feat(opencode): cache-aligned compaction to reuse prefix cache

[lloydzhou]

Issue for this PR

Close #25120

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Compaction currently builds its own LLM request with an empty system prompt, no tools, and a filtered message history. Because the request structure differs from normal chat requests, none of the historical messages hit the provider's prompt cache — they're all charged at full input price.

This PR aligns the compaction request to share the exact same prefix as the main agent loop: same system prompt, same tool definitions, same message serialization. The compaction request becomes indistinguishable from a normal chat request up to the point where the summary instruction is appended. This lets the provider serve [system] + [tools] + [dropped messages] from cache, cutting compaction cost by roughly 90%.

See the design rationale: Cache-Aligned Summarization

Cost example (Claude Sonnet 4 pricing, compacting 45K tokens of history):

	Tokens	Without cache alignment	With cache alignment
System prompt	~2K	Full: $0.006	Cached: $0.0006
Tools	~3K	Full: $0.009	Cached: $0.0009
Dropped messages	~40K	Full: $0.120	Cached: $0.012
Summary instruction	~200	Full: $0.0006	Full: $0.0006
Total	~45.2K	$0.136	$0.014

~90% cost reduction per compaction. The cached portion (system + tools + dropped messages) drops from $3/MTok to $0.30/MTok — only the short summary instruction pays full price.

Conditions for the optimized path (otherwise falls back to original behavior unchanged):

• Main model and compaction model share the same provider + model ID
• The original user message is not a json_schema structured output request

Changes:

• prompt.ts: Extract resolveStreamContext (shared system+tools resolution), make processor optional in resolveTools so compaction can call it without a live message handle, compute resolved context before calling compaction
• compaction.ts: Accept optional resolved context with agent/system/tools/user. When present, skip hidden filtering and stripMedia/toolOutputMaxChars so serialized output matches the main loop exactly. Set toolChoice: "none" to prevent tool execution.

How did you verify your code works?

• bun typecheck passes in packages/opencode
• Fallback path (model mismatch / json_schema) is completely unchanged from current behavior
• Reviewed that serialized [system] + [tools] + [messages] prefix is identical between compaction and main loop when resolved context is used

Screenshots / recordings

N/A — backend-only change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

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

Based on my search results, no duplicate PRs were found. The only PR matching these searches is PR #25100 itself (the current PR being analyzed). While there are related PRs that touch on cache optimization and compaction functionality (such as PR #24842 about caching messages and PR #14743 about Anthropic prompt cache hit rates), they address different aspects and are not duplicates of this specific cache-aligned compaction feature.

No duplicate PRs found

[lloydzhou]

dev模式bun run dev使用mimo-v2.5-pro进行测试，读取sqlite数据库记录显示最后一条 compaction 的 cache.read = 14,656，命中缓存。
缓存命中率 = 14,656 / 17,267 ≈ 84.9%

sqlite3 /Users/xxxx/.local/share/opencode/opencode-local.db "SELECT json_extract(data, '$.modelID') as model, json_extract(data, '$.agent') as agent, json_extract(data, '$.role') as role, json_extract(data, '$.tokens') as tokens, datetime(json_extract(data, '$.time.created')/1000, 'unixepoch', 'localtime') as created FROM message ORDER BY json_extract(data, '$.time.created') DESC LIMIT 20

mimo-v2.5-pro|compaction|assistant|{"total":17267,"input":784,"output":513,"reasoning":1314,"cache":{"write":0,"read":14656}}|2026-05-01 19:27:27
|build|user||2026-05-01 19:27:27
mimo-v2.5-pro|build|assistant|{"total":14870,"input":140,"output":15,"reasoning":59,"cache":{"write":0,"read":14656}}|2026-05-01 19:27:17
|build|user||2026-05-01 19:27:17
mimo-v2.5-pro|build|assistant|{"total":14788,"input":13644,"output":13,"reasoning":107,"cache":{"write":0,"read":1024}}|2026-05-01 19:27:00
|build|user||2026-05-01 19:27:00
mimo-v2.5-pro|compaction|assistant|{"total":1478,"input":939,"output":513,"reasoning":26,"cache":{"write":0,"read":0}}|2026-05-01 19:26:14