context-management（上下文管理）

Context Management

Manage your context window, survive compaction, persist state across turns.

<when_to_use>

• Planning long-running or multi-step tasks
• Coordinating multiple subagents
• Approaching context limits (degraded responses, repetition)
• Need to preserve state across compaction or sessions
• Orchestrating complex workflows with handoffs

NOT for: simple single-turn tasks, quick Q&A, tasks completing in one response

</when_to_use>

Claude Code operates in a ~128K token context window that compacts automatically as it fills. When compaction happens:

What survives:

• Task state (full task list persists)
• Tool results (summarized)
• User messages (recent ones)
• System instructions

What disappears:

• Your reasoning and analysis
• Intermediate exploration results
• File contents you read (unless in tool results)
• Decisions you made but didn't record

The consequence: Without explicit state management, you "wake up" after compaction with amnesia — you know what to do, but not what you've done or decided.

Tasks: Your Survivable State

Tasks are not just a tracker — they're your persistent memory layer. Use TaskCreate, TaskUpdate, TaskList, and TaskGet to manage state that survives compaction.

What Goes in Tasks

Task Discipline

Exactly one in_progress at any time. Call TaskUpdate to mark in_progress before starting.

Mark complete immediately. Don't batch completions — TaskUpdate with completed as you finish.

Include agent IDs for resumable sessions in task metadata.

Expand dynamically. TaskCreate as you discover work; don't front-load everything.

Action-oriented subjects. Use verbs: "Implement X", "Fix Y", "Review Z"

Status Flow

pending → in_progress → completed
                ↓
         (blocked → TaskCreate for blocker, add blockedBy)

If blocked, don't mark complete. Create a new task for the blocker and link with addBlockedBy.

Pre-Compaction Pattern

As context fills, ensure tasks capture progress. Use TaskUpdate to add details to in_progress task description:

Task: Implementing token refresh
Description:
  - Refresh endpoint: POST /api/auth/refresh
  - Token rotation: enabled
  - Refresh window: 5 minutes before expiry
  - Remaining: validation logic

Decisions embedded in completed task descriptions. Current state detailed in active task. Future work queued as pending.

<pre_compaction>

Pre-Compaction Checklist

Run through this when context is filling (you'll notice: slower responses, repetition, degraded reasoning):

1. Capture progress — What's done? TaskUpdate completed tasks with outcomes in description.
2. Record decisions — What did you decide? Why? Put in task descriptions.
3. Note current state — Where exactly are you in the current task? TaskUpdate the in_progress task with specifics.
4. Queue discovered work — What did you find that needs doing? TaskCreate as pending.
5. Mark dependencies — What needs what? Use addBlockedBy in TaskUpdate.
6. Include agent IDs — Any background agents? Record IDs in task metadata.

Example: Before Compaction

Bad (state will be lost):

- [x] Research auth approaches
- [ ] Implement auth
- [ ] Test auth

Good (state survives):

- [x] Research auth approaches → middleware + JWT (see src/auth/README.md)
- [ ] [in_progress] Implement JWT refresh flow
    - Using jose library (already in deps)
    - Endpoint: POST /api/auth/refresh
    - Handler started in src/auth/refresh.ts:15
    - Remaining: validation logic, token rotation
- [ ] Add refresh flow tests (after impl)
- [ ] [reviewer] Security review auth module (after tests)

</pre_compaction>

Delegation for Context Preservation

Main conversation context is precious. Every file you read, every search result, every intermediate thought consumes tokens. Subagents run in isolated contexts — only their final output returns.

Default Stance

If a task can be delegated, delegate it.

Delegation Decision Tree

Task arrives
├── Exploration/research? → Explore agent (always)
├── Multi-file reading? → Subagent (summarizes for you)
├── Independent subtask? → Background agent
├── Specialized expertise? → Domain agent (reviewer, tester, etc.)
└── Simple, focused, single-file? → Main agent (maybe)

Note: Plugin agents require the plugin:agent-name format (e.g., outfitter:reviewer). Built-in agents (Explore, Plan, general-purpose) work without prefix. Use /agents to see available agents.

