[FEATURE] Allow to resume interrupted generation without creating a new conversation turn

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When a response is interrupted (usage limit, network failure, etc.), the only recovery is to add a new user turn (using existing --continue or --resume and then writing an additional user message such as "continue"). This changes the conversation semantics and relies on the model inferring that it should continue the previous response, while also needlessly wasting a small amount of tokens.

Proposed Solution

Add a command (e.g. claude --resume-generation) that reconstructs the KV cache by replaying the transcript and partial assistant output, then continues the interrupted assistant turn without adding a new user message.

Alternative Solutions

_No response_

Priority

Medium - Would be very helpful

Feature Category

API and model interactions

Use Case Example

1. Usage limit interrupts conversation
2. Network failure
3. Accidentally closed claude terminal

Additional Context

_No response_