Kit CLI Navigator

Use the kit command-line interface to interrogate repositories with deterministic tooling before handing context back to Claude. kit bundles structure-aware commands (file trees, symbol extraction, semantic search, dependency graphs, exports) that make “context engineering” repeatable for any language mix. See the reference sheet when you need flag-level detail.

When to activate

• You or the user need a fast mental model of an unfamiliar repo.
• The task is “find where this behavior lives”, “trace symbol usage”, or “surface related code” or any similar request about any codebase.
• Preparing briefs for refactors, dependency upgrades, incident reviews, or onboarding guides.
• Creating exportable context bundles (JSON, DOT, Markdown) for Claude or collaborators.

Prerequisites

1. Install/upgrade kit (requires Python ≥3.9):

   pipx install cased-kit  # or: pip install --upgrade cased-kit
   kit --version
   

2. Run commands from the repository root (or pass an absolute path).
3. For kit search-semantic, ensure sentence-transformers is available; kit prompts for installation if missing.
4. Optional: set KIT_CACHE_DIR when working across large monorepos to reuse indexes.

Core workflow

Always pair this skill with subagent-orchestration: dispatch each kit CLI command to a focused subagent, have it run the command from the repo root, and post a concise result bundle (command, exit status, key findings, saved file paths) back to the main agent thread before proceeding.

1. Confirm scope

   • Ask which repo/directory matters and whether remote refs/tags are relevant.
   • Record success criteria (e.g., “find queue consumers”, “map Terraform modules”).

2. Baseline snapshot

   • Subagent: kit git-info /abs/path → capture branch, SHA, remotes and summarize back.
   • Subagent: kit file-tree /abs/path --path src --output tree.json → structural overview (pipe through jq/less for quick scans) and link the output file.
   • Subagent: kit file-content /abs/path package.json (or other manifest) to read dependencies without opening an editor; paste only the relevant snippet.

3. Surface APIs and hotspots

   • Subagent: kit symbols /abs/path --format table to list top-level functions/classes.
   • Narrow with --file when the repo is large; switch to --format json when Claude needs machine-readable structure (delegate each targeted run to its own subagent).
   • Subagent: kit usages /abs/path ConfigLoader --type class to see definitions + references and return a bulleted digest.
   • Subagent: kit chunk-symbols /abs/path path/to/file.py for precise copy/paste-ready blocks; attach only the relevant chunks.

4. Search by text AND meaning

   • Subagent: run literal filters (kit search /abs/path "retryPolicy" --pattern "src/**/*.ts" or kit grep /abs/path "TODO" --include "*.py") and post paths with short excerpts.
   • Subagent: run semantic recall kit search-semantic /abs/path "where do we sanitize user input" --top-k 10 --chunk-by lines; include embedding settings + highlights.
   • Subagent: kit context /abs/path api/routes.py 128 (± surrounding lines) instead of quoting entire files; return condensed snippets.

5. Understand relationships

   • Subagent: kit dependencies /abs/path --language python --format dot --output deps.dot to visualize imports/modules (render with Graphviz or open in VS Code plug-ins) and record where the asset lives.
   • For Terraform, spin a dedicated subagent with --language terraform and --cycles to smoke-test architectural issues.
   • Subagent: kit index /abs/path --output repo-index.json builds a full-text + symbol inventory for bulk analysis; report file size + location.

6. Package and hand off context

   • Subagent: run exports (kit export /abs/path index out/index.json or kit export /abs/path symbol-usages out/queue.json --symbol QueueWorker) and confirm saved path + checksum.
   • Subagent: create slices for Claude using kit chunk-lines ... --max-lines 80, returning only the refined snippets.
   • Store outputs under debug/ or another ignored directory so they do not pollute Git, and have the subagent mention the folder.

7. Cache + iterate

   • Subagent: kit cache warm /abs/path (see reference.md) so heavy operations reuse embeddings/ASTs, then report cache status.
   • Keep a scratchpad of helpful commands for the session log; reuse them as new questions arise.

Examples

Example 1: Trace where auth tokens are validated

Goal: “Find every spot we touch validateToken and summarize behavior.”

1. kit symbols /repo --file src/auth/index.ts --format json → locate symbol signature.
2. kit usages /repo validateToken --type function --output tmp/validate.json.
3. kit context /repo src/auth/middleware.ts 57 for a targeted snippet you can quote back to the user.
4. Summarize findings + attach the JSON so Claude can ingest it without rerunning expensive scans.

Example 2: Prepare a refactor brief for Terraform modules

1. kit file-tree /repo --path infra/terraform --output tmp/tree.json.
2. kit dependencies /repo --language terraform --cycles --llm-context --output tmp/terraform.md.
3. kit search-semantic /repo "state bucket" --top-k 5 --chunk-by symbols.
4. Use the markdown export + semantic hits to outline which modules share remote state and where to sand down coupling.

Example 3: Build a context pack for Claude

1. kit index /repo --output tmp/index.json.
2. kit export /repo symbol-usages tmp/payment_init.json --symbol PaymentClient.
3. kit chunk-lines /repo services/payment.py --max-lines 60 --output tmp/payment_chunks.json.
4. Feed the JSON bundle into Claude when answering synthesis questions (“Why do payments fail on retries?”).

Best practices

• Prefer absolute repo paths so skills run consistently (kit ... /Users/bleikamp/work/app).
• Use subagents for every CLI run so commands execute in parallel without blocking; ensure each subagent returns the command, exit status, and key findings to the main thread.
• Capture command output snippets in markdown fences when relaying back to Claude; keep them concise.
• Combine literal + semantic search: run kit search first for obvious hits, fall back to kit search-semantic for conceptual matches.
• Export artifacts once per session and reuse them; kit’s deterministic CLI makes outputs easy to diff/share.
• Keep raw command transcripts outside SKILL context (store in debug/kit/ and reference paths).

Common pitfalls & recoveries

• Running from the wrong directory → Always pass the repo path explicitly or cd into it before invoking kit.
• Huge outputs flooding Claude → Use --format json + summarize, or kit chunk-lines to slice manageable blocks. Use subagents whenever possible to avoid polluting the main thread.
• Semantic search unavailable → Install sentence-transformers (kit prompts) or fall back to textual search.
• Dependency visualization fails → Ensure Graphviz (brew install graphviz) is available when using --visualize.
• Stale caches → kit cache clear /repo if outputs look outdated before rebuilding.
• Silent subagents → If a subagent finishes without reporting, explicitly ask for its summary so the main thread has authoritative context.

References

• Command matrix and additional flags: reference.md
• Kit CLI docs and release notes: https://kit.cased.com