不同 Agent 上下文压缩的实现对比

User Message
    ↓
Step 1: LLM 分析问题
    ↓
Step 2: 调用 search_logs
    ↓
Step 3: LLM 阅读日志并发现数据库超时
    ↓
Step 4: 调用 query_metrics
    ↓
Step 5: LLM 综合结果
    ↓
Assistant Message

You are summarizing a conversation to preserve context for continuing work later.

**Critical**: This summary will be the ONLY context available when the conversation resumes. Assume all previous messages will be lost. Be thorough.

**Required sections**:

## Current State

- What task is being worked on (exact user request)
- Current progress and what's been completed
- What's being worked on right now (incomplete work)
- What remains to be done (specific next steps, not vague)

## Files & Changes

- Files that were modified (with brief description of changes)
- Files that were read/analyzed (why they're relevant)
- Key files not yet touched but will need changes
- File paths and line numbers for important code locations

## Technical Context

- Architecture decisions made and why
- Patterns being followed (with examples)
- Libraries/frameworks being used
- Commands that worked (exact commands with context)
- Commands that failed (what was tried and why it didn't work)
- Environment details (language versions, dependencies, etc.)

## Strategy & Approach

- Overall approach being taken
- Why this approach was chosen over alternatives
- Key insights or gotchas discovered
- Assumptions made
- Any blockers or risks identified

## Exact Next Steps

Be specific. Don't write "implement authentication" - write:

1. Add JWT middleware to src/middleware/auth.js:15
2. Update login handler in src/routes/user.js:45 to return token
3. Test with: npm test -- auth.test.js

**Tone**: Write as if briefing a teammate taking over mid-task. Include everything they'd need to continue without asking questions. No emojis ever.

**Length**: No limit. Err on the side of too much detail rather than too little. Critical context is worth the tokens.

You MUST use this format for your response (you may omit sections that aren't applicable). Do not include the <template> tags in your response.

## Objective

- [one or two brief sentences describing what the user is trying to accomplish]

## Requirements

- [constraints, preferences, requirements, and scope boundaries stated by the user, or "(none)"]

## Decisions

- [decisions already made and why, or "(none)"]

## Work State

Break the objective into smaller goals and report which are completed, which are being worked on, and which are blocked.

### Completed

- [goals that have been completed; otherwise "(none)"]

### Active

- [goals currently being worked on; otherwise "(none)"]

### Blocked

- [anything blocking progress, and why; otherwise "(none)"]

## Next Move

1. [ordered list of next actions, or "(none)"]

## Relevant Files

List the files and directories, other than the current working directory, that another agent would need to open to continue this work. Include at most 15, most important first. Do not list every file that was read or changed. Include paths outside the current working directory when relevant. If none, write "(none)".

- \`[file or directory path]\`: [brief reason it matters]

## Important Context

- [facts the next agent cannot continue without and cannot easily find on its own; or "(none)"]

Rules:

- Keep each section concise. Use terse, single-line bullets, not prose paragraphs or nested lists.
- Prefer short references over detailed restatement. It is fine to leave out information the next agent can recover from the code or the files listed above.
- Preserve exact file paths, symbols, commands, error strings, URLs, and identifiers.
- Carry forward only user questions or requests that remain unanswered or require further action. Do not repeat ones that newer history has answered or resolved. Preserve exact wording when carrying one forward.
- Preserve consequential workflow state, including whether changes are uncommitted, committed, pushed, under review, or merged.
- Do not mention the summary process or that context was compacted.`