External provider models: sessions create but model calls never execute with valid model IDs

[jacobabiola]

Description

When using a direct @ai-sdk/anthropic provider model in agent.ts, Eve sessions are created (HTTP 202) but model calls never execute. The session immediately parks with a continuationToken and no turn processing occurs. There are zero errors, zero log activity, and zero diagnostics.

Critically, this only happens with valid model IDs. When using a deliberately invalid model ID (e.g., claude-haiku-4-5-WRONG), Eve does fire the model call and hits the Anthropic API (which correctly returns not_found_error). This confirms the provider, network path, and env vars are all functional.

Reproduction

agent/agent.ts

import { defineAgent } from "eve";
import { anthropic } from "@ai-sdk/anthropic";

export default defineAgent({
  model: anthropic("claude-haiku-4-5"), // valid model ID — never executes
  // model: anthropic("claude-haiku-4-5-WRONG"), // invalid — DOES fire API call
});

agent/instructions.md

Say HELLO in one word.

Steps

1. Set ANTHROPIC_API_KEY and ensure ANTHROPIC_BASE_URL points to https://api.anthropic.com/v1
2. Run eve dev --no-ui --port 3001
3. Send any message via the Eve Client:

const client = new Client({ host: "http://127.0.0.1:3001" });
const session = client.session();
const resp = await session.send({ message: "Say HELLO" });
// resp = { continuationToken: "...", sessionId: "..." }
// No text, no model call in logs

4. Check /tmp/eve-dev.log — zero activity beyond server startup

Actual behavior

• Session created (HTTP 202) with continuationToken
• Response in ~40-170ms (no model call occurred)
• No model call entries in logs
• No AI_ errors in logs
• Continuing the session also parks immediately
• Streaming the session returns no events

Expected behavior

Model call fires, Anthropic API responds, turn completes, text returned.

Additional finding: subagent configs empty

When using subagents (with coordinator pattern), subagent agent.ts files are discovered (appear in compiled manifest) but their config is always {}:

Eve 0.11.6: past-appearances: config={}
Eve 0.13.8: past-appearances: config={}
Eve 0.14.0: past-appearances: config={}

Root agent config IS read correctly. This was tested across all three versions.

Environment

• OS: macOS (Darwin 24.3.0)
• Node: v24.17.0
• Next.js: 14.2.35
• Eve versions tested: 0.11.6, 0.13.8, 0.14.0
• @ai-sdk/anthropic: 3.0.85
• ai: 7.0.0-beta.178

What we verified works (independently)

• curl to api.anthropic.com/v1/messages with the same API key and model IDs ✅
• generateText from ai with anthropic("claude-haiku-4-5") (outside Eve) ✅
• Eve compilation — 0 diagnostics, 0 errors, 0 warnings ✅
• Eve server startup — listening on port 3001 ✅
• Session creation — returns 202 with sessionId ✅
• With invalid model ID: API call DOES fire (hits api.anthropic.com/v1/messages) ✅

Note

The ANTHROPIC_BASE_URL environment variable had to be explicitly overridden because Claude Code (the AI coding tool) sets it to https://api.deepseek.com/anthropic globally. We override it to https://api.anthropic.com/v1 when starting Eve. This was verified in the Eve process environment before each test.