Code Explorer
Understand how an existing feature works before changing it. Code Explorer helps map a code path; its explanation still needs to be checked against your repository.
Is it right for your task?
- Use it when
- You are onboarding to a repository or preparing a change that crosses several files.
- Skip it when
- You already know the single function you need, or are looking for proof that a system is secure or correct.
- Start simpler
- Ask Claude to explain one file first. Escalate to feature exploration only when the behavior crosses multiple components.
Try it: map a feature before editing it
Example brief. Adapt it to your own test project.
Use the code-explorer agent to explain how one existing user-facing feature works in this repository. First ask me for its route or entry point if it is unclear. Trace the path from user action through validation to data access and response. Cite current file paths and line numbers. Separate directly observed code from inferences. Identify one failure path, relevant tests and unanswered questions. Do not edit files, install anything or contact external services. Stop after the explanation; do not begin implementation.
What to expect
A bounded feature map with citations you can open, plus clear unknowns. There is no universal answer key because repositories differ; the checks below are the acceptance criteria, not a claim that this agent passed them.
Check the result
- Every cited path exists at the recorded revision and the referenced lines support the associated claim.
- At least one real call chain connects the entry point to an output or side effect; a list of filenames alone is insufficient.
- An error path is grounded in code rather than invented expected behavior.
- Check git status and the diff afterward. The exercise requests no changes; this is not a permission-enforced guarantee.
Before you start
- Claude Code and a repository you are permitted to share with its environment.
- The publisher's Feature Development plugin. This agent is part of that plugin, not the PR Review Toolkit.
- A specific feature, route or entry point to investigate; avoid asking for an entire large repository in one pass.
- Claude Code
- The source defines a Sonnet subagent inside the Feature Development plugin.
- Claude web
- This is not a connector or a hosted repository browser. The plugin setup described here targets Claude Code.
Setup and first use
- Open a disposable checkout or a clean working tree and record its current revision.
- Follow the linked Feature Development plugin setup. Inspect its agent definition and permissions before enabling it.
- Check /agents after installation, then explicitly request Code Explorer for one feature.
- Run the example using a route that actually exists in your repository. Do not launch the full feature-development workflow when you only want an explanation.
- Open the cited files yourself and follow one successful path and one failure path. Save unresolved questions before planning changes.
Access, cost, and limits
What access are you granting?
The published tool list includes file inspection and web tools, plus shell-output/process tools. Treat the task's no-edit request as intent, not isolation. Review Claude Code permissions, protect secrets and deny unnecessary network access.
What does it cost?
Exploration consumes Claude usage under your account. Keep the scope small and review your usage limits; no fixed token cost or completion time is promised.
Limitations to keep in mind
- Static inspection can miss dynamically configured behavior or services outside the checkout.
- Paths and line numbers become stale after code changes; record the revision with your notes.
- The parent plugin also contains design and implementation phases. This guide only covers exploration, and does not claim independent execution or measured accuracy.
If something goes wrong
The agent is unavailable
Check the Feature Development plugin installation, active scope and /agents list. Reopen the session if the plugin was just enabled.
The explanation is too broad
Name one entry point and one question. Ask it to stop at the first external boundary and list the missing evidence.
A reference does not support the claim
Give the specific mismatch and ask for corrected evidence. Do not carry an unsupported architectural claim into your implementation plan.
Sources and review scope
Publisher: Anthropic. This editorial guide checks published documentation, not runtime behavior. The examples and acceptance checks are our suggested exercises.
- Code Explorer definition
Agent purpose, model, declared tools and citation-oriented output. Moving source checked on the date above.
- Feature Development plugin
Parent plugin, manual exploration and the wider development workflow.
Reviewed 2026-09-21. Recheck due 2026-10-21, or sooner if setup changes. No installation or workflow execution is claimed.
Report an issue with this guide