Skip to content

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.

TypeScript
5.5 KB
Raw Blame

You are helping select and then brainstorm the scope of the NEXT release for the Open Second Brain project. Do not write code. Do not write a final design. Your job has two parts: (1) recommend the single most useful RELATED cluster of low-priority triage tasks to ship as one release, and (2) for that recommended cluster, produce architectural variants.

Selection objective

From the triage column below, pick a COHERENT cluster of tasks that:

  • Share a subsystem or a single architectural theme (a "related cluster", not a grab-bag).
  • Are LOW / minor priority on this board. Priority scale here is: 4 = highest, 0 = unprocessed/lowest. Prefer p0-p1 clusters; p2 only if tightly related to the chosen theme.
  • Maximize USEFULNESS to the project. Usefulness is the most important criterion by far. The number of files touched and exact feature count are the WEAKEST signals - do not optimize for them.
  • Are buildable additively, without architectural unknowns that need a separate ADR, and without depending on reverse-engineering an external/unstable data format (treat "needs reverse-engineering of an external transcript/session format" as a blocking risk).

Explicitly AVOID recommending:

  • Large architectural subsystems the board's validator flagged as "needs ADR / design discussion" (PDF export, voice/STT-TTS, whiteboard, 15-adapter architecture, universal agent runtime contract, LLM tracing, inbox-pill UI, subagent orchestration).
  • Thin passthroughs to an external binary that would silently no-op when that binary is absent (e.g. a reindex --force verb that just shells out to an absent graphify) - the project forbids misleading fallbacks.
  • Anything gated on an unmerged upstream PR.

Project context

  • Name: Open Second Brain. Language/runtime: TypeScript on Bun. Storage: SQLite (bun:sqlite, FTS5 + sqlite-vec). Surfaces: an MCP stdio server, an o2b CLI, and an Obsidian/Markdown vault.
  • The kernel is provider-agnostic: it NEVER calls an LLM itself. Any AI extraction must be agent-driven (via an MCP tool or a lifecycle hook); the kernel only stores already-extracted structured data.
  • Hard guarantees the project holds itself to: byte-identical reads when a new feature flag/surface is unused; language-agnostic behavior (NO hardcoded natural-language word lists in any language - use structural signals, frontmatter fields, document-frequency, or agent/LLM extraction); no misleading fallbacks; no hardcoding; no as TypeScript cast crutches.
  • Two relevant registry-driven subsystems already exist and are explicitly designed for easy extension:
    1. src/core/install/adapters/ - writes the Open Second Brain MCP servers (and optional lifecycle hooks) into a third-party coding agent's own config. Present targets: aider, copilot-cli, cursor, gemini-cli, generic, grok, kiro, opencode, pi. A thin adapter is ~25 lines via the shared _json-mcp.ts base; bespoke hook+MCP adapters (grok) are larger. NOTE: the adapter-registration import list is currently duplicated verbatim across 4 CLI entry files (install.ts, init-interactive.ts, uninstall-target.ts, update.ts) - a latent DRY problem.
    2. src/core/brain/sessions/ - session-import adapters that read an agent's transcript files and ingest them. Present: claude, codex, grok, hermes, opencode. Registry self-documents the 3-step add procedure. Adding a runtime here requires KNOWING that runtime's on-disk transcript format.

Recent shipped releases (most recent first):

  • v1.10.0 Recall & Working-Memory Quality - selectable recall profiles, usage decay, co-occurrence auto-relate, file-context recall
  • v1.9.0 Brain Portability & Interop - bank export/import, page contract, brain_create_note, in-process SDK
  • v1.8.0 Indexer Durability & Resilience - cooperative abort, graceful watch shutdown, resumable reindex
  • v1.7.0 Knowledge Provenance - ingest, research, NER, derived facts, owner-scope, standing-query
  • v1.6.0 Vault Integrity & Trust; v1.5.0 Search & Recall Quality; v1.4.0 native Grok integration; v1.3.0 opencode integration + Continuity/Hygiene/Freshness

Constraints for the chosen cluster:

  • Do not change existing public APIs in a breaking way.
  • No new external runtime dependencies unless unavoidable.
  • Every new surface must be byte-identical-when-unused and language-agnostic.

Triage column (every task with body and comments)

<<<TRIAGE [The full triage column - 42 tasks with bodies and comments - was fed in verbatim at runtime via kanban.ts triage. Omitted from the committed audit trail to avoid persisting a board snapshot in the repo; the live board is the source of truth.] TRIAGE

Required output format

Part 1 - Scope recommendation

Candidate clusters

List 2-4 candidate clusters you considered. For each: the cluster name, the task IDs in it, the shared theme, and a one-line usefulness/risk note.

Recommended cluster

  • Name:
  • Task IDs:
  • Why this cluster (3-5 sentences, usefulness-first):
  • Why not the others (1-2 sentences each):
  • Key risks and how to neutralize them:

Part 2 - Architectural variants for the recommended cluster

Produce exactly 3 distinct architectural variants for IMPLEMENTING the recommended cluster. For each variant:

Variant N:

  • Approach: 2-3 sentences.
  • Trade-offs: bullet list of pros and cons.
  • Complexity: small | medium | large
  • Risk: low | medium | high

After the three variants:

Recommended: Variant N

Rationale: 2-3 sentences, considering the project context and constraints above.

Output nothing outside of these sections.