feat: add experimental compaction prompt and preserve prefix support

[dixoxib]

Fixes #11497

Summary

• Add OPENCODE_EXPERIMENTAL_COMPACTION_PRESERVE_PREFIX flag to preserve agent prefix cache during compaction
• Add OPENCODE_EXPERIMENTAL_COMPACTION_PROMPT environment variable for customizable compaction prompts
• Update compaction logic to use detailed session summary format when no custom prompt is provided

Why Prefix Preservation Matters

When enabled, OPENCODE_EXPERIMENTAL_COMPACTION_PRESERVE_PREFIX reuses the original agent's prefix cache (tools + system prompts) instead of switching to the generic compaction agent. This provides significant benefits for providers that implement prefix caching:

1. Cache Hit Optimization: Providers like Anthropic Claude and OpenAI GPT-4 use prefix caching to reduce token usage cost and latency. By preserving the agent prefix during compaction, context generates 99% cache hits instead of 100% cache miss tokens, preventing new cache build up, improving performance (especially with slower local LLM) and reducing costs up to 90%. (https://platform.openai.com/docs/guides/prompt-caching).

2. Consistent Tool Context: Maintaining the same tool definitions and system prompts ensures the LLM continues with the exact same capabilities and behavior context, avoiding context-switching overhead.

3. Seamless Continuation: The session continues with identical agent characteristics, preserving specialized instructions or model-specific optimizations from the original session.

Changes

1. Flag additions: Two new experimental flags for compaction configuration
2. Prefix cache preservation: When enabled, compaction reuses the original agent's prefix cache (tools + system prompts)
3. Customizable default prompt: Option for individual session summary prompts adapted to project workflow
4. API export: Made SessionPrompt.resolveTools exportable for reuse in compaction

Testing

Enable with:

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

Note

This is an experimental feature behind feature flags, allowing for gradual rollout and user feedback before potential stabilization.

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search results, I found one potentially related PR that might be addressing a similar compaction issue:

PR #11453 - fix(opencode): prevent context overflow during compaction
#11453

This PR appears to be related to compaction and context handling, which is closely connected to the new compaction prompt and prefix preservation features in PR #11492.

However, the other compaction-related PRs (#4710, #4616, #7104, #7824) appear to be older and addressing different aspects of the compaction feature (token metadata, context arguments, hybrid pipelines, branching).

PR #11492 is the current PR you're checking, so it's excluded from the duplicate consideration.