Open GSD — AI Coding Agent Meta-Framework That Beats Context Rot
Open GSD — AI Coding Agent Meta-Framework That Beats Context Rot
You've been there. You start a session with Claude Code or Codex, explain your architecture, write some solid code. Three hours and 50 tool calls later, the agent starts hallucinating — forgetting what file it just edited, suggesting things you already tried, and losing the plot entirely. That's context rot: the quality degradation that accumulates as an AI agent fills its context window with conversation history.
Open GSD (Git. Ship. Done.) fixes this. It's not another AI coding agent — it's a meta-framework that wraps existing agents (Claude Code, Codex, Cursor, Copilot, Cline, Gemini CLI, and more) in a disciplined phase loop. Each phase runs heavy research, planning, and execution work in fresh-context subagents while keeping your main session lean. The result: no context rot, structured verification, and code that actually ships.
Since its release in late May 2026, Open GSD has garnered over 6,200 GitHub stars and an active Discord community. It's MIT-licensed and works with any major AI coding runtime.
Why It's Trending
- Context rot is solved — every subagent starts with a clean 200K-token context window (up to 1M for supported models). No accumulated session junk degrading quality.
- Universal runtime support — works with Claude Code, Codex, Cursor, Copilot, Cline, Windsurf, Gemini CLI, Kimi CLI, Kilo, OpenCode, Antigravity, Trae, Augment Code — any runtime where AI agents run.
- Structured phase loop — Discuss → Plan → Execute → Verify → Ship. Each step catches a specific failure mode that the previous step can't prevent.
- File-based state — all project state lives in
.planning/as Markdown and JSON. No database, no server. Survives context resets and spans multiple sessions. - 33 specialized agents — researchers, planners, executors, verifiers, debuggers, security auditors, UI auditors — each with focused prompts and tool permissions.
- Spec-driven development — requirements → research → plans → execution → verification pipeline. Every phase has clear acceptance criteria.
- Active community — Discord, growing plugin ecosystem, active development with frequent releases.
Architecture Overview
The architecture follows a layered design. At the top, commands (/gsd-* slash commands or $gsd-* skill invocations) serve as user entry points. These trigger workflows — orchestration .md files that load context, spawn specialized agents, and manage state transitions. The CLI Tools Layer (gsd-tools.cjs) provides utility commands for context loading, model resolution, state management, and plan verification. Everything reads and writes to the .planning/ File System — the persistent state layer that survives session resets.
Specialized agents fall into several categories: Researchers (4 parallel — stack, features, architecture, pitfalls), Synthesizers (combines research into summary), Planners (decompose into bounded work units), Checkers (verify plan correctness), Executors (implement plans in parallel waves), Verifiers (confirm acceptance criteria met), and Shippers (create PRs, archive phases).
The phase loop works through .planning/ which carries state across the loop — CONTEXT.md from Discuss, RESEARCH.md and PLAN.md from Plan, SUMMARY.md from Execute, VERIFICATION.md from Verify — all persisted for cross-session access.
Prerequisites
- Node.js 18+ (for the
npx @opengsd/gsd-coreinstaller) - One of the supported AI coding runtimes: Claude Code, Codex, Cursor, Copilot, Cline, Windsurf, Gemini CLI, Kimi CLI, Kilo, OpenCode, or Antigravity
- Git (for version control and PR creation)
- A terminal emulator with sufficient context window support (recommended: 200K tokens minimum)
Installation
GSD Core installs with a single command:
npx @opengsd/gsd-core@latest
The installer prompts you for:
- Your runtime (Claude Code, Codex, Cursor, Copilot, Cline, etc.)
- Whether to install globally or locally in the current project
For runtimes without Node.js, see the install-on-your-runtime guide.
Starting a New Project
Once installed, initialize a greenfield project:
cd my-new-project
/gsd-new-project
The command walks you through a structured Q&A session (based on questioning.md philosophy) to capture your idea, then spawns 4 parallel researchers (stack, features, architecture, pitfalls), a research synthesizer, and sets up your .planning/ directory.
Onboarding an Existing Codebase
For brownfield projects:
cd your-existing-repo
/gsd-onboard
This generates a codebase map (technology stack, architecture patterns, conventions, testing setup, known concerns) and creates your initial .planning/ structure.
The Phase Loop in Practice
Each milestone repeats the same five-step loop:
1. Discuss
/gsd-discuss "Add Redis caching layer to the API gateway"
The Discuss step captures implementation decisions before planning begins. Which Redis client library? What cache invalidation strategy? Per-route or global TTL? Edge cases? Output: CONTEXT.md in the phase directory.
2. Plan
/gsd-plan
Runs a sequence of fresh-context subagents: a researcher investigates the ecosystem and records findings in RESEARCH.md, a planner reads both research and CONTEXT.md to produce PLAN.md files, and a plan-checker verifies plans are complete, consistent, and within scope.
