[Bug] Switching models mid-session can trigger false pi-ai context overflow

[iiijiashu]

Summary

Switching models inside an existing long-running DSH session can immediately produce:

pi-ai detected context overflow for model "claude-opus-5"

while the original model can still use the same session history.

Reproduction

1. Continue a DSH session with model A until it has substantial context.
2. Confirm model A still works.
3. Switch the same session to model B (observed with claude-opus-5).
4. Send the next turn.
5. DSH reports a pi-ai context overflow.

The important part is that this is the same session/history; only the selected model changes.

Actual behavior

The same session surface is accepted by the original route, but after a model switch the target route is re-evaluated against a newly resolved contextWindow. If that value is missing, fallback, or too small, pi-ai can locally classify the request as CONTEXT_WINDOW_EXCEEDED.

The error text comes from the pi-ai-side overflow classification path rather than necessarily from the upstream provider.

Expected behavior

On model switch, DSH should resolve the target route's real effective context window before the request. If the current surface does not fit, it should explicitly compact or warn before sending. A catalog miss or fallback should not silently create a misleading local overflow when the provider/model actually supports a larger window.

Suspected cause

This appears closely related to #7213. The model-switch path makes the issue more visible: the history is unchanged, but llm-pi-ai uses the switched-to model metadata for isContextOverflow(...).

If the new route resolves to a fallback/default (for example 262144) or otherwise incorrect value, the same conversation can be classified as overflow immediately after switching models even though the previous route worked and the provider may support a larger context.

Suggested improvements

• Show the resolved contextWindow and its source in diagnostics (explicit model config / provider default / built-in catalog / fallback guess).
• Perform a context compatibility check when changing models.
• Compact before the first request when the target window is genuinely smaller.
• Avoid silently using a guessed fallback window for local overflow classification when route metadata is missing.
• Include measured input/context usage and the resolved window in the error message.

Environment

• DSH: 0.1.5-rc.1
• Observed switched-to model: claude-opus-5
• Existing long-running session, model changed in-place
• No credentials or provider secrets included

[ASCLL0100]

DeepSeek Harness

版本 V0.2.1-rc.2

也同样有这个问题

[TangQi001]

Me too, I switch my model and get a same issue.