Copilot CLI 1.0.97 rechecks cached MCP tool definitions: Sume

Copilot CLI 1.0.97-0 fixes a case where a cached MCP tool call could run against an out-of-date definition: "Cached MCP tool calls recheck their definition when a concurrent tool-list refresh changes the catalog." With Sume, the practical advice is to let the agent read tools_schema for a tool before its first call in a session, and to treat a changed tool list as a reason to read it again.

Sume's hosted MCP server is at https://mcp.sume.com/mcp. These notes cover only what the docs and the release say.

What is a cached MCP tool call?

The release does not define the term beyond that sentence, so this post does not guess at internals. What it says is enough: Copilot holds a tool call prepared from a definition, the tool list can be refreshed at the same time, and when the catalog changes the definition is checked again.

For a server like Sume, whose visible tools depend on your credential, a refresh is not exotic. It happens when you reconnect with a different scope.

Why does the Sume tool list change between sessions?

The MCP overview and the tools and gates page describe access by credential. An OAuth token with only mcp:read sees read-only tools. A token with mcp:read and mcp:write sees the full hosted toolset. An API key is equivalent to full write access.

So the same server shows different catalogs to a read-only sign-in and a write sign-in. Reconnect after changing Write at consent, and the list changes.

How do I keep arguments valid for a Sume tool?

Sume ships three discovery tools: tools_list shows what your credential can reach with safety metadata, tools_schema returns a specific tool's contract, and mcp_health checks readiness and auth state. The docs' Playbook B tells an agent to call tools_schema, then generation_admission_preview or dry_run=true, then submit with a fresh idempotency_key.

• Read tools_schema for the exact tool before its first call in a session.
• After any reconnect, call tools_list again and compare it to what the agent planned to use.
• Do not copy arguments from an older session's transcript; the schema is the source of truth.

What about tools that are not in the list at all?

The docs state that hosted MCP does not have full parity with the HTTP API. The Sume Image 1.0 and Video 1.0 calls (images_create and videos_create) are REST-only. If an agent cannot find a tool, check catalog_list and the REST docs before concluding the connection is broken.

How do I confirm the fix on my machine?

Connect Sume with a read-only sign-in and list the tools with tools_list. Note the count and the names. Reconnect with Write enabled at consent and list again; the paid tools should now appear. In a single long Copilot session, ask the agent to call tools_schema for generate_image both times and compare the contract it reports.

If a call fails after a catalog change, the error will usually be a schema or scope problem rather than a network one. A missing Write scope returns insufficient_scope, which tells you to reconnect rather than retry. Keep the retry rule in mind for paid calls: reuse the same idempotency_key for the same request, and use a new key only when you change the request.

The release notes give no timing for how often the tool list refreshes, so do not build on an assumption about it. Reading the schema before the first call in each session costs one read and removes the question.

• Check 1: count tools under Read only, then under Read and Write.
• Check 2: call tools_schema for one generation tool after each reconnect.
• Check 3: run dry_run=true before the first paid call of the session.