[FEATURE]: propagate W3C trace context (traceparent/baggage) in MCP tools/call _meta per SEP-414 #46856

Activity

1. github-actions

   assigned

   nexxeln

   on Sep 2, 2026
2. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed request, @Yagi-Michael!

   To restate it: when OpenCode calls an MCP tool, it would also send the W3C trace context of the active tool-call span in CallToolRequest.params._meta. That means traceparent, optionally tracestate, and baggage, as reserved by SEP-414. MCP servers, including stdio ones, could then parent their spans onto OpenCode's span.

   Right now callTool in packages/core/src/mcp/client.ts only adds _meta["ai.opencode/sessionID"]. A plugin can't add these keys either: tool.execute.before can only change tool arguments, not the MCP request's _meta. That's why your workaround has to go through args._otel.

   Related:

   • feat(core): pass session context to MCP tools #46008 / [FEATURE]: Pass the active session ID to MCP tool calls via _meta #45997: added _meta.sessionID on MCP tool calls
   • [FEATURE]: propagate W3C trace context on outbound MCP HTTP/SSE requests #28295: traceparent as an HTTP header for remote MCP transports (closed for inactivity)

   Maintainers will look at this and decide whether to take it on.
3. tiennv147 commented on Nov 6, 2026

   

   tiennv147

   on Nov 6, 2026

   More actions

   +1. Here is a reproduction on 2.1.2 with an MCP gateway (LiteLLM 1.105.0), and why gateways also need the HTTP header, not only _meta. Checked again on v2.2.5: MCP.callTool still adds no attributes, and the MCP transport still doesn't inject per-call context.

   Setup: opencode run with OTEL_EXPORTER_OTLP_ENDPOINT set. Remote MCP server = LiteLLM's MCP gateway (/mcp/, streamable HTTP), Code Mode on. LiteLLM exports its own OTel v2 spans to the same backend.

   What happens: the model call links up, but the MCP call doesn't.

   trace A (opencode)                                  trace B (litellm), new root, no parent
   SessionStep.attempt                                 POST /mcp                              10.48 s
   ├─ http.client POST ──► POST /v1/chat/completions   └─ tools/call deepwiki-ask_wiki_question 10.47 s
   │                       └─ chat claude-haiku-4-5        └─ MCP send tools/call ask_wiki_question
   └─ Tool.execute (execute)
      └─ Tool.execute
         └─ MCP.callTool                     10.50 s   ◄── same call, separate trace
   

   • LLM requests go through Effect's HttpClient, which injects traceparent from the current span, so LiteLLM's POST /v1/chat/completions nests under http.client POST. 👍

   • MCP requests go through the MCP SDK transport with McpOAuth.loggedFetch (packages/core/src/mcp/oauth.ts). It calls the global fetch inside Effect.runPromiseWith(context captured in connect()), so:

     1. no traceparent header is added at all, and
     2. even if propagation were added inside loggedFetch, it would use the connection-time context, not the MCP.callTool span's.

     The result is that every MCP HTTP request starts a fresh trace on the server.

   • tools/call already carries _meta: {"ai.opencode/sessionID": …} (thanks for feat(core): pass session context to MCP tools #46008). The trace context could ride along in the same place.

   Why _meta alone isn't enough for gateways: LiteLLM 1.105 parents its MCP span to its own POST /mcp HTTP server span, which takes its parent from the traceparent header. A traceparent in params._meta (SEP-414) is recorded only as a span link, by design (litellm/integrations/otel/plumbing/context.py). So:

   • _meta (this issue) → a stdio server can correlate, and a gateway gets a link.
   • HTTP header ([FEATURE]: propagate W3C trace context on outbound MCP HTTP/SSE requests #28295) → a remote server or gateway becomes a child in the same trace.

   Both would be ideal: a header for remote transports, _meta for all.

   Suggested fix, along the lines of PR #28299:

   • Capture the current span when callTool / readResource / getPrompt runs.
   • Inject W3C traceparent / tracestate from it into the request headers. For example, pass per-request headers through the transport's requestInit, or have the fetch wrapper read the span from an AsyncLocalStorage set around the SDK call instead of the connect-time context.
   • Add the same values under params._meta per SEP-414.

   Happy to test a build against LiteLLM's MCP gateway.

   Related: the Tool.execute and MCP.callTool spans also lost their tool.name / tool.call_id attributes in v2, so even inside OpenCode's own trace you can't see which tool ran. Filed as #62305.