Context Research Plan

Last Updated: 2026-03-23

Goal

Research how mainstream agents assemble and manage context within a conversation: 1. How tokens are generated and streamed 2. How context (system prompt, messages, tool results) is stitched together before each LLM call 3. How different agents handle token budget, compaction, and truncation

Completed

• [x] Clone open source repos as submodules (Pi, OpenClaw, Gemini CLI, Codex, OpenCode, claude-code-system-prompts)
• [x] Study Pi: core agent loop and context compaction → pi.research.md
• [x] Study OpenClaw: ContextEngine plugin architecture and assemble() flow → openclaw.research.md
• [x] Study Gemini CLI: context management implementation → gemini-cli.research.md
• [x] Study Claude Code: system prompts, compaction, sub-agents (via community extraction) → claude-code-context.research.md
• [x] Study Codex: dual compaction, context manager (open source, Rust) → codex-context.research.md
• [x] Study OpenCode: two-phase compaction, fork/revert, resumable sub-agents → opencode.research.md
• [x] Write per-project research documents (7 total)
• [x] Cross-project comparison → context.summary.md
• [x] Cross-domain findings (Memory × Context) → findings.md

Open Areas

Discussed but not deeply researched

Topic	Status	Notes
Token streaming mechanics	Mentioned in goal #1, not studied	How each agent handles SSE/WebSocket/stdio streaming. Lower priority — this is transport, not context management
Anthropic "Effective Context Engineering" article	Read and referenced	Could do detailed breakdown: official recommendations vs what agents actually implement
Prompt placement empirics	Identified as gap in findings	No agent does A/B testing. AI Muse 18-model benchmark is the closest. Would need own experiments to go further
Knowledge graphs for context	Identified as gap in findings	Graphiti is Memory-side only. No one applies graph structures to context management
Compaction quality measurement	Identified as unsolved problem	No standard metric exists. Could survey academic work on summarization evaluation

Not yet started

Topic	Priority	Notes
Learning (continual learning)	TODO	New research direction. Focus on academic survey + any production systems doing user-level adaptation. See CLAUDE.md
Cross-domain article	Future	Synthesis article covering Memory × Context × Learning findings. Not urgent — deepen understanding first