Gemini history rehydration drops valid thought-signature replay; stock Anthropic also hits duplicate step IDs across resumed runs

[kaisewhite]

Summary

When an AgentKit run is resumed from persisted history inside an Inngest function, Gemini tool-call transcripts that were stored with a valid thought_signature still fail on the next agent.run() with error making inference request: unsuccessful status code: 400.

Separately, a clean non-Gemini control run on stock @inngest/agent-kit@0.13.2 with Anthropic (claude-haiku-4-5) fails with:

Duplicate step ID "routine-runner-seed-session-spike" detected across parallel chains

followed by the same step.ai 400.

Environment

• @inngest/agent-kit@0.13.2
• Inngest dev server local
• Bun runtime for the app
• Inngest function owns the agent run and passes step into agent.run(..., { step })
• History is persisted via history.get / history.appendResults

Gemini repro

Goal: validate a long-running agent session that spans sleepUntil boundaries while preserving the same transcript.

High-level flow:

1. agent.run() executes inside an Inngest function.
2. Agent calls a balance tool.
3. Tool result is returned successfully.
4. Agent emits CHECK 1: ....
5. Function sleeps with step.sleepUntil(...).
6. On wake, a second agent.run() is invoked with the same AgentKit history thread.
7. Second inference fails with error making inference request: unsuccessful status code: 400.

The important part: the persisted history row does contain the original Gemini tool call with a real thought_signature.

Stored history excerpt:

{
  "output": [
    {
      "role": "assistant",
      "type": "tool_call",
      "tools": [
        {
          "id": "get_account_balance",
          "name": "get_account_balance",
          "type": "tool",
          "input": {
            "reason": "Fetch account balance for CHECK 1"
          },
          "thought_signature": "<real signature present>"
        }
      ],
      "stop_reason": "tool"
    }
  ],
  "toolCalls": [
    {
      "role": "tool_result",
      "tool": {
        "id": "get_account_balance",
        "name": "get_account_balance",
        "type": "tool"
      },
      "type": "tool_result",
      "content": {
        "data": {
          "cash": 50000.03,
          "equity": 50000.03,
          "buying_power": 200000.12
        }
      },
      "stop_reason": "tool"
    }
  ]
}

The same run also has the subsequent assistant text turn stored correctly:

{
  "output": [
    {
      "role": "assistant",
      "type": "text",
      "content": "CHECK 1: equity 50000.03, cash 50000.03, buying_power 200000.12",
      "stop_reason": "stop"
    }
  ],
  "toolCalls": []
}

So the signature is not absent in storage. The failure appears during AgentKit's history rehydration / replay-to-provider-request mapping for the next agent.run().

Anthropic control repro

To avoid Gemini signatures entirely, I ran the same spike against stock AgentKit with:

• provider: Anthropic
• model: claude-haiku-4-5
• 3 iterations
• 1-minute sleeps owned by the executor, not a wait tool

That run fails on the first balance tool call with app-side output:

[tool:get_account_balance] fetched account snapshot.
Duplicate step ID "routine-runner-seed-session-spike" detected across parallel chains

and Inngest dev logs:

error handling generator response: error making inference request: unsuccessful status code: 400

This was on unpatched @inngest/agent-kit@0.13.2.

Why this seems upstream

• Gemini: the durable store contains the needed thought_signature, but the resumed run still 400s.
• Anthropic: the same high-level executor-loop pattern hits duplicate AgentKit/Inngest step IDs on stock AgentKit before the session spike can complete.

That points to AgentKit replay/orchestration rather than application-level state loss.

Request

Could you take a look at:

1. Gemini history rehydration for persisted tool-call transcripts containing thought_signature.
2. Whether agent.run() inside resumed Inngest executions needs unique inference step IDs across turns / wakes, especially when the same agent name is reused.

If helpful I can provide a smaller standalone repro extracted from the app, but the diagnosis above is already from a minimal durable-session spike.

[linear-code]

EXE-2068