3. Execute
/gsd-execute
Plans are grouped into dependency waves. Executors in the same wave run in parallel with fresh 200K-token contexts. Each executor has:
- The specific
PLAN.mdto execute - Project context (
PROJECT.md,STATE.md) - Phase context (
CONTEXT.md,RESEARCH.mdif available)
Executors write code and commit atomically. Later waves wait for earlier ones to complete.
4. Verify
/gsd-verify
A verifier agent reads the phase goal, CONTEXT.md decisions, the plans, and execution summaries — then checks that what was built matches what was intended. Produces VERIFICATION.md and generates targeted fix plans if discrepancies exist.
5. Ship
/gsd-ship
Creates the pull request, archives the phase artifacts, and updates STATE.md to mark the phase complete. The loop then begins again for the next phase.
Quick Commands Reference
| Command | Purpose |
|---|---|
/gsd-new-project |
Start a greenfield project |
/gsd-onboard |
Onboard an existing codebase |
/gsd-discuss |
Capture implementation decisions |
/gsd-ui-phase |
Optional UI design step (between Discuss and Plan) |
/gsd-plan |
Research, decompose, and plan |
/gsd-execute |
Execute plans in parallel waves |
/gsd-verify |
Verify built work against phase goals |
/gsd-ship |
Create PR and archive phase |
/gsd-status |
Show current phase and project state |
/gsd-quick |
Quick work below the phase-loop threshold |
Key Concepts
Fresh-Context Subagents
Every subagent spawned by an orchestrator gets a clean context window (up to 200K tokens). This is the core mechanism that prevents context rot. An executor that runs with 180K tokens of accumulated session history is a degraded executor; one that starts clean and reads only what its plan requires operates at full capacity.
Wave Execution Model
During execute-phase, plans are grouped into dependency waves:
- Wave 1: Plans with no dependencies run in parallel
- Wave 2: Plans depending on Wave 1 run after Wave 1 completes
- Wave 3: Plans depending on Waves 1 and 2 run last
This maximizes parallelism while ensuring correctness. Each executor in the same wave touches non-overlapping concerns, with --no-verify commits to avoid hook contention and STATE.md file locking for safe parallel state writes.
File-Based State
All state lives in .planning/ as human-readable Markdown and JSON:
PROJECT.md— Project overview and goalsREQUIREMENTS.md— Feature requirementsROADMAP.md— Milestone and phase trackingSTATE.md— Current position in the loopphases/— Phase-specific artifacts (CONTEXT, RESEARCH, PLAN, SUMMARY, VERIFICATION)config.json— Runtime configuration
This means state survives context resets (/clear), is inspectable by both humans and agents, and can be committed to git for team visibility.
Adaptive Context Enrichment (1M Models)
When the context window is 500K+ tokens (1M-class models like Opus 4.6, Sonnet 4.6), subagent prompts are automatically enriched — executor agents receive prior wave SUMMARY.md files and phase CONTEXT.md/RESEARCH.md for cross-plan awareness, and verifier agents receive all PLAN.md, SUMMARY.md, CONTEXT.md, and REQUIREMENTS.md files.
Use Cases
- Greenfield projects — structured development from idea to shipped product
- Brownfield refactoring — systematic codebase improvements with verification gates
- Debugging complex issues — structured debug sessions with subagent isolation
- Code review pipelines — automated multi-agent review workflows
- Team collaboration — shared
.planning/state committed to git for team visibility
Comparison with Other Tools
| Feature | GSD Core | Claude Code (raw) | Codex (raw) | Cursor |
|---|---|---|---|---|
| Context rot prevention | ⭐ Subagents + phase loop | ❌ None | ❌ None | ⭐ Partial (tabs) |
| Multi-agent orchestration | ⭐ 33 agents | ❌ Single agent | ❌ Single agent | ❌ Single agent |
| Runtime support | ⭐ 10+ runtimes | ⚠️ Claude only | ⚠️ Codex only | ⚠️ Cursor only |
| Cross-session memory | ⭐ .planning/ files |
❌ None | ❌ None | ⚠️ Limited |
| Verification pipeline | ⭐ Built-in verifier | ❌ Manual | ❌ Manual | ❌ Manual |
| Spec-driven development | ⭐ Full pipeline | ❌ None | ❌ None | ❌ None |
| Open source | ⭐ MIT | ❌ Proprietary | ❌ Proprietary | ❌ Proprietary |
Resources
- GitHub: github.com/open-gsd/gsd-core
- Website: opengsd.net
- Documentation: github.com/open-gsd/gsd-core/tree/main/docs
- Discord: discord.gg/mYgfVNfA2r
- npm Package: @opengsd/gsd-core
- Architecture Deep Dive: docs/ARCHITECTURE.md
- Phase Loop Explanation: docs/explanation/the-phase-loop.md
- License: MIT