#general
Oct 13October 21, 2026
Hi all 👋 Kun here. We've spent the last few weeks running Archestra as the front door for our team's AI work, now on 1.4.0-rc.55: the built-in chat agent with Splunk and ServiceNow tools, Claude Code and OpenCode as runtime agents, and our own model service behind the LLM proxy. Claude did most of the hands-on testing, with a human supervising.
The deck is a tutorial we use to teach our own team: how an agent works, where its pieces should live, and how a team shares one case (shared project, private working chats, every call under the person's name). It walks through the Studio pages, the model and tool logs, delegation to Claude Code, and how outside agents can join.
What held up: per-person attribution on every model and tool call we made, from chat, personal gateways and runtime pods, delegated runs included.
Fixed by rc.55 (we re-measured):
• An Admin can now invite people as Member, Editor or Admin
• A Claude Code run's first request now carries the gateway's tools
• An oversized tool result is saved to a sandbox file instead of ending the chat turn
Notable ones still open:
• A run started with start_run joins no project, so files it saves land in no project or chat
• Project files are one flat list: no folders, and a saved file keeps only its base name
• savefile and editapp take "exactly one of" several optional fields; OpenAI models fill all of them, so every call is refused
• Chat sends no reasoning_effort for gpt-6 models, so they can't use tools without a proxy shim
• OTLP export is mostly database and auth spans with SQL text, unless Sentry is on
• The LLM and MCP logs can't be searched by content
• Sandbox results are reused across sandboxes and users with the same command history: #8662
We're filing 10 issues on GitHub with reproductions over the next few days. If you'd like the full field report (15 issues, reproductions, and what changed from rc.26 to rc.55), send me a message and I'll share it.
One heads-up for anyone on older notes: since rc.36 the per-tool call policies are gone, and Guardrails (OpenAPPA) replace them.
Read-only live mirror of Archestra.AI Slack
👋Join the discussion withAI enthusiasts!Hi all 👋 Kun here. We've spent the last few weeks running Archestra as the front door for our team's AI work, now on 1.4.0-rc.55: the built-in chat agent with Splunk and ServiceNow tools, Claude Code and OpenCode as runtime agents, and our own model service behind the LLM proxy. Claude did most of the hands-on testing, with a human supervising.
The deck is a tutorial we use to teach our own team: how an agent works, where its pieces should live, and how a team shares one case (shared project, private working chats, every call under the person's name). It walks through the Studio pages, the model and tool logs, delegation to Claude Code, and how outside agents can join.
What held up: per-person attribution on every model and tool call we made, from chat, personal gateways and runtime pods, delegated runs included.
Fixed by rc.55 (we re-measured):
• An Admin can now invite people as Member, Editor or Admin
• A Claude Code run's first request now carries the gateway's tools
• An oversized tool result is saved to a sandbox file instead of ending the chat turn
Notable ones still open:
• A run started with start_run joins no project, so files it saves land in no project or chat
• Project files are one flat list: no folders, and a saved file keeps only its base name
• savefile and editapp take "exactly one of" several optional fields; OpenAI models fill all of them, so every call is refused
• Chat sends no reasoning_effort for gpt-6 models, so they can't use tools without a proxy shim
• OTLP export is mostly database and auth spans with SQL text, unless Sentry is on
• The LLM and MCP logs can't be searched by content
• Sandbox results are reused across sandboxes and users with the same command history: #8662
We're filing 10 issues on GitHub with reproductions over the next few days. If you'd like the full field report (15 issues, reproductions, and what changed from rc.26 to rc.55), send me a message and I'll share it.
One heads-up for anyone on older notes: since rc.36 the per-tool call policies are gone, and Guardrails (OpenAPPA) replace them.
1 reply
Hey there! Real glad that you're filing these issues!
Unless we are dealing with some more sensitive stuff or unless the reports are massive, I think you can add most of the report directly in the issues.