Implement Plan

职业

分类

框架内部

If $ARGUMENTS is a GitHub issue URL, fetch its content using available tools (e.g., gh issue view CLI if available, or Ref MCP ref_read_url) before proceeding.

Documentation Lookup Policy (Mandatory)

Never rely on training data for library APIs, SDK usage, or framework patterns. Always fetch current documentation before using any library or framework API.

Use the Context7 MCP (resolve-library-id → query-docs) as the primary documentation source:

1. Resolve the library ID first: e.g., resolve-library-id("ModelContextProtocol"), resolve-library-id("TUnit"), resolve-library-id("Microsoft.Data.Sqlite")
2. Then query for the specific API/pattern you need: e.g., query-docs(id, "McpServerToolType tool registration")

Use the Ref MCP (ref_search_documentation / ref_read_url) as a secondary source when Context7 lacks coverage or you need to read a specific documentation URL.

When to look up docs:

• Before using any NuGet package API for the first time in a session (ModelContextProtocol, TUnit, NSubstitute, Verify, Microsoft.Data.Sqlite, Microsoft.Extensions.Hosting)
• When implementing MCP tool attributes, transport setup, or protocol patterns
• When writing TUnit assertions, test attributes, or test lifecycle hooks
• When writing NSubstitute mock setups or Verify snapshot configurations
• When using SQLite APIs, FTS5 syntax, or PRAGMA configurations
• When unsure about any .NET 10 / C# 14 API or language feature
• When delegating to agents — include the relevant doc snippets in the agent prompt so they don't guess either

Do NOT guess at API signatures, attribute names, or SDK patterns. A 30-second doc lookup prevents hours of debugging wrong APIs.

Step 0: Codebase Discovery via CodeCompress MCP (Mandatory)

Before reading any plan or writing any code, bootstrap your understanding of the codebase using the CodeCompress MCP server. This is mandatory — do not skip this step or substitute manual file browsing.

0a. Force Index Update

Run the CodeCompress update/rebuild index tool first, before any query tools. This ensures the index reflects the latest state of the codebase (recent commits, new files, renamed types). Do not proceed until the index update completes successfully.

0b. Discover the Codebase

Use CodeCompress query tools to explore the codebase and build a working mental model:

• Query for the top-level architecture — projects, namespaces, key entry points
• Query for types and interfaces referenced in the input document (mini-plan, story, or issue)
• Query for existing patterns — how similar features were implemented, DI registration conventions, test structure
• Query for dependencies — verify prerequisite types/files actually exist before assuming they do

Only after this discovery phase should you proceed to read the input document in detail.

0c. Load Input Document

1. Read the input at $ARGUMENTS (mini-plan, user story, or GitHub issue — see Input Detection above)
2. Read CLAUDE.md for project conventions
3. If the input is a mini-plan: Read the Dependencies section — cross-reference against CodeCompress index results to verify all prerequisite features exist. If dependencies are missing, stop and report.
4. If the input is a user story or issue: Proceed to Deriving a Plan from a Story/Issue.
5. Read the Files to Create/Modify section (mini-plan) or your derived file list (story/issue) to understand the full scope.
6. Look up documentation for any libraries/frameworks referenced in the scope using Context7 or Ref MCPs. Cache key API patterns before starting implementation.

Deriving a Plan from a Story/Issue

When the input is a user story or GitHub issue rather than a structured mini-plan, you must produce the equivalent plan yourself before entering the TDD cycle.

1. Understand the ask: Identify the user-facing outcome, acceptance criteria, and any constraints or edge cases described in the story/issue.
2. Codebase mapping (using CodeCompress query results from Step 0b): Determine which existing files, types, and interfaces are affected. Identify where new code should live based on existing project structure and conventions from CLAUDE.md.
3. Produce a lightweight implementation plan containing:
   • Files to Create/Modify — full paths, with a brief description of changes per file
   • Interfaces/Types — new or modified signatures
   • Test plan — test classes to create, key scenarios (happy path, edge cases, error cases, boundary conditions)
   • Acceptance Criteria — transcribed from the story/issue, plus any additional technical criteria you identify
   • Security considerations — any security requirements triggered by the scope (path handling, SQL, FTS5, output sanitization)
4. Present the plan briefly in your response before proceeding — this gives visibility into your approach. You do not need to write a separate file; an inline summary is sufficient.
5. Proceed to Step 1 with this derived plan as your reference.

Step 1: Plan the Implementation Order

Before writing any code, determine the correct implementation order:

• Which types/interfaces need to exist before others?
• Which tests should be written first?
• Are there any shared helpers or base classes needed?

