Skip to content

Repository files navigation

Grimdex

The Grimoire Index for coding — a tool-agnostic, file-first knowledge base for AI pair programming. One markdown repo holds your programming decisions, rules, and lessons; every AI coding agent you use (Claude Code, Codex, Gemini, GitHub Copilot, Cursor, local models) reads it before working and records what it learns back into it. A scheduled "librarian" routine keeps it "aggressively maintained".

Expectation-setting: this is an early, experimental personal project, published because the pattern is useful — not a turnkey product. Windows + PowerShell 7 first; the conventions are portable, the scripts currently are not.

The idea

  • GRIMDEX.md is the law — the only file agents always read. Kept deliberately tiny. Everything situational lives in universal/playbooks/ and loads only at its moment (starting a project, compacting a session, ending a project).
  • Decisions are the product. Every significant choice becomes a decision record — what was chosen, the alternatives, the reasoning — in that project's tier under projects/<id>/. See examples/.
  • Rules earn their way in. Cross-project rules are proposed into a staging inbox (universal/promotions/). A daily sweep (a headless agent session on a scheduler) conflict-checks candidates against the law and two append-only ledgers: PROMOTIONS-LOG.md (past admissions/rejections) and RIPPEDPAGES.md (everything ever removed, with the why). Clean additions auto-inscribe with a full paper trail; conflicts and removals always wait for the human.
  • Every tool is a contributor. wire-project.ps1 injects a marked pointer block into a repo's CLAUDE.md, AGENTS.md, GEMINI.md, GROK.md, .cursorrules, and .github/copilot-instructions.md, so whatever agent opens the project finds the knowledge base and the contribution rule.
  • Cold-start seeding. Rules normally earn their way in through observed failures — a fresh instance has none yet. SUGGESTED-RULES.md carries seven community-converged conventions (each independently advocated by 3+ prominent practitioners), the reasoning behind each, and a paste-into-your-agent prompt that files them through the normal promotions inbox.
  • Multi-machine consistency. Run one knowledge base across several machines via a hub-and-spoke, single-writer model: a designated hub is the only writer of rules/law; spokes pull read-only and propose edits with scripts/propose-rule.ps1 (filed as a rule-sync proposal in universal/promotions/), which the hub reviews and publishes with scripts/review-rule-proposals.ps1. GitHub is the exchange — no merge conflicts on shared rules by construction.
  • Model labor-split. Route each task to the right worker for your rig — bulk builds to your highest-capacity coder, reviews to your strongest model, docs to your cheapest — via a config/fleet.json policy and a tested selector (scripts/fleet-lib.ps1). Includes a user-invoked multi-model "full counsel" cross-review. See labor-split.md.
  • Learn mode. An optional teaching layer: the agent narrates the why and defines the jargon as it works, so a learner builds real vocabulary. A surfacing ledger (scripts/learn-lib.ps1 + a transcript parser) tapers each term deterministically — full definition, then a reminder, then only occasionally — so it fades instead of nagging. See learn-mode.md.
  • Compaction safety. A PreCompact hook (scripts/precompact-guard.ps1) blocks a conversation compaction while work is uncommitted or unpushed, so context is never compressed before the closeout runs. Fail-safe: it only ever blocks on a confirmed unsaved-work signal. See compact-guard.md.

Where Grimdex sits — the three-layer ecosystem

Grimdex governs how coding work is done. Two sibling layers cover the rest: Grimlore (optional) remembers why/who/context, and Grimdex Baton decides what happens next and executes it. Grimdex resists bloat precisely because those layers exist to take what doesn't belong here.

The authoritative brief — the Law, definitions, boundary tests, promotion policy, and the list of assumptions to discard — is docs/2026-08-14-grimdex-ecosystem-architecture.md. Grimlore's opening design draft is docs/2026-08-14-grimlore-spec.md (spec only — nothing built yet).

Boundary test, in one line: changes how work is done → Grimdex. Explains why/who → Grimlore. Makes something happen → Baton.

Quick start

git clone https://github.com/Ryfter/Grimdex.git
pwsh Grimdex\setup.ps1                      # verify structure, report state
pwsh Grimdex\scripts\wire-project.ps1 -ProjectDir <your-project>
pwsh Grimdex\scripts\install-schedule.ps1   # optional: daily sweep + weekly audit

Your knowledge accumulates in projects/ and universal/ — keep your instance repo private; it will contain your decision history and preferences. This public repo is the engine/template only.

Requirements: PowerShell 7+, git. The scheduled routines additionally expect a headless-capable AI CLI (built against Claude Code); everything else is plain files. Tests: pwsh scripts/test-<area>.ps1 (10 suites).

Layout

See GRIMDEX.md — it is the canonical description, and the file your agents read first.

MIT licensed.

About

The Grimoire Index for coding - a tool-agnostic, file-first knowledge base of programming decisions, rules, and lessons. One markdown store shared by every AI coding agent, aggressively maintained by scheduled sweep and audit routines.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages