feat(mcp): Phase 4 — oneshot pipeline workflows via MCP

[staging-devin-ai-integration]

Summary

Follow-up from #358 (Phase 1 landed in #359).

Phase 4 extends MCP to support oneshot (batch processing) pipelines, enabling AI agents to submit files for processing and retrieve results.

Proposed tools

• run_oneshot — submit a pipeline YAML + input file(s) for batch processing, return the output
• list_oneshot_jobs — list active/recent oneshot jobs
• get_oneshot_status — poll job status and retrieve results

Design considerations

• Oneshot pipelines use streamkit::http_input / streamkit::http_output synthetic nodes
• Input data needs to be passed as base64 or multipart within the MCP tool call
• Results should be returned inline for small outputs, or as MCP resources for large outputs
• Must respect max_concurrent_oneshots limits
• File-path security checks still apply

Dependencies

• Phase 1 MCP infrastructure (feat: implement MCP server #359)

[staging-devin-ai-integration]

Closing: the original framing had media flowing through MCP tool calls, which doesn't make sense — MCP is control plane, not data plane. Refiling with a corrected scope.