kiwiskil 0.4.2

Your codebase, understood by any LLM. Generate a checked-in wiki and skill files from any repo — no cloud, no lock-in.

A ranked, checked-in map any AI agent reads instead of crawling your source — no cloud, no lock-in.

Why · Install · Quick start · What you get · CLI · Configuration · Contributing

---

kiwiskil turns any codebase into a static, checked-in map that any AI agent can navigate and debug — fast, and with a fraction of the tokens of reading source. It parses your code into a call graph, ranks what matters with PageRank, and writes it all to plain markdown in your repo. No cloud service, no vector database, no running server, no lock-in — the map is just files an agent reads directly, and a git hook keeps it current.

---

Why kiwiskil

Code-context tools tend to be either dumb committed blobs (whole-repo dumps) or smart graphs behind a service (vector/graph DBs you have to run). kiwiskil is the missing third thing — a smart, ranked, relational map that's just committed files.

Also: OKF-native markdown (+ SCIP IDs for IDE interop), and six languages — Python, JS/TS, Go, Java, Ruby, Rust — via tree-sitter, no build step.

---

How it works

---

Install

pip install kiwiskil

Quick start

# in any git repo
kiwiskil init       # config + git hook + CLAUDE.md / AGENTS.md
kiwiskil run        # build the map → wiki/ + .indexer/skills/codebase.md

That's it. Every commit re-indexes what changed automatically. No API key needed if you're signed in to the claude CLI.

What you get

• wiki/INDEX.md — a system overview, key flows, the PageRank repo map (most load-bearing symbols first), and the core abstractions.
• wiki/<group>.md — per module: symbols with one-line descriptions, and each symbol's callers, calls, and blast radius ("editing this affects…").
• .indexer/skills/codebase.md — a skill file that teaches any agent to navigate via the map instead of reading source.
• .indexer/manifest.json — every file → its wiki page and component IDs (with SCIP descriptors).

All plain markdown, OKF-framed, checked into your repo. See this repo's own wiki/ for a live example.

CLI

kiwiskil run                     # incremental + deep enrichment (default)
kiwiskil run --force             # full re-index
kiwiskil run --skip-deep         # structural only, faster
kiwiskil run --smart             # verify + repair/fill anything broken or missing
kiwiskil run --smart --dry-run   # report drift only; exits non-zero → CI drift-gate
kiwiskil status                  # last indexed commit, stale files, stats

Configuration

.indexer.toml (created by init):

[llm]
provider = "anthropic/claude-sonnet-4-6"  # any LiteLLM model; leave api_key_env unset to use the claude CLI
api_key_env = "ANTHROPIC_API_KEY"

[indexer]
map_tokens = 1024          # token budget for the ranked repo map
merge_threshold = 2        # min files under a folder before it gets its own page

[hooks]
pre_commit = true
deep = true                # narrative / flows / constraints (set false for speed)

No API key? If the claude CLI is installed and signed in, kiwiskil uses your session — zero config. Otherwise any LiteLLM provider works (OpenAI, Gemini, Ollama, local). With no LLM at all, you still get the full structural map (symbols, call graph, repo map, blast radius); only the written descriptions are skipped.

Languages

Python, JavaScript/TypeScript, Go, Java, Ruby, Rust — via tree-sitter, no build step. A file whose grammar isn't installed is skipped gracefully, never a crash.

Loading the skill

Point your agent at .indexer/skills/codebase.md — it's plain markdown that activates on any codebase question. kiwiskil init already wires up CLAUDE.md and AGENTS.md; for other tools, drop the file into your rules/instructions path (.cursor/rules/, .windsurfrules, .github/copilot-instructions.md, …).

License

MIT · backed by xysq.ai