Context-Saving Patterns

Research delegation — Instead of reading 10 files:

{
  "description": "Find auth implementation",
  "prompt": "Locate authentication-related files, summarize the auth flow",
  "subagent_type": "Explore"
}

Main agent receives: concise summary, not 10 file contents.

Parallel review — Instead of sequential analysis:

// Single message, multiple calls, all run_in_background: true
{ "subagent_type": "outfitter:reviewer", "run_in_background": true, "prompt": "Security review..." }
{ "subagent_type": "outfitter:analyst", "run_in_background": true, "prompt": "Performance review..." }

Main agent: stays lean, collects results when ready.

Background execution — For independent work:

{
  "subagent_type": "outfitter:tester",
  "run_in_background": true,
  "prompt": "Run integration tests for auth module"
}

Continue other work; retrieve with TaskOutput later.

Task Integration

Track delegated work in tasks:

[analyst] Research caching strategies (pending, metadata: {taskId: "def456"})
[engineer] Implement cache layer (pending, blockedBy: analyst task)
[reviewer] Review cache implementation (pending, blockedBy: engineer task)
[tester] Validate cache behavior (pending, blockedBy: reviewer task)

When background agents complete, TaskUpdate status and process results.

What NOT to Delegate

• Direct user Q&A needing conversation history
• Simple edits to files already in context
• Final synthesis requiring your judgment

<cross_session>

Cross-Session Patterns

For work spanning multiple sessions, use episodic memory MCP server.

Prerequisites: Cross-session patterns require an episodic-memory MCP server to be configured. If unavailable, skip this section — Tasks handle single-session persistence.

Saving State

At session end or before long pause:

{
  "tool": "episodic-memory:save",
  "content": {
    "task": "Implementing auth refresh flow",
    "status": "in_progress",
    "completed": ["JWT validation", "Refresh endpoint structure"],
    "remaining": ["Token rotation logic", "Tests", "Security review"],
    "decisions": {
      "library": "jose",
      "algorithm": "RS256",
      "refresh_window": "5 minutes"
    },
    "files_modified": ["src/auth/refresh.ts", "src/auth/middleware.ts"],
    "next_steps": "Implement token rotation in refresh.ts:42"
  }
}

Restoring State

At session start:

{
  "tool": "episodic-memory:search",
  "query": "auth refresh implementation"
}

Then reconstruct tasks from saved state using TaskCreate.

When to Use Cross-Session

• Multi-day projects
• Complex refactors with many steps
• Work that will be interrupted
• Handing off to future sessions

For single-session work, Tasks alone suffice.

</cross_session>

Workflow Integration

At Task Start

1. TaskCreate with initial scope
2. If complex: use Plan subagent to explore, preserve main context
3. TaskUpdate first task to in_progress

During Execution

1. TaskUpdate as work progresses
2. Delegate exploration to subagents
3. Mark completed immediately (no batching)
4. TaskCreate for discovered work
5. Note decisions in completed task descriptions

Approaching Compaction

1. Run pre-compaction checklist
2. Ensure current state captured in in_progress task description
3. Record any background agent IDs in task metadata

After Compaction

1. TaskList to read task state (it persists)
2. Resume from in_progress task
3. Use saved details to continue without re-exploration

At Task Completion

1. TaskUpdate final tasks to complete with outcomes in description
2. If multi-session: save to episodic memory
3. Report summary to user

ALWAYS:

• Use Tasks for any work over 2-3 steps
• TaskUpdate before significant actions
• Mark completed immediately, not batched
• Include agent IDs in task metadata for resumable sessions
• Delegate exploration to subagents (preserves main context)
• Record decisions in completed task descriptions
• Run pre-compaction checklist when context fills

NEVER:

• Rely on conversation history surviving compaction
• Keep large research results in main context (delegate or summarize)
• Have multiple in_progress tasks simultaneously
• Stop early due to context concerns (persist state instead)
• Batch multiple completions together
• Leave tasks vague ("do the thing" → "Implement refresh endpoint")

• task-patterns.md — deep patterns and templates
• delegation-patterns.md — context-preserving delegation
• cross-session.md — episodic memory integration
• subagents skill — agent orchestration patterns