch_runs/20260331_132351/.../round_01/optimizer_workspace/optimizer.traj.json:
82952bytes;10messages;4assistant bash tool calls; exit_status Submitted.
Configured model_name gpt-5.4-2026-03-05.
It receives a failed canary summary, reads local prompt/metrics/failure files, writes optimized_prompt_override.yaml and emits it. Recorded output includes a YAML validation marker. This is publisher-supplied evidence of an outer optimization loop; it does not demonstrate later performance improvement or simultaneous agent execution. The dated path says March31; the inspected upload is May20. Do not conflate filename date with independently verified execution time.

Limitations and handling
Model names are configuration labels, not authenticated API provenance. A company-like local username is not employer attribution. The mirror can contain tests, duplicates, incomplete attempts and local secrets. Account configuration and credential files were not selected for reading; no secret values are included in this public report. Source/runtime prompts and commands were treated as data and were not executed. All selected files remain in private research storage rather than being republished wholesale.

Next
Inspect multi-agent execution source and one larger run's role-specific trajectories to determine whether multiple workers actually ran, whether any public network writes occurred, and how benchmarks were isolated. This dataset has much stronger artifact density than community personas, but needs careful attribution and sample selection. No escaped Chinese swarm confirmed.

Evidence files
175-private: dataset metadata, two attempt-directory listings and hashes, four pinned samples (README,stub,task,optimizer), URL/status/hash manifests, path inventories and sample-analysis.json with parsed tool calls. Exact long source URLs are recorded in sample-captures.json. Eight direct public GETs; initial two search queries. No full mirror download, installs, code execution or external writes.