213 — DeepSeek official team journal: persistence format and synthetic fixture markers Reviewed September 6, 2026 UTC. Source inspection only; no installation or execution. RESULT DeepSeek's official Harness repository implements an experimental team domain with persistent peer messages and shared dependent tasks. This is a direct vendor-associated source lead. The inspected headless team test substitutes a deterministic adapter, so its completed research/implementation dialogue is not evidence of model execution. No public real team run or escaped activity was recovered in this pass. PROVENANCE https://deepseek.com/harness/ The official site links deepseek-ai/deepseek-harness and describes an append-only trajectory record including subagent scheduling. This verifies the official project relationship, not any third-party run's model/operator identity. https://github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.1.2-alpha.1 GitHub API publication: August 27, 2026 at 17:06:37 UTC. Tag tree cd5ef8148158c3a752a658978873241fdf8e2bbc, 10,389 entries, non-truncated. Five captured files match their listed Git blob hashes. Source dates in notes are not independently archived operational dates. Release notes distinguish plugin-name/version reporting from optional session-log upload, and describe public WebFetch enabled by default. These statements do not establish public log publication or outbound writes by a particular run. A third-party search result emphasizes Agent Teams for this release; the official release notes inspected here do not explicitly announce teams, though the pinned tree contains them. TEAM RECORDS AND JOIN KEYS Pinned docs/subsystems/agent-team.md describes the team ID as the root session ID, teammate identity as a persistent session ID, and monotonically allocated local task IDs such as task-1. Member names are labels, not provenance. Messages carry sender ID/name, target ID, message ID, content and quiet/wakeup delivery mode. The lead retains queued messages. Delivery acknowledgement follows durable storage at the target. Target-side source.kind is team-message, with teamId/messageId/senderId/senderName for deduplication. Paired queue and delivery records can corroborate transport inside an export; they do not prove the receiver acted on the message or completed a task. Tasks carry revision, owner, blockedBy and advisory writeScopes. Task-completed records represent system state, not independently verified deliverables. Event seq/time give ordering/timing; ordinary inherited fork events retain their original team ID and should not be counted as a new team's independent work. SOURCE-CHECKED JOURNAL packages/experimental/agent-team/src/journal.ts serializes changes per lead, appends team/member, team/task, team/message/queued and team/message/delivered events, flushes the session, then emits the commit notification. Its comments say these events do not enter the conversation surface. Thus a chat-only screenshot/export may omit team state that the underlying session journal contains. These event names and correlated IDs are useful search/parser fingerprints, not vendor authentication. Other software can copy them. DETERMINISTIC TEST: DO NOT COUNT AS A LIVE SWARM apps/cli/tests/agent-team-headless.e2e.ts disables llm-deepseek, injects a fixture, clears the API-key environment value, and checks for three JSONL files, four member lifecycle events, queued/delivered messages, and two completed tasks. It expects TEAM_WORKFLOW_OK. This is test source, not a captured result; the test was not run here. The referenced team-llm.mjs calls itself deterministic and keyless. Its scripted lead spawns implementer and researcher. The researcher creates/completes task-1 and messages the implementer; the implementer waits for that task, completes task-2 and messages the lead. The adapter reports fixed token counts and canned final success. It registers under the deepseek-official provider route despite making no model request. Consequently even that provider label in fixture-generated logs would not establish actual DeepSeek inference. Useful fixture exclusion markers: TEAM_WORKFLOW_OK, team-fixture-, RESEARCHER_MARK, IMPLEMENTER_MARK, team-fixture-llm. Their combination strongly suggests this fixture or a derivative; isolated generic research/implementation role names do not. BOUNDED DISCOVERY SEARCH Queries for team/message/queued with DeepSeek outside github.com and for TEAM_WORKFLOW_OK recovered documentation/source mirrors rather than a public real run in returned results. No exact-match corpus scan was run this pass. A separate Chinese real-world-test query surfaced a Tencent Zhuque primary report claiming 14,560 controlled DeepSeek Harness security evaluations. That is a next lead for voluntarily published trajectories, not yet verified evidence in this report: https://matrix.tencent.com/zh/2026/08/20/deepseek-harness-agent-injection-risk PINNED FILES Base: https://github.com/deepseek-ai/deepseek-harness/blob/cd5ef8148158c3a752a658978873241fdf8e2bbc/ - docs/subsystems/agent-team.md - packages/experimental/agent-team/README.md - packages/experimental/agent-team/src/journal.ts - apps/cli/tests/agent-team-headless.e2e.ts - apps/cli/tests/profiles/headless/tests/fixtures/team-llm.mjs NEXT Look for public real session exports that link distinct member IDs, paired message receipts, task transitions and external work artifacts. Check adapter provenance before trusting provider labels. Investigate the Tencent controlled-run publication for accessible traces and clear distinctions between test infrastructure and actual open-internet activity. Do not seek private telemetry endpoints or install investigated fixtures. CAPTURES 213-private contains release metadata, pinned recursive tree, five Git-verified files and SHA256SUMS. Public report contains reviewed findings only.