feat(tool): forward LLM tool-call id to MCP servers via _meta

[alenkacz]

What

Threads the originating LLM tool-call id (llm.ToolRequestPart.ID) from the agent loop down to the MCP server on every tools/call:

• tool.WithCallID / tool.CallIDFromContext — new context helpers (tool/callid.go).
• Registry.Execute stashes req.ID on the execution context immediately before invoking a tool (tool/registry.go). Harmless for tools that ignore it.
• The MCP client reads it and forwards it in the tools/call request _meta under the domain-namespaced key redpanda.com/llm-tool-call-id, exported as mcp.MetaKeyLLMToolCallID (tool/mcp/tools.go).

Why

Today, when an agent runs LLM and MCP calls through a gateway (e.g. Redpanda's aigw), the gateway's MCP tool-execution span carries the tool name but not the which-tool-call id, so it can't be tied back to the specific model turn that requested it — only grouped by conversation. Forwarding the id lets the gateway stamp gen_ai.tool.call.id on its span, enabling a precise tool-use ↔ invocation join (and lets the transcripts consumer dedup the gateway span against the in-agent execute_tool span by conversation + tool.call.id).

_meta is the MCP protocol's sanctioned extensibility slot: servers that don't understand the key ignore it, and the tool's own arguments are never touched. When the context carries no id (e.g. a direct ExecuteTool call outside the registry), the _meta key is simply omitted — no behavior change.

Tests

• TestWithCallIDRoundTrip, TestRegistryExecuteExposesToolCallID — context plumbing.
• TestExecuteToolForwardsLLMToolCallID — asserts the id reaches the server in _meta when present, and is absent when the context carries none.

Consumer side (separate, in cloudv2)

The matching aigw change (read params._meta → stamp gen_ai.tool.call.id on the MCP span) is a follow-up in the cloudv2 repo; this PR is the producer half and is a no-op until that lands.

🤖 Generated with Claude Code

[blacksmith-sh]

Found 1 test failure on Blacksmith runners:

Failure

Test	View Logs
github.com/redpanda-data/ai-sdk-go/tool/TestRegistry_WebfetchToolWithLLM_Integration	View Logs