SuperAPI AI Directory

eve-compat

General & other

View repository

Updated 10-16 01:28

About this skill

Deploy eve projects on Zavu

Zavu compiles an eve-layout project into a first-class Zavu agent. The user's files are never edited: agent/instructions.md becomes the system prompt, each agent/tools/<name>.ts becomes a tool named <name>, and Zod inputSchema objects are converted to JSON Schema automatically.

Commands

# Import from GitHub and deploy in one step (also works for Zavu-native repos):
npx zavudev import owner/repo --sender sndr_123
npx zavudev import owner/repo#branch --root apps/agent   # monorepo / branch
# Private repos: set GITHUB_TOKEN in the environment (sent only to GitHub).

# Deploy a local eve project (detected by agent/instructions.md):
cd my-eve-agent && npx zavudev deploy --sender sndr_123

The sender can come from --sender (inlined at compile time) or the SENDER_ID function secret (npx zavudev fn secrets set SENDER_ID sndr_123). Without either, the agent deploys but stays disconnected until a sender is attached in the dashboard.

What maps

eve Zavu
agent/instructions.md system prompt (required)
agent/agent.ts string model agent model; non-string configs (provider instances, defineDynamic) fall back to the default with a warning
agent/tools/*.ts (default-export defineTool) tools, named by filename stem
Zod / JSON Schema inputSchema JSON Schema parameters
package.json dependencies installed server-side; the eve package itself is compiled away

What does NOT run on Zavu (always tell the user)

The deploy prints a compatibility report; these produce explicit warnings:

  • agent/channels/ — unnecessary: Zavu senders are the channels.
  • agent/schedules/ — cron is not available on Zavu Functions yet.
  • agent/skills/ — not loaded at runtime; fold content into instructions.md or a knowledge base.
  • agent/subagents/, agent/sandbox/, agent/connections/, evals/ — not supported. Agents that depend on them should run on eve's own runtime.
  • Tool approval gates and toModelOutput — ignored per tool, with warnings.
  • agent.ts options reasoning, compaction, limits, modelOptions, outputSchema — ignored, with warnings.

Tools run on every channel — plain text (SMS/WhatsApp/Telegram), voice, and flow tool steps — with up to 5 tool rounds per reply. The one exception is agents test: a dry run never executes tools (that would fire real webhooks) and says so in its warnings.

Requirements and troubleshooting

  • Compiling an eve layout needs esbuild resolvable in the project (npm install -D esbuild). npx zavudev import adds and installs it automatically; only --no-install skips that.
  • inputSchema must be an object schema — z.object({...}) or a JSON Schema with type: "object". Exotic Zod types fall back to a permissive schema with a warning in the build logs.
  • After deploying: npx zavudev agents list for the agent id, then npx zavudev agents test <agentId> --message "hi" for a dry run that delivers nothing and charges nothing.
  • Redeploy after edits with npx zavudev deploy — the reconcile summary (+ / ~ / = / -) is the source of truth for what changed.