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.
CodeGraph and MCP operational readability variants
Consultant variants
Variant 1: In-place surface augmentation
Approach: Extend the two existing modules directly without adding new surfaces. codegraph.ts gains Cargo-workspace detection that enriches the existing code_graph CheckResult with crate count and workspace-membership readability (reading Cargo.toml [workspace] members and, when the partner CLI is present, passing through its status), while communities.ts gains an opt-in batchSize option that chunks label propagation and cluster-note materialization with per-batch failure isolation. Defaults stay unset, so doctor output and community detection remain byte-identical until a flag is used.
Trade-offs:
- Smallest footprint and fastest to ship; no new CLI verbs, MCP tools, or schema tokens to document and version.
- Couples two unrelated concerns (partner readability, graph robustness) onto pre-existing functions, and Rust readability stays buried inside
o2b doctoroutput rather than being independently queryable. - Partial-failure resilience added inside
detectCommunitiesrisks diluting its current deterministic, fully-in-memory contract that downstream tests rely on. Complexity: small Risk: low
Variant 2: Dedicated operational-readability reporting layer
Approach: Add a new read-only surface (o2b partner codegraph report plus a brain_codegraph_report MCP tool) that aggregates partner code-graph status, detected Cargo-workspace crate membership, and any partner-reported crate dependency counts into one structured, schema-versioned report with honest absent-data fields and typed errors. Community labeling robustness is delivered as a separate, option-gated batched pass over the existing detection, leaving the default path untouched. Each concern lives in its own discoverable surface.
Trade-offs:
- Operational readability becomes a first-class, queryable artifact for agents and operators rather than a doctor side effect, matching the cards' stated intent.
- Honest-absent-data and typed-error reporting fit the project's existing conventions and stay additive, so existing reads are byte-identical.
- More new surface area to register, document, changelog, and version-sync than Variant 1; two cards land as two distinct surfaces rather than one cohesive change. Complexity: medium Risk: low
Variant 3: Unified batched-graph-pass abstraction with crate edges
Approach: Introduce a shared "batched graph pass with partial-failure accumulation" primitive consumed by both subsystems: community detection runs chunked, fault-isolated batches, and a new crate-graph projection extracts Graphify-style crate_depends_on edges from detected Cargo workspaces into Open Second Brain's graph schema as a new typed edge. This generalizes batching as a cross-cutting concern and brings Rust workspace dependencies in as first-class graph structure.
Trade-offs:
- Maximizes reuse and treats batching/partial-failure as one well-tested primitive across the graph layer.
- Adds a new
crate_depends_onedge type to the graph schema, a non-additive change that touches indexing, recall, and schema-pack surfaces, raising migration and regression surface. - Presupposes deeper Graphify-style edge mapping the constraints caution against, and risks recasting Open Second Brain's optional partner as an owned ingestion dependency. Complexity: large Risk: high
Recommended: Variant 2
Rationale: The cards are explicitly framed as operational readability rather than new graph machinery, so a dedicated read-only report surface delivers their value where operators and agents can find it, without burying it in doctor output as Variant 1 does. It stays additive and option-gated with typed errors and honest absent-data, preserving byte-identical default behavior per the constraints. It deliberately avoids Variant 3's new crate_depends_on schema edge and Graphify-as-dependency overreach, keeping the partner integration detection-only. The modest extra surface area is the right cost for a discoverable, versioned artifact that fits existing conventions.
Final decision
Chosen variant: Variant 2, Dedicated operational-readability reporting layer.
I agree with the consultant recommendation. It best matches the locked scope by improving operator and agent visibility through additive report surfaces, while keeping partner integration optional and read-only. It also prevents the release from overreaching into new graph schema edges or LLM community labeling, both of which would increase risk beyond the two triage cards.