Experimental compaction improvements (prefix preservation + prompt customization)

[dixoxib]

Problem

When sessions are compacted, the system currently switches to a generic "compaction" agent, losing:

1. Original agent's tool definitions and system prompts
2. Specialized instructions or model-specific optimizations
   → 3. Prefix cache benefits for providers that implement caching (Claude, GPT-4, etc.)

Proposed Solution

Add experimental feature flags for:

1. OPENCODE_EXPERIMENTAL_COMPACTION_PRESERVE_PREFIX - Preserve original agent's prefix cache during compaction
2. additionally OPENCODE_EXPERIMENTAL_COMPACTION_PROMPT - Allow customizable compaction prompts

Benefits

• Performance: Cache hits for providers with prefix caching reduce token usage costs and latency up to 90%
• Flexibility: Users can customize compaction prompts for their workflow
• Gradual rollout: Experimental flags allow testing before stabilization

Testing

export OPENCODE_EXPERIMENTAL_COMPACTION_PRESERVE_PREFIX=true
export OPENCODE_EXPERIMENTAL_COMPACTION_PROMPT="Custom prompt if needed"

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Auto-compaction causes the subagent to lose its original context, possibly the system prompt as well #6535: Subagent loses original context and system prompt during compaction
• Soft compaction / AI global workspace metabolism #3032: Soft compaction / AI global workspace metabolism - proposes alternative compaction strategy to preserve context
• [FEATURE] Sliding window context management for long-running sessions #4659: Sliding window context management for long-running sessions - addresses context preservation across compactions
• [FEATURE]: Add systemPrompt option to SDK like Claude Agent SDK #7351: Add systemPrompt option to SDK - relates to the customizable compaction prompt aspect

Feel free to ignore if none of these address your specific case.