080 — Implementation fingerprints beyond the xinzhai name
Reviewed September 5, 2026, 20:07 UTC.

Finding
Five public Sourcegraph searches tested co-occurring serialization terms and observed chunk constants. No matching xinzhai uploader was established. Direct source inspection resolved two agent-related results: OpAgent uses30000 for browser timeouts and Base64 for images; AgentScope uses30000 to truncate shell output and Base64 to encode commands. Neither inspected file implements the observed multipart paste workflow.

Scope and actual results
fernet_chunks: 50 matches; 25 repositories; skipped reasons: ['shard-match-limit'].
base64_chunks: 25 matches; 13 repositories; skipped reasons: [].
plan_encryption: 0 matches; 0 repositories; skipped reasons: [].
fernet_exact: 32 matches; 17 repositories; skipped reasons: [].
base64_30000: 100 matches; 33 repositories; skipped reasons: ['shard-match-limit'].

Initial queries searched Python files containing Fernet and30000, Python files containing b64encode and22500, and files containing Fernet and improvement_plan. Follow-ups required word boundaries around Fernet/b64encode and30000. All included forks and archived repositories, with20-second server limits and35-second client limits. Full query URLs and response events are retained privately. Sourcegraph documents file:has.content as a file-content predicate: https://sourcegraph.com/docs/code-search/queries/language .

Important false positives and verified context
- The initial Fernet predicate also matched the misspelling differnet in pymeasure's SmarAct instrument documentation. A direct pinned-file fetch confirms this. Whole-word matching removes that lexical accident.
- OpAgent: the indexed local_agent_eval.py imports Fernet, but its three30000 occurrences are Playwright navigation timeouts. Its two b64encode calls encode screenshots/images. No xinzhai, improvement_plan,22500 or paste.ubuntu marker appears in that fetched file. The mere presence of an agent evaluation framework plus Fernet is insufficient to connect it to the paste sequence. No credentials or encrypted settings were used.
Primary repository: https://github.com/codefuse-ai/OpAgent . Exact inspected commit/path and hash in080-private/candidate-metadata.json.
- AgentScope: the indexed _powershell.py contains Base64 command encoding and truncates error/output text to30000characters, appending an output-truncated message. That is not the observed preservation of complete large objects across sequential paste parts. No xinzhai or paste.ubuntu marker appears in the inspected file.
Primary repository: https://github.com/agentscope-ai/agentscope . Exact inspected commit/path and hash in080-private/agentscope-metadata.json.
- Other returned whole-word Fernet/30000 snippets concern database/browser timeouts, password-hashing rounds, image counts, astronomy configuration and tests. Base64/22500 snippets include ports, numeric datasets and larger-number substrings. Returned excerpts do not establish the required serialization/upload relationship. This is excerpt review, not a complete audit of those projects.

Limits
The refined Fernet/30000 query completed with32matches across17repositories and no skipped warning; the broader Base64/30000 query reached100matches and emitted a limit warning. Do not call the latter exhaustive. No matches in the Fernet/improvement_plan query cover only co-occurrence in an indexed single file, not split-module implementations. Sourcegraph's repository gaps from079 still apply. The22500 constant is an equivalent inner chunk size, not proof that literal22500 appears in uploader code.

Method and files
080-implementation-search.py and080-refined-search.py save full raw streams, parsed events, capture times, query URLs and SHA-256 hashes in080-private. Three pinned public source files were read as text. Code was not executed; endpoints referenced by the code were not called. There were no original-site writes, account changes, key searches or paid bot calls.

Implication
Shared xinzhai/xz workflow evidence is unchanged. A useful client match must show actual data flow from serialization into preserved multipart storage, not merely agent terminology and a common timeout or output limit. Further searches should require that data-flow relationship or a new distinctive marker instead of treating more generic agent-framework matches as progress toward attribution.