tech-advisor - Retrospective

What Went Well

Research Orchestration: Using a specialized researcher agent to investigate Linux CLI editing tools was highly effective. It allowed me to focus on high-level integration while the sub-agent handled the “deep dive” into specific tool commands and pros/cons.
Reference Code Integration: Leveraging the existing mcp-genmedia-go repository provided a concrete foundation for the tech stack. It ensured that the advice given to teams was grounded in production-ready Go code rather than just theoretical API specs.
Incremental Refinement: The process of drafting the tech-stack.md and then refining it based on specific feedback from Preston worked smoothly. Each iteration added high-signal technical depth (e.g., specific FFmpeg patterns like two-pass GIFs and standardization).

Initial Context Fragmentation: Information about the project was spread across several files (agents.md, project-brief.md, team-definition.md, etc.). While I was able to synthesize it, a single consolidated “Project Bible” might have saved some initial reading time.
Hardware Ambiguity: Early in the session, the availability of GPU acceleration was unclear. This required an explicit check (ls /dev/nvidia*) to confirm a CPU-only environment, which fundamentally shifted the recommendations for editing tools (favoring lightweight FFmpeg/MLT over heavy Blender renders).

Agent Syncing: Waiting for the editing-tool-researcher to complete its report was a minor bottleneck. While it performed well, the inherent latency of starting and waiting for sub-agents means this pattern is best suited for large, discrete tasks rather than rapid-fire queries.
Read Permissions: I encountered some initial confusion regarding file paths outside the standard /workspace, but was able to resolve this by using run_shell_command instead of standard file-reading tools when necessary.

Go as Primary Language: Decided to mandate Go for the agents’ generative media tooling. The alternative was Python, which is common in AI, but Go was chosen to align with the provided mcp-genmedia-go reference implementations and provide better performance/concurrency for orchestration tasks.
Tool Hierarchy: Recommended MoviePy as the primary tool for Editors (best abstraction for LLMs) but kept MLT (melt) and FFmpeg as recommended alternatives for more complex or performance-critical work.

Standardized Hardware Manifest: Provide a SYSTEM.md or similar file in the root that explicitly lists available hardware (CPU count, RAM, GPU, OS version). This would eliminate the need for agents to “sniff” the environment.
Integrated Tooling Docs: The compositing_recipes.md in the reference toolkit is excellent. Future projects should prioritize having these “recipe” style docs early on, as they are very easy for agents to parse and implement.