Local-first 🧠memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.
Agent capability + CLI integration - variants
Consultant output
See cli-output/claude.md for the primary consultant output.
Parsed variants
Variant 1: Independent additive layers
Implement the capability filter, inherited --json, and shell completions as three independent additions.
- Complexity: medium
- Risk: low
- Trade-off: low blast radius, but completions and JSON discovery can drift because they duplicate command metadata.
Variant 2: Single declarative manifest spine
Introduce one registry that drives MCP tool capability predicates, CLI flags, help JSON, and completions.
- Complexity: large
- Risk: high
- Trade-off: best anti-drift property, but too much refactor and too much coupling for this scope.
Variant 3: Phased - shared CLI registry now, ADR-gated capability layer
Use a CLI-only manifest for JSON discovery and completions, then add a separate transparent MCP capability filter layered on top of static scope.
- Complexity: medium
- Risk: low
- Trade-off: two registries instead of one, but each owns a smaller and clearer boundary.
Orchestrator decision
Chosen: Variant 3.
The CLI tasks have a real shared abstraction: command/flag metadata. A manifest lets help --json, fallback JSON handling, and shell completions use the same source of truth without forcing the MCP tool registry through a broad refactor. The runtime capability check stays separate, additive, and transparent: it filters after static ToolScope and exposes a report through the MCP surface and probe path so tools are never hidden without a way to inspect why.