Repository navigation
Aventide - Taking on client management with eve in production #5811
christianquiver
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone!
I'm Christian Quiver, cofounder of Aventide (https://aventide.ai). We're two independent business owners building AI client management for coaches, consultants and designers, including people who run more than one business. Our assistant runs on Eve 0.83.3 in production.
Here's what we built and what we learned, in case it's useful.
What it does
agent/tools/, found by Eve's file-based discovery.eveChannel), Slack, Telegram, and WhatsApp as our owndefineChannel.approval: always()and 12 more ononce(), so a draft for a client waits in the chat for the owner. Our product rule: "Anything Aventide drafts for a client waits for your approval; automations you switch on run on their own."What we learned
1. Put the approval in the tool, and decide who may answer it.
Approval lives in the tool definition, not in prompt text. From eve 0.72 only the requesting principal can answer, but Telegram and Slack buttons answer as the platform account, so every press was refused. We wrap each policy in
ownerApproval(): the requester, or a button press from the same linked platform account. A test fails the build for any tool that skips the wrapper.2. 0.83 durable callbacks bite wrapped policies.
Right after our 0.83.3 deploy, four tools vanished from every session with
callback "approvalResponse" does not have a durable descriptor. A policy made by a function call is invisible to Eve's transform.defineDurableCallbackfixed it.3. Cap tool output in a model middleware, not in the tool.
toModelOutputruns once. When a native-channel session replays its history, the rawexecute()result is what gets re-sent, and one tool had shipped 1.1 MB. AwrapLanguageModelmiddleware now caps every tool-result part at 24k chars right before the model call, so it catches fresh calls, deferred-tool results and replayed history alike.4. Guard tool discovery in CI.
eve buildturns every top-level entry ofagent/tools/into a tool, and it fails afternext buildhas already passed. A__tests__/folder shipped in a merge and our alias served 502s for 38 minutes. A small script now runs on PRs and in the merge queue.5. Native sessions never end on their own.
Phone-channel sessions replay the whole transcript every turn and only compact near 90% of a 1M window. We roll over to a fresh session after 6 idle hours or 90k input tokens, with the cursor stored in Postgres so it survives cold starts.
6. Share the prompt cache across surfaces.
Our tools are ~33k tokens, 77% of a cold start, and were being cached once per prompt variant. One breakpoint on the last eager tool makes one shared entry. We also pin the Gateway's provider order to Anthropic, because a cache entry written on one upstream is invisible to the others.
7. Keep traces clean twice.
Our Agent Runs export policy redacts inputs and outputs, and a test pins that no channel ever reports
audience: "public". Slack is DM-only for that reason.Happy to answer questions, and thank you to the Eve team.
Christian
All reactions