This document outlines the planned development roadmap for gstack, organized by development phase. It tracks the evolution from a headless browser utility into a comprehensive suite of 30+ specialized cognitive modes for AI agents.
The roadmap reflects the strategic direction of gstack: providing specialized cognitive modes for AI agents like Claude Code, Gemini CLI, and Codex. gstack is currently in Phase 7 (v1.49.0.0), focusing on Memory Management, Fidelity Gap reduction, and Cross-Model Determinism.
Sources: TODOS.md1-20 package.json3
Phase 7 (Current) addresses the "Fidelity Gap"—ensuring models like Codex and Claude follow complex multi-step instructions with 100% reliability—and a major push for Browser Memory Leak Prevention surfaced during engineering reviews.
Sources: TODOS.md3-9 package.json3
Phase 1-6 established the core architecture, the skill library, and the "Fix-First" heuristic for automated code reviews.
| Feature | Status | Implementation Location |
|---|---|---|
| Fix-First Heuristic | ✅ Complete | review/SKILL.md.tmpl120-145 |
| Base Branch Detection | ✅ Complete | ship/SKILL.md.tmpl31-33 |
| Test Failure Triage | ✅ Complete | ship/SKILL.md.tmpl164-168 |
| Distribution Pipeline Check | ✅ Complete | ship/SKILL.md.tmpl100-125 |
| Slop Scan (Advisory) | ✅ Complete | review/SKILL.md.tmpl77-89 |
| Skill Template System | ✅ Complete | scripts/gen-skill-docs.ts1-10 |
The system now utilizes a sophisticated template pipeline that injects shared behaviors (like {{PREAMBLE}} and {{TEST_BOOTSTRAP}}) into skill-specific templates. The gen:skill-docs script package.json16 is responsible for regenerating SKILL.md files from templates.
Sources: ship/SKILL.md.tmpl29-144 package.json16
Phase 7 focuses on memory efficiency in the gbrowser infrastructure and the nuances of AI interaction.
Planned Features:
| Feature | Priority | Implementation Detail |
|---|---|---|
| MV3 SW Memory Profile | P2 | Call Target.getTargets to profile extension service workers TODOS.md10-19 |
| GPU Memory Breakdown | P2 | Surface SystemInfo.getInfo for GPU and Memory.getDOMCounters TODOS.md39-48 |
| Architectural Cleanup | P3 | Replace per-page listeners with Target.setAutoAttach browser-wide handler TODOS.md69-80 |
| Peak-RSS Reproducer | P3 | Real-Chromium test with 500 concurrent fetches to verify leak fixes TODOS.md102-116 |
| Context Lanes | P3 | /context-save --lane A for parallel workstreams TODOS.md5-15 |
| Pacing Overhaul | P0 | Reduce "interruption volume" via phase field in logs TODOS.md37-50 |
| Diff-based Testing | ✅ Shipped | E2E_TOUCHFILES mapping for selective test execution test/helpers/touchfiles.ts34-40 |
Sources: TODOS.md1-135 test/helpers/touchfiles.ts1-106
The roadmap is guided by Cognitive Load Reduction and Execution Integrity.
Key Technical Decisions:
| Decision | Rationale | Implementation Evidence |
|---|---|---|
| Diff-based selection | Drastically reduce CI time by only running tests for modified components | test/helpers/touchfiles.ts4-7 |
| Non-Interactive Ship | /ship means "DO IT" — only stop for critical gates like merge conflicts | ship/SKILL.md.tmpl37-41 |
| Memory Attribution | Ensure diagnostics capture memory where 160GB leaks actually live (Native/GPU) | TODOS.md50-55 |
| Idempotency | Re-running /ship runs all verifications but skips already-completed actions | ship/SKILL.md.tmpl62-70 |
Sources: test/helpers/touchfiles.ts1-26 ship/SKILL.md.tmpl35-70 TODOS.md39-55