abstract

A config repo that turns OpenCode into a small academic research lab: a kernel of invariants, a library of skills (writing, literature, experiments, story-keeping, word export, presentations, and more), and a catalog of sidekick agent briefs (internal and external reviewers, literature lookup and review, visual review, nlpatch). One primary session per project orchestrates through files (tickets, reports, memos, models) and flat subagent delegation.

Status: early. See MEMORY.md for the design rationale and decision tail, docs/prompting-2026-10-11.md for the skills-model survey and design, and manifesto.md for the philosophy.

Layout

• kernel.md -- the always-on instructions mounted per project
• skills/ -- the skill library; staging/ holds the immature (lint)
• agents/ -- sidekick briefs (markdown agent files)
• opencode.template.jsonc -- the project mount (instructions + skills + mcp)
• src/cli.ts -- the abstract CLI, init only
• typ2docx/, tools/, lint/ -- the Word converter, the reference-stock patches, the experimental style linter
• repertoire/ -- the convention corpus pipeline and query path
• docs/ -- dev-facing design documents

Usage

Install and link once:

cd /path/to/abstract && bun install && bun link

Then scaffold a research project:

abstract init <dir>   # writes opencode.jsonc, symlinks the sidekick briefs
                      # into .opencode/agents/, scaffolds the notes skeleton

Open the project in OpenCode. The kernel and the skill library mount from the project's config; sidekicks resolve through the symlink farm. Files are memory; sessions are a cache.

Memory (hindsight)

The hindsight plugin is per-machine and manual:

bun x @vectorize-io/hindsight-coding-agents install opencode2

That writes one plugin entry into ~/.config/opencode/opencode.json. Its config is one user-level file, ~/.hindsight/coding-agent.json; merge in the stanza from hindsight.example.json (the memory server URL, optInOnly, and the mapPathToBank entry routing ~/projects/research to the shared research bank). Memory is then inert outside research projects and no per-project env vars are needed. Note: the plugin's autoUpdate re-stages itself via npm; set "autoUpdate": false to stay bun-only.