Gemini continuation and /compact fail with "Tool name is required" on mixed-model history

No source files or tests are named. Start by tracing cross-model history normalization and the Gemini request adaptation, including how custom.name is handled during normal requests and /compact. Done means supported model switching, resume, rewind, and compaction no longer fail on historical apply_patch calls, with a regression test covering the mixed-model history.

Describe the bug

A long-lived, mixed-model Copilot CLI session fails to continue with Gemini:

400 Tool name is required for Gemini requests
{"message":"Tool name is required for Gemini requests","code":"invalid_tool_call_format"}

Multiple normal prompts fail before any new tool executes. Manual /compact also fails with the same error, so compaction does not provide recovery in this session.

The captured failed compaction request provides a useful triage lead: its history contains three GPT-originated type: "custom" / custom.name: "apply_patch" calls among ordinary function calls. All tool results have matching, preceding calls with nonempty names when both call types are accounted for. This is not evidence of a missing tool name in the persisted call itself.

Suspected area: cross-model history normalization / Gemini request adaptation. This is a hypothesis, not a confirmed root cause or an assertion that two independent bugs exist.

Affected version

• Copilot CLI 1.0.86, verified from the running process's loaded .../1.0.86/prebuilds/darwin-arm64/runtime.node.
• Session originally created under 1.0.83 and subsequently resumed across updates; the version that introduced the failure is unknown.
• Model at failure: gemini-3.8-flash.
• Earlier turns used gpt-6-astra, including built-in apply_patch.
• macOS 26.7, Apple Silicon (arm64), tmux.

Steps to reproduce the behavior

Observed workflow, not yet a clean-room/minimal reproducer:

1. Continue a long-lived session that has used both GPT and Gemini, with previous compactions and tool calls.
2. GPT turns in that history use built-in apply_patch, persisted as custom tool calls.
3. Continue later with Gemini. Conversation-only rewinds and process resume/restart occurred before the reported failures. Their causal role has not been isolated; switching models alone is not claimed to reproduce this.
4. Submit a normal follow-up prompt. It fails with invalid_tool_call_format.
5. Retry later: the same error recurs.
6. Run /compact: it also fails, with session.compaction_complete.success: false and error: "host-rethrow: 400 Tool name is required for Gemini requests".

The real environment includes user hooks, plugins, and MCP servers. These have not been eliminated in a fresh configuration. This report does not establish reproducibility in a stock installation.

Expected behavior

A session created through supported model switching, tool use, resume, and rewind should remain usable. Historical custom tool calls should be translated into a supported representation when replayed to another provider, or the CLI should provide an actionable compatibility/recovery error without silently discarding history.

Additional context

Captured request metadata

For the failed manual compaction, model.model_call_failure records:

{
  "model": "gemini-3.8-flash",
  "status": 400,
  "api_endpoint": "/chat/completions",
  "transport": "http",
  "failure_kind": "api",
  "badRequestKind": "structured_error",
  "errorCode": "invalid_tool_call_format"
}

Parsing that event's JSON-encoded requestMessages locally:

Check	Result
Messages	506
Assistant tool calls	258
Function calls with nonempty function.name	255
Custom calls with nonempty custom.name	3, all apply_patch
Tool-result messages	258
Results without a matching preceding named call	0

The three custom calls are at zero-based message indices 85, 107, and 115; their matching results immediately follow. All 258 result messages omit a standalone name; each has a tool_call_id that resolves to its preceding call. This alone is not asserted to be invalid in the OpenAI-shaped intermediate format.

Synthetic structural illustration only (not an executed reproducer; IDs, input, and output replaced):

[
  {
    "role": "assistant",
    "content": "",
    "tool_calls": [{
      "id": "custom_call_example",
      "type": "custom",
      "custom": {
        "name": "apply_patch",
        "input": "<synthetic patch placeholder>"
      }
    }]
  },
  {
    "role": "tool",
    "tool_call_id": "custom_call_example",
    "content": "<synthetic result placeholder>"
  }
]

Please check whether the Gemini adapter handles custom.name, rather than assuming every historical call has function.name. The captured request does not establish which field the server rejected; it is not a capture of the final native Gemini payload.

Trace correlation

UTC time, 2026-09-19	Operation	Request ID	Service request ID
21:44:48	Normal prompt: HTTP 400	7B0D:345026:2C3E730:36BEE94:6AAF024C	138e7b2d-aebe-4123-97f2-8b61e81f383e
22:07:00	Later normal prompt: HTTP 400	7F6E:178245:2FFDC9D:3C2C1FB:6AAF077F	4b9439d5-7085-490b-8815-a43414f9cea4
22:07:20	Manual compaction: HTTP 400	4DFE:3B6359:2FDC2F9:3C0B645:6AAF0797	90eafc3c-9e21-4633-a126-9f3bfc45a047

Important limits / ruled-out inference

• Some persisted event parentId links are dangling, but following those links alone incorrectly suggested an orphaned image-view result. The actual captured request contains both image-view calls, their names, and their results. Therefore, an orphan caused by rewind is not established.
• Successful later Gemini turns exist after the original custom-tool turns. This is not established as an immediate model-switch failure; resume/replay or runtime-version differences may matter.
• No controlled retry with only the custom calls normalized has been performed. No successful model-switch workaround is claimed.
• The live session was inspected read-only. No repair, event deletion, or history rewrite was performed for this investigation.
• No original conversation, image, patch contents, filesystem paths, or configuration is attached. Only structural findings and request-correlation metadata are included.

Related but not established as duplicates: broader Gemini 400 report and older tool-use/result corruption report.