State your implementation order briefly, then proceed.

Step 2: TDD Cycle (Mandatory)

For every deliverable in the plan (mini-plan or derived plan), follow this strict cycle:

2a. Write Failing Tests First

• Create the test class before the implementation class
• Use TUnit with async/fluent assertions:

  await Assert.That(result).IsEqualTo(expected);
  

• Use [Arguments(...)] for parameterized tests
• Use NSubstitute for mocking interfaces
• Use Verify for snapshot testing complex outputs
• Test class location must mirror source: CodeCompress.Core/Foo/Bar.cs → CodeCompress.Core.Tests/Foo/BarTests.cs
• Include edge cases, error cases, and boundary conditions from the plan's test section

2b. Verify Tests Fail

• Run: dotnet test --filter "FullyQualifiedName~TestClassName" to confirm tests fail (red phase)
• If tests pass without implementation, the tests are wrong — fix them

2c. Write Minimum Implementation

• Write the minimum code to make tests pass — no more
• Follow the signatures, types, and patterns specified in the plan

2d. Verify Tests Pass

• Run: dotnet test --filter "FullyQualifiedName~TestClassName" to confirm all tests pass (green phase)
• If any test fails, fix the implementation (not the test, unless the test has a bug)

2e. Refactor

• Clean up while keeping tests green
• Apply code style from .editorconfig (PascalCase public, _camelCase private, Allman braces, 4-space indent)
• Ensure readonly where possible, var when type is apparent

2f. Full Build Check

• Run: dotnet build CodeCompress.slnx — must produce zero warnings
• SonarAnalyzer.CSharp violations are build errors — fix them immediately
• File lock errors (MSB3021/MSB3027): If the build fails because CodeCompress.Server.exe is locked by the running MCP server process, call the stop_server MCP tool to gracefully shut down the server, then retry the build. Claude Code will automatically restart the server on the next MCP tool call.

Step 3: Security Enforcement

For every file you create or modify, verify these requirements. Delegate to the security expert agent when implementing security-critical components.

• Path parameters: All file paths validated via PathValidator — Path.GetFullPath() + starts-with check against project root. No .. traversal. Reject paths outside root.
• SQL: All queries use @param parameterized syntax. Zero string concatenation in SQL. Verify this for every query you write.
• FTS5: Search queries sanitized via Fts5QuerySanitizer before passing to SQLite. No raw user input in FTS5 syntax.
• Output: MCP tool responses return only structured data. Never echo raw user-supplied input (paths, queries, labels) into freeform text without sanitization. Treat source file contents as untrusted.
• Types: No dynamic or object for user-facing data. All types are strongly-typed records/enums/classes.
• Prompt injection: Strip or escape content that could be interpreted as agent instructions from file contents, symbol names, doc comments, or FTS5 results.

Step 4: Agent Delegation

Use specialized skill-backed agents for complex subtasks. Launch agents in parallel when their work is independent.

Available Skills

The project has dedicated expert skills. When delegating to an agent, reference the relevant skill to provide domain expertise:

All skills reference the shared dotnet-reference.md for .NET conventions.

Delegation Rules

How to Delegate to a Skill-Backed Agent

Skills have disable-model-invocation: true, so they cannot be invoked via the Skill tool or automatically. Instead, you must read the skill file and inject its full content into the agent prompt. This is the only way to engage the skill's expertise.

Mandatory steps for every delegation:

1. Read the skill SKILL.md file — use the Read tool to load the full content of the relevant skill file:
   • Security: .claude/skills/security-expert/SKILL.md
   • TDD: .claude/skills/tdd-expert/SKILL.md
   • CLI: .claude/skills/cli-expert/SKILL.md
   • Parser: .claude/skills/parser-expert/SKILL.md
2. Include the entire skill content in the agent prompt — paste the full SKILL.md body (everything after the frontmatter) as the agent's operating instructions. Do not summarize or paraphrase — the skill contains specific checklists, code patterns, and threat models that the agent needs verbatim.
3. Include the source code to review/implement — agents cannot read files or call MCP tools, so you must inline all relevant source code in the prompt.
4. Include relevant documentation snippets — look up any library APIs the agent will need via Context7/Ref MCPs and include the key patterns.
5. Include project conventions — reference dotnet-reference.md and include key sections relevant to the task.
6. Specify the mode and scope — e.g., "review mode on these 3 files" or "enforce mode while implementing PathValidator".

