Repository navigation
Replies: 3 comments
那个
|
|
补上你要的原文与层位——先认两处修正,再说收窄后的诉求。 层位不在编排器 HTTP 层(编排器从不发起搜索调用)。401 出现在 web_search 工具结果里,同一记录带结构化错误字段。来源:session journal(逐 turn, 工具结果文本(逐字):
同记录的结构化面:
两处修正(对原帖)
计数按 存活下来的诉求(收窄为"把已有的码变成可观测的降级")
版本确认 0.2.0-rc.2(同你的基线注)。编排器场景=外部多智能体运行器(loopx 族),这条正是"码存在但 host 不可观测"造成代价的场景,已按你的建议写明。 |
|
跟进:根因在我方侧已定位——DeepSeek 账户欠费(我们今日内部确认;账面状态不在编排器可见面内,此前只能从 401 反推)。 值得记录的是:欠费在该端点呈现为与 key 写错完全相同的 |
Uh oh!
There was an error while loading. Please reload this page.
(Issues are disabled on this repo — posting here per the discussion channel. Happy to move it wherever the maintainers prefer.)
Environment
What happened (measured, not speculated)
During a fully-instrumented research run (2026-10-06, 17 fetch attempts by the agent), every
web_searchinvocation returned HTTP 401 on the configured Messages API endpoint — the entire search channel went down at once. Root cause on our side: no ARK credential configured (ARK_API_KEYnever set). Two observations that we believe are product gaps rather than our misconfiguration:No typed degradation event. The failure was only observable because the model itself reported it in prose. The runtime emitted no structured "search unavailable" event that a host/orchestrator could consume — downstream tooling cannot distinguish "search down" from "agent chose not to search".
No key-presence check / actionable error. A missing credential surfaces as a bare 401 mid-run instead of a startup-time "web_search requires ARK_API_KEY" warning. Our agent silently degraded to direct URL fetching + marking claims as unverified (14 claims in that run) — correct agent behavior, but avoidable damage.
Impact
Suggestions (directions, not prescriptions)
search_backend_degradedevent when the search tool errors N times.Happy to share run journals/CLI failure logs. The runtime itself is excellent — the turn journal + receipt/lineage model has given us five fully auditable runs.
All reactions