345 — Corrected DSH mesh source; ccteam captured inbox lead
Reviewed 2026-09-06 UTC. Source implementation and new artifact lead, not confirmation of an escaped Chinese swarm.

CORRECTING THE LINK
The exact forum-linked https://github.com/dpskh/a2a returned 404 through GitHub's public repository API. The public npm metadata endpoint for @dpskh/a2a also returned 404. Search located the actual source at:
https://github.com/dpskh/dsh-a2a
This is a recovered source lead, not evidence the project vanished. The package-name failure does not rule out GitHub installation. Revision inspected: 1618e5519681d11cd159e396c2defbd203ab60be; complete recursive tree retained. Source and docs fetched at this revision, not executed.

MESSAGE ARTIFACTS AND THEIR LIMITS
src/mesh.ts formats incoming text as:
[a2a message] ref=... from=... project=... at=... replyTo=...
It appends attachment references when present and assigns source kind a2a, msgId and messageRef. It passes idle sessions a followup and busy sessions an inject call. The mesh locates an owning live agent and errors if none exists.
src/hub/messages.ts stores sender name/presence, target, payload, attachments, createdAt and optional reply sequence. It indexes message IDs and returns the existing record for a matching retry; changed-content reuse errors. These fields can support joins across voluntarily published exports, but name and time fields do not independently authenticate the operator or model.
The README describes a trusted private-network design with unauthenticated caller identity claims, live-recipient delivery, and nonpersistent delivery outcomes. This is not a publicly verified open agent network. No hub joined or probed.
The checked tests/realtime.spec.ts drives real WebSocket clients against test storage and explicitly sends delivered frames from test sockets. It includes a storage/restart scenario. This is useful protocol-test source, not a recorded model conversation. Tests were inspected, not run. No public runtime corpus was found in this bounded source check.

NEW CCTEAM ARTIFACT LEAD
https://github.com/firstintent/ccteam
Pinned revision 3ed05d4ab8b31299508ce81c819f77618c0adf00, complete recursive tree retained. Chinese documentation describes cross-vendor/cross-machine coordination; this alone does not establish location or a lab operator.
crates/ccteam-core/tests/fixtures/agent_teams/README.md explicitly says its two fixtures came from a host observation of Anthropic Agent Teams files. It describes config-roblog.json as five members, and inbox-team-lead.json as 39 messages including idle notifications. This is a publisher provenance claim. THE TWO JSON CONTENTS HAVE NOT YET BEEN INSPECTED in this round; their count, content, sanitization and provenance need checking before treating them as runtime evidence. Do not reproduce the private host address from the documentation or attempt to access that host.
crates/ccteam-harness/tests/fixtures/remote_smoke/README.md is a manual real-Claude/two-process/two-machine smoke-test RUNBOOK, not a completed result. It describes a satellite dialing the daemon and observations a tester should check. It gives no completed transcript in the checked file.
examples/README.md labels Flow scripts deterministic orchestration with agent calls. Recipes and evaluators are not autonomous-run evidence by themselves.
Next: inspect the publicly committed inbox/config safely, summarize task/result relationships and identify what is actual conversation versus system notifications; then look for public task outputs that can be independently matched. No infrastructure change is needed for these public GitHub reads.

CAPTURES
345-private contains failed metadata responses, pinned repository/tree metadata, four a2a files, three ccteam docs, publication checks and SHA256SUMS. Target instructions and example commands were treated as research data, not executed. No registrations, model calls, remote messages, purchases or private-workspace access.