[FEATURE] Reduce initial system prompt token overhead (~68k tokens before first user message) #26661

Problem

In a fresh OpenCode session with a standard setup (Sisyphus agent, ~100 skills, 50+ tools), the token count before the user sends a single message is approximately 68,000 tokens. This is ~99% system overhead.

Breakdown:

• Skill descriptions: 100+ skills with full descriptions inlined in system prompt
• Tool JSON Schemas: 50+ tool definitions (Playwright, AST, Spec-workflow, etc.)
• Agent behavior instructions: Full Phase 0-3 orchestration rules
• Using-superpowers skill: Injected inline at session start

For comparison, Claude Code's system prompt is ~10-15k tokens for the first message.

Impact

With a typical 128k context window:

• Effective conversation space: ~60k tokens (less than half the window)
• Faster context exhaustion → more frequent compressions → potential loss of context
• Higher API costs per conversation turn (system prompt sent every time)
• Shorter productive sessions before hitting context limits

Proposed Solutions

1. Lazy-load skill descriptions

Currently all 100+ skill names + descriptions are inlined. Only inject skill names initially, and load full descriptions on-demand when skill tool is invoked. This could save ~10-15k tokens.

2. Tool schema compression

• Only inline tool schemas for tools the current agent type is likely to use
• Or use abbreviated schemas with a "full schema on demand" mechanism
• Some tools (e.g., full Playwright suite with 20+ methods) are rarely all used in one session

3. Agent instruction modularization

The Sisyphus Phase 0-3 behavior instructions are verbose. Consider:

• Loading detailed instructions only when the agent type is Sisyphus
• Providing a compressed version for simpler agent types (Sisyphus-Junior, explore, etc.)
• Subagents already have scoped instructions; the orchestrator pays the full cost

4. Skill content injection optimization

The using-superpowers skill content is injected with <EXTREMELY_IMPORTANT> tags at session start. Consider deferring this until a skill is actually matched/needed.

Related Issues

• [FEATURE]: Implement Dynamic/Lazy Loading for MCP Tool Schemas to Prevent Context Bloat #17482: Lazy loading for MCP tool schemas
• [FEATURE]: Move git/PR instructions out of bash tool description to save ~1.7K tokens per request #21345: Move git/PR instructions out of bash tool descriptions

Environment

• OpenCode v1.14.45
• Windows 11
• Agent: Sisyphus with ~100 installed skills