Overview
Most AI coding agents start every session from scratch. They forget your project's constraints, skip clarification, and jump straight into editing files. Ged fixes that with a structured workflow and durable memory layer that sits on top of your existing editor.
The agent clarifies your intent before touching code, writes a
spec and task list, implements in bounded slices, and verifies
each slice before moving on. Everything it learns about your
project (architecture decisions, standards, glossary) persists
in a .ged/ directory as human-readable markdown.
Two implementations ship in this monorepo: GedPi for the Pi editor and GedCode for OpenCode. Both share the same workflow philosophy and memory format.
The Workflow
Ask one focused question at a time until behavior, constraints, non-goals, and success criteria are concrete.
Write a real spec, task list, and test plan in
.ged/. No edits until these exist.
Work bounded task slices with test-driven discipline. One slice at a time, verified before moving on.
Run planned checks, clean-context review the diff, then update state so the next session picks up cleanly.
Packages
A batteries-included Pi package with the full Ged workflow always active. Drop it into any project and the agent starts clarifying, planning, and implementing in bounded slices out of the box.
- Repo map indexes source files, ranks by structure and recent activity, and injects codebase awareness into prompts
- Skill discovery finds and installs relevant skills automatically, creates project-local skills when none exist
- Bundled extensions include web search, native micro-UI via Glimpse, git diff review, prompt-template commands, and theme support
- Subagent mode delegates to read-only scouts, planners, and verifiers while the main brain stays the sole writer
- Auto-updater checks for new versions on startup, prompts to install and restart
npm install -g gedpi
# run in any project
cd your-project && gedpi
Give your coding agent a memory and a process. GedCode layers the Ged workflow on top of OpenCode so sessions stop forgetting and the agent stops charging ahead without a plan.
- Plan-before-edit guard blocks file edits until SPEC.md, TASKS.md, and TESTS.md have real content
- Durable project memory specs, tasks, tests, decisions, standards, and session summaries persist on disk between runs
- TDD discipline behavior-changing slices go through red-green-refactor, recorded in TESTS.md
- Disciplined diagnosis bugs route through reproduce, minimize, hypothesize, instrument, fix, regression-test
- Token savings RTK compresses bash output by 60-90% automatically
- Zero-impact isolation runs in its own config sandbox, your normal OpenCode setup is untouched
curl -fsSL https://raw.githubusercontent.com/edgyarmati/ged-mono/main/install.sh | bash
# then run
gedcode
Durable Memory
Both GedPi and GedCode use a three-tier memory architecture
under .ged/. All memory is project-scoped and
stored as human-readable markdown you can edit, review, and
commit alongside your code.