Question 1

How can I renew an expired OAuth token for an existing provider account in Manifest?

Accepted Answer

There is no built-in "Renew Auth Token" action yet. The current workaround is to delete the expired provider account and add it again via the OAuth flow. This will re-authenticate and update the credentials, but you will need to reapply any routing settings, agent-specific selections, and account labels. A feature request exists for a non-destructive renewal flow.

Question 2

Why does DeepSeek tool calling fail with 'reasoning_content must be passed back' error in Manifest?

Accepted Answer

This is a known bug (#2047) where Manifest drops the required reasoning_content field in multi-turn conversations. Workaround: remove DeepSeek models from any tier that uses tool calling (e.g., 'complex' tier). This avoids the error but loses reasoning-quality responses. A permanent fix to preserve reasoning_content is pending.

Question 3

Why are tokens and costs showing as 0 for successful LLM requests in Manifest?

Accepted Answer

Upgrade to Manifest v6.8.2, which fixes missing streamed usage requests for dynamic custom providers. If tokens still show 0 after upgrading, the upstream provider (e.g., LLM7.io) is not returning a `usage` block in its API response. Manifest records exact tokens from that block; if absent, it reports 0 and cannot calculate cost.

Question 4

How to fix Gemini streaming tool calls being merged, changing completion IDs, or returning finish_reason "stop" instead of "tool_calls"?

Accepted Answer

Upgrade to a release containing the PR "fix: keep Gemini stream state across SSE events." The bug is in `transformGoogleStreamChunk` in `google-adapter.ts`: it recreates stream state for every Gemini event. After upgrading, state is maintained once per SSE stream, so: (1) `tool_calls[].index` increments across events instead of resetting to 0; (2) one completion ID and `created` timestamp are reused for all deltas; (3) if any tool call was emitted, the terminal finish reason is `tool_calls`, even when a usage-only trailer follows; and (4) finish/usage chunks are emitted only once. If you cannot upgrade, move the completion ID, `created` value, next tool-call index, and `hasToolCalls` flag outside the per-event transformation so they persist for the full stream.

Question 5

How to fix Gemini `functionDeclarations` being recorded as unnamed empty tools in llm-gateway?

Accepted Answer

Upgrade to a version containing the PR "fix: keep tool calls and Gemini tools in recorded conversations." The bug is in `extractRequestTools` in `packages/shared/src/chat-message.ts`: it normalizes `tool.function` and flat tool fields but does not handle Gemini's `functionDeclarations` array.

As a workaround, add a Gemini-specific branch that expands each declaration:

```js
return tools.flatMap((tool) => {
  if (Array.isArray(tool.functionDeclarations)) {
    return tool.functionDeclarations.map((fn) => ({
      type: "function",
      function: {
        name: fn.name,
        description: fn.description,
        parameters: fn.parameters,
      },
    }));
  }

// Keep the existing normalization for tool.function and flat tool fields.
});
```

After the fix, a Gemini request containing `functionDeclarations: [{ name: "lookup", ... }]` is recorded as a named function tool with its description and JSON Schema parameters instead of `[{"type":"function","function":{}}]`.

Manifest

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects