claude-code - 💡(How to fix) Fix "Can't rewind to this message" after long session with multiple MCP disconnect/reconnect cycles

Summary

After a long-running session (estimated 12+ hours of active use) with multiple MCP server disconnect/reconnect cycles, the rewind feature surfaced a yellow "Can't rewind to this message." toast for a message that was clearly visible in the transcript UI. There was no indication beforehand that the message wasn't rewindable.

Environment

• Platform: macOS
• Working directory: a large monorepo with several MCP servers configured (Figma, code-review-graph, scout, bulletin-board, swc-docs, spectrum-design-data, Brave AppleScript, computer-use, and others)
• Session age at error: approximately 12+ hours of active use across many tool calls and file edits

MCP server churn during this session

At least 3 disconnect/reconnect events observed:

• mcp__Figma__* — disconnected and reconnected multiple times (one event dropped 3 tools, a later event dropped 4 tools; both sets eventually came back)
• mcp__code-review-graph__* — disconnected with 30 tools dropped, later reconnected
• A final reconnect surfaced 34 deferred tools at once (30 code-review-graph + 4 Figma)

Message being rewound to

A message earlier in the session containing a research prompt I had drafted. The visible text in the rewind-target area included phrases like "a recommendation tuned to your actual use case." The message was unambiguously present in the transcript view; it was roughly ~10 turns back from the current turn, well inside the visible scrollback.

Expected

Either (a) rewind succeeds, or (b) the UI marks not-rewindable messages distinctly (greyed rewind affordance, tooltip) so the user doesn't attempt an action that will fail.

Actual

Yellow toast: "Can't rewind to this message." No upfront indication of which messages are rewindable. No actionable error context.

Likely root cause (speculation)

One or more of:

1. Rewind index gets corrupted/truncated when MCP servers reconnect mid-session — state sync between the client transcript view and the rewind buffer drifts.
2. Auto-compression has fired silently on prior turns, removing them from the rewindable buffer while leaving them visible in the UI.
3. Finite rewind window with no UI surface for the boundary.

Repro suggestion

1. Start a Claude Code session in a project with multiple MCP servers (3+ servers, including at least one that occasionally disconnects)
2. Use the session actively for several hours through many turns and tool calls
3. Have at least 2 MCP disconnect/reconnect events occur naturally during the session
4. Attempt to rewind to a message ~10–20 turns back

This is likely intermittent and may be hard to reproduce deterministically.

Severity / impact

Low-to-medium. The toast is dismissible and doesn't crash anything, but the workaround is "restart the session," which loses in-flight scratchpad state. For long, expensive sessions this is real friction.