[Bug] Mid-session model switch fails: `input[87]` missing required field `summary` (HTTP 400, non-retryable)

[Bug] Mid-session model switch to third-party model fails: input[87] missing required field summary (HTTP 400, non-retryable)

Summary

Switching models mid-session (GLM -> third-party model via an OpenAI-compatible proxy) makes every subsequent request fail with:

Upstream request failed: [invalid_request_error] `input[87]` missing required field `summary`

The failure happens at request time (iteration 0, before any generation) with HTTP 400 retryable: false, so retrying with the same payload fails identically. Starting a fresh session directly on the target model works fine, so the model/upstream itself is healthy — the problem is the request payload ZCode builds from cross-model conversation history.

Environment

• ZCode Desktop 3.11.2 (build 3.11.2.6792), macOS Apple Silicon
• Source model: builtin:zai-coding-plan/GLM-5.3-Flash
• Target model: custom provider opencode-go, model opencode-go/muse-spark-1.3-contributor, via local OpenAI-compatible proxy http://127.0.0.1:10100/v1 (SSE transport)

Reproduction

1. Start a session on GLM-5.3-Flash and hold a long conversation (in my case ~48 messages / ~196 parts, including tool calls and a few model switches back and forth).
2. Switch model mid-session to the third-party model.
3. Send any message -> immediate error banner: TerminalStreamChunkError: Upstream request failed: [invalid_request_error] input[87] missing required field summary, stack pointing at handleStreamErrorEvent / runStreamText in resources/glm/zcode.cjs.
4. Retrying fails the same way; only a fresh session on the target model works.

Evidence (from ~/.zcode/cli/log/zcode-2026-09-17.jsonl)

• session.model.updated at 2026-09-17T03:35:29Z: GLM-5.3-Flash -> opencode-go/opencode-go/muse-spark-1.3-contributor
• First failure at 2026-09-17T03:37:25Z (turnNumber: 4, iteration: 0 — fails on the first request of the turn, i.e. payload validation, not generation):
  • model.request.failed: reason: invalid_request, retryable: false, statusCode: 400, statusMessage: Upstream request failed: [invalid_request_error] input[87] missing required field summary
  • turn.failed with AiSdkModelAdapterError / ProviderBusinessError, responseStatus: 400, providerKind: openai-compatible
• Second identical failure at 2026-09-17T03:38:30Z on the next attempt (requestIds fe2cfbc0-… and a9132c71-…).

Analysis (my read, for triage)

summary is a field on reasoning-type items in the Responses API shape. After a mid-session switch, ZCode replays the full history — including the previous model's reasoning blocks — into input[]. Item 87 (a reasoning block from the GLM side, which carries no summary in the form the new target expects) fails the target's strict validation. Because it is a 400 non-retryable error, the retry loop correctly gives up — but the user is left with a permanently broken session and no in-app recovery path.

Suggested fix

When building the Responses input[] after a model switch (or whenever history contains reasoning items from a different model family), sanitize cross-model history: drop reasoning items that lack the target-required summary field, or convert them to plain text/context items the target accepts — same class of fix as the existing attachment handling on model switch (#188). Alternatively, warn before switching when the history contains incompatible blocks, or offer one-click "compact & switch".

Workaround (for other users hitting this)

Do not switch to the new model mid-session — start a new task/session directly on the target model. If you must switch, compact/summarize or rewind first to drop raw reasoning blocks.

Logs

Relevant log lines and request IDs above; happy to provide the full excerpt or a diagnostic archive on request.

Start with resources/glm/zcode.cjs, especially the request path around runStreamText and error handling in handleStreamErrorEvent. Reproduce the mid-session model switch using the provided history and logs, then trace how reasoning items enter the Responses input; done means the target model accepts the request after switching without breaking existing history behavior.