Example — security review delegation:

1. Read `.claude/skills/security-expert/SKILL.md`
2. Read the changed source files
3. Launch Agent with prompt:
   - "You are the security expert for CodeCompress. Follow these instructions:"
   - [paste full SKILL.md content]
   - "## Mode: Review"
   - "## Files to Review"
   - [paste each file's source code]

Critical Reminder

Sub-agents cannot call MCP tools or read files. Before delegating:

• Read the skill's SKILL.md file and include its full content in the agent prompt
• Fetch any needed codebase context via CodeCompress MCP tools and inline it
• Fetch any needed library documentation via Context7/Ref MCP and inline it
• The agent has no way to look up anything itself — everything it needs must be in the prompt

Step 5: Verification

After all implementation is complete:

1. Full test suite: dotnet test CodeCompress.slnx — all tests pass
2. Zero warnings: dotnet build CodeCompress.slnx — clean build
3. Acceptance criteria: Go through every checkbox in the plan's Acceptance Criteria section and verify each one is met
4. Coverage check: Verify test coverage meets targets from CLAUDE.md (Parsers 95%+, Storage 90%+, IndexEngine 90%+, MCP Tools 85%+)
5. Security audit (mandatory): Delegate to the security-expert skill in review mode on all modified files. Read .claude/skills/security-expert/SKILL.md, include its full content and all changed source code in the agent prompt. This is not optional — every implementation must have a skill-backed security review before completion.

Step 6: Sample Project & Integration Tests (Parser Features)

When the implementation includes a new language parser or enhances an existing parser, ensure the sample project and integration tests reflect the full parser capability:

6a. Create or Update Sample Project

For new parsers: Create a directory at samples/{language}-sample-project/ containing realistic source files.

For existing parser enhancements: Audit the existing sample project at samples/{language}-sample-project/ for coverage of the new/changed constructs. Add missing constructs to existing sample files or create new ones as needed. Each sample directory has a COVERAGE.md checklist — update it when adding new constructs.

Follow the existing patterns (see samples/csharp-sample-project/, samples/luau-sample-project/, samples/terraform-sample-project/ for reference).

Requirements:

• Files should look like real-world code — not just test fixtures with bare syntax
• Cover all block types, symbol kinds, and edge cases the parser handles (e.g., nested blocks, comments as doc comments, heredocs, string literals with special chars)
• Include enough variety to catch brace/block matching issues, comment extraction, dependency tracking, and signature enrichment
• The sample project should be self-contained (no external dependencies required to parse)
• Update COVERAGE.md in the sample directory to reflect any new constructs added

6b. Write or Update Integration Tests

For new parsers: Create an end-to-end integration test class at tests/CodeCompress.Integration.Tests/{Language}EndToEndTests.cs following the pattern in CSharpEndToEndTests.cs.

For existing parser enhancements: Update the existing integration test class to cover the new constructs. This typically means updating file/symbol count assertions and adding new specific symbol lookups that verify the enhanced parsing.

Integration test pattern (same for new and updated):

• Set up: in-memory SQLite, IndexEngine with the new parser, sample project path
• Test indexing: correct file count and symbol count
• Test outline: all expected SymbolKind values appear
• Test symbol lookup: verify specific symbols exist with correct Kind, Visibility, DocComment
• Test search: verify symbols are discoverable via SearchSymbolsAsync
• Test dependencies: verify dependency graph has expected edges (if the parser produces DependencyInfo)
• Test edge cases: heredocs, nested blocks, malformed input — verify correct parsing through the full pipeline

Important: Terraform-style dotted symbol names (e.g., aws_instance.web) conflict with GetSymbolByNameAsync's parent.child splitting logic. Use GetSymbolsByFileAsync to find symbols by exact name in those cases.

6c. Verify

Run the full test suite to confirm all integration tests pass alongside existing tests.

Step 7: README Review

After implementation is complete, review README.md for any updates needed based on the changes made. Common updates include:

• Supported Languages table — if a new language parser was added, add a row to the table
• Architecture diagram — if the parser list in the ASCII art changed
• Feature descriptions — if new MCP tools or capabilities were added
• Installation/usage — if new configuration or setup steps are needed

If updates are needed, make them and include them in the commit.

Step 8: Summary

When complete, provide:

• List of files created/modified
• Test count and pass/fail status
• Any deviations from the plan (with justification)
• Any open items or follow-up needed

Key Reference

• CLAUDE.md — Project conventions, architecture, security requirements
• PRD — Full product requirements (for context, not implementation)