session.metadata.getContextAttribution / getContextHeaviestMessages fail with "unknown variant memory" when memories are loaded #5325

Summary

Extensions calling session.rpc.metadata.getContextAttribution() or getContextHeaviestMessages() get an error in a session where Copilot Memory is loaded (a <memories> block is present). A snapshot taken at startup, before any messages existed, succeeded in the same session.

Error

Request session.metadata.getContextAttribution failed with message: unknown variant `memory`, expected one of `skill`, `subagent`, `mcp`, `tool` at line 1 column 17
Request session.metadata.getContextHeaviestMessages failed with message: unknown variant `memory`, expected one of `skill`, `subagent`, `mcp`, `tool` at line 1 column 17

Environment

• Copilot CLI 1.0.105, macOS
• Calling extension: https://github.com/kjetiljd/copilot-context-breakdown-plugin

Steps to reproduce

1. Start a session with Copilot Memory enabled and memories loaded.
2. From an extension, call session.rpc.metadata.getContextAttribution() after the first assistant message.
3. Call session.connection.sendRequest("session.metadata.getContextAttribution", { sessionId }) directly. It fails identically, so the typed SDK wrapper is not the cause.

Expected

The calls return entries, as the SDK docs say: entry kind is "not a closed set, tolerate unknown values".

Actual

Both RPC calls fail, so extensions cannot read context attribution while memories are loaded.

Notes

• The accepted list in the error (skill, subagent, mcp, tool) differs from the documented entry kinds (skill, subagent, mcpServer, tool, system, toolDefinition, plugin). It looks like a separate, narrower source enum that does not yet include memory, possibly a rollout in progress.
• The error format (at line 1 column 17) looks like a serde-style deserialization error.
• Running /context directly in the same session did not show the error.
• I found no mention of a fix in 1.0.106-1 or 1.0.106-2 release notes, and no existing issue matching this error.