context-discipline（语境纪律）

Context Discipline

Core Principles

1. MCP tools first, file reads second — A Roslyn MCP query costs 30-150 tokens. Reading a file costs 500-2000+ tokens. For navigation and understanding, always try MCP tools before opening files. Only read files when you need to modify them or MCP tools don't provide enough detail.
2. Lazy load everything — Don't read files "just in case." Don't load all skills upfront. Don't explore directories you aren't about to modify. Load information at the moment you need it, not before.
3. Subagents are context isolation chambers — Every subagent gets its own context window. Offload exploration, research, and analysis to subagents. They process information and return a summary — your main context stays clean.
4. Summarize and discard — After exploring a subsystem, summarize what you learned in a few lines. The summary is what stays in context, not the raw file contents. Think of it as compressing information.
5. Know your budget — A 200k token window sounds large but fills fast. A typical.cs file is 500-2000 tokens. Loading 50 files can consume half your budget. Plan your reads like you plan your sprints — deliberately.

Patterns

MCP-First Navigation

Always prefer MCP tools for understanding code structure:

TASK: Understand how OrderService works

EXPENSIVE APPROACH (file reads):
1. Read src/Orders/OrderService.cs          → ~1200 tokens
2. Read src/Orders/IOrderService.cs         → ~300 tokens
3. Read src/Orders/OrderRepository.cs       → ~800 tokens
4. Read src/Orders/Models/Order.cs          → ~600 tokens
Total: ~2900 tokens consumed

TOKEN-EFFICIENT APPROACH (MCP-first):
1. find_symbol "OrderService"               → ~50 tokens (file path + line)
2. get_public_api "OrderService"            → ~120 tokens (method signatures)
3. find_references "OrderService"           → ~80 tokens (who uses it)
4. get_type_hierarchy "OrderService"        → ~60 tokens (inheritance chain)
Total: ~310 tokens consumed — 9x cheaper

Only THEN read the specific method you need to modify: ~200 tokens
Grand total: ~510 tokens vs 2900 — 5.7x savings

Subagent Offloading Decision Matrix

Decide when to offload to a subagent vs. handle in main context:

USE A SUBAGENT WHEN:
- Exploring an unfamiliar part of the codebase (> 3 files to read)
- Researching a question that requires reading docs or multiple files
- Running analysis that produces verbose output (test results, diagnostics)
- Comparing approaches that require loading multiple examples
- Any task where the journey is verbose but the answer is concise

STAY IN MAIN CONTEXT WHEN:
- Modifying a file you've already read
- Quick lookups (1-2 MCP queries)
- Tasks where you need to see prior conversation context
- Writing code that builds on discussion with the user

Example subagent delegation:

TASK: "Find all places where we handle authentication"

MAIN CONTEXT APPROACH (expensive):
- Read 8 files, consume ~8000 tokens
- All that content stays in context forever

SUBAGENT APPROACH (efficient):
- Spawn subagent: "Find all authentication handling in this codebase.
  Use find_symbol and find_references for auth-related types.
  Return: file paths, line numbers, and a 1-line summary per location."
- Subagent returns ~200 tokens of summarized findings
- Main context stays clean

Context Budget Planning

Before a complex task, estimate token spend per phase:

UNDERSTAND (~5k): MCP queries (~300) + subagent exploration (~500) + reserve
PLAN (~2k):       Discussion + plan documentation
IMPLEMENT (~15k): Read files to modify (~3k) + write code (~5k) + iteration (~7k)
VERIFY (~3k):     Build + test + diagnostics + format check
REMAINING: ~175k for conversation — comfortable

File Reading Prioritization

When you must read files, prioritize by impact:

PRIORITY 1 — Files you will modify
Read fully. You need exact content to make correct edits.

PRIORITY 2 — Files with interfaces/contracts you must satisfy
Read the interface/base class. Skip implementation details.

PRIORITY 3 — Files for reference patterns
Use get_public_api first. Only read if the API surface isn't enough.

PRIORITY 4 — Files for general context
Use subagent to summarize. Don't read in main context.

NEVER READ:
- Entire directories "to understand the project" — use get_project_graph
- Test files for context (unless modifying tests) — use get_test_coverage_map
- Generated files (.designer.cs, migrations) — use get_diagnostics for issues
- Package/config files unless specifically needed

Context Pruning & Large Codebase Strategy

WARNING SIGNS (take action if any apply):
- Read 10+ files, 50+ exchanges, forgetting earlier details, re-reading files

RECOVERY: Summarize in 5-10 lines → subagents for remaining exploration →
MCP-only for new lookups → reference prior line numbers → suggest new session if needed

LARGE CODEBASES (50+ projects):
get_project_graph → identify 2-3 relevant projects → find_symbol for key types →
get_public_api for interfaces → read ONLY files you'll modify → subagents for cross-cutting

NEVER: Read every file to "understand" a project, load all skills upfront,
open a file to find one function (use find_symbol)

Anti-patterns

Reading Entire Files for One Function

// BAD — read 1500 tokens to find one 10-line method
Read: src/Orders/OrderService.cs (full file, 80 lines)
*Only needed the ProcessOrder method on line 42*

// GOOD — targeted approach
MCP: find_symbol "ProcessOrder" → "src/Orders/OrderService.cs:42"
Read: src/Orders/OrderService.cs lines 42-55 → ~200 tokens

Loading All Skills Upfront

// BAD — dump 15 skills into context at session start
"Load: modern-csharp, ef-core, minimal-api, testing, docker,
 authentication, logging, caching, messaging, resilience..."
*15 skills × ~300 tokens each = ~4500 tokens before any work starts*

// GOOD — load skills as topics arise
Session start: modern-csharp (always relevant)
User asks about EF: load ef-core
User asks about tests: load testing
*Only pay for what you use*

Not Using Subagents for Exploration

// BAD — explore in main context, polluting the window
Read 12 files across 4 projects to understand auth flow
*~15,000 tokens consumed, all staying in context*

// GOOD — subagent explores, returns summary
Subagent: "Trace the authentication flow from login to token validation.
           Return: the flow as numbered steps with file:line references."
*~300 tokens in main context*

Loading Everything Because the Window Is Large

// BAD — "200k tokens is huge, let's load everything"
Read all 30 files in the Orders module
Read all 15 test files
Read the entire docker-compose.yml
Read all migration files
*80k tokens consumed before writing a single line of code*

// GOOD — minimum viable context
MCP: get_project_graph (solution shape)
MCP: find_symbol (locate target types)
Read: 2-3 files you'll actually modify
Subagent: summarize anything else you need
*~3k tokens consumed, 197k remaining for actual work*

Decision Guide