---

name: kb description: "Manages the Morphir knowledge base under kb/ — OKF bundles and concept documents. Use when adding content to a bundle, creating a new bundle, checking the knowledge base for conformance or provenance drift, building or querying its SQLite index, managing intent through its lifecycle, recording or reading architectural decisions as Decision Records, or navigating, searching and listing its bundles, concepts and links." allowed-tools: Bash(.claude/skills/kb/kb *), Bash(cat *), Bash(ls *), Bash(find *), Bash(git *), Read, Edit, Write metadata: version: 0.7.0

kb — Morphir Knowledge Base Assistant

Manages kb/, the Open Knowledge Format knowledge base. Bundles live under kb/bundles/, optionally grouped a level deeper. kb/AGENTS.md is the source of truth for the conventions; this skill automates the mechanical parts of following them.

The kb command

Everything runs through the launcher in this directory. It wraps Mill's single-file Scala scripting — no build file, no install step. The first run resolves dependencies and compiles; later runs are incremental.

.claude/skills/kb/kb list

.claude/skills/kb/kb check --verbose

Every command accepts --json. Progress output goes to stderr, so --json on stdout is clean and pipeable — prefer it when you need to consume the result rather than read it.

Full flag reference: → references/commands.md

| Command | Does | | ------- | ---- | | list | Bundles and their concept counts; --bundle X lists that bundle's concepts | | show --path /x.md | One document: frontmatter, outbound links, heading outline | | search --query X | Search titles, descriptions, tags and paths; --body to include prose | | check | Conformance and provenance findings; non-zero exit on errors | | index | Builds the SQLite index; --status reports its freshness | | refresh | Both kinds of derived state; narrow with refresh markdown / refresh db | | query --sql | Read-only SQL over that index | | sync … | Mirroring an upstream repository into a bundle — status, pull, push, diff | | intent … | Intent lifecycle — new, list, show, the transition verbs, check | | new-bundle | Scaffolds a bundle with index.md and log.md | | add-concept | Scaffolds a concept and wires it into the index and log |

When to use what

Adding content to an existing bundle. Run add-concept to create the file and wire it up, then write the body yourself. The scaffold deliberately leaves a TODO comment rather than plausible-looking prose.

→ references/authoring.md before writing the body.

Writing or reviewing prose. Every concept body is written in one of three registers, chosen by its type:. The style cards are tool-neutral and live at .agents/skills/kb/styles/; voice.md there applies to all registers and includes the banned-pattern list (no em-dashes, no AI filler), diagrams.md says when a Mermaid diagram or SVG should replace prose narration of a flow or structure, and altitude.md sizes documents to capability stories: one narrative-home Design Note per capability in flight, with every fine-grained concept reachable from it.

| type: | Register | Card | | ------- | -------- | ---- | | Playbook, tutorial, orientation, guidance | article | styles/article.md | | Design Note, Decision Record, Intent, synthesis | white-paper | styles/whitepaper.md | | Reference, Specification Section, Glossary, Data Dictionary, version-pinned notes | reference | styles/reference.md |

Delegate drafting to the kb-writer subagent and review to the kb-reviewer subagent (.claude/agents/kb-writer.md, kb-reviewer.md), naming the target file and register in the dispatch. Writing inline is fine for small edits; still apply the cards. Style applies to new and touched content only; do not sweep existing prose.

Creating a bundle. Run new-bundle, then add it to the Bundles table in kb/README.md and to the group's README.md if it is in a group. The command reminds you; it does not edit those files, because their wording is a judgement call.

Checking the knowledge base. Run check. It reports structural problems (missing type, broken links, unindexed concepts, frontmatter that does not parse) and provenance drift (commit-pinned sources whose reference checkout has moved on). Nothing here touches the network.

→ references/checks.md for the catalogue and how to fix each finding.

Searching and locating. search scans the markdown and is always current. For anything heavier — full-text search over bodies, "what links here", orphaned concepts, tag or provenance distributions — build the SQLite index once and query it.

.claude/skills/kb/kb index

.claude/skills/kb/kb search --query "entry point" --index

The index is derived state under .dev/kb/index.db, gitignored, and rebuilt from the markdown. It has no automatic invalidation — kb index --status lists files changed since the last build.

Keeping derived state honest. kb refresh does both halves in one pass: it rewrites index bullets that have drifted from their concept's description, then rebuilds the SQLite index if anything changed.

.claude/skills/kb/kb refresh --dry-run

Narrow it when you only want one half — kb refresh markdown or kb refresh db, equivalently --no-db and --no-markdown. Reach for it after editing descriptions or adding concepts, and before relying on a query. --add-missing also appends entries for unindexed concepts, which is opt-in because it has to pick a section.

→ references/index-db.md for the schema, the views, and worked queries.

Managing intent. Features, enhancements and bugs are recorded as prose in the intent bundle, with a lifecycle whose obligations are enforced — most importantly, releasing requires linking the Capability it produced.

→ the intent skill for the process; references/commands.md for flags.

Recording a decision. Architectural decisions are the knowledge base's third register, alongside Intent and Capability: past-tense, immutable, and superseded rather than edited. kb decision list and kb decision show read them; kb check validates their supersession links.

→ references/decisions.md for the frontmatter, the checks and how to supersede one.

Mirroring an upstream repository. A bundle may declare a sync.yaml and carry upstream's own files rather than a paraphrase of them. Markdown lands as concepts with an injected, fenced block of kb-owned frontmatter; everything else lands as byte-identical assets. kb sync push deletes exactly that fenced region, so what goes back upstream is what came from it.

.claude/skills/kb/kb sync status

→ references/sync.md for the manifest, the lockfile and the state model.

Finding divergence in the content. check finds mechanical inconsistency. Contradictions between what two concepts assert — the thing that actually matters in a knowledge base — cannot be detected by a script.

→ references/divergence.md for that procedure.

Rules that the tooling assumes

• A bundle root is a directory whose index.md carries okf_version. That is how bundles are discovered.
• Only index.md and log.md are reserved. Every other .md file inside a bundle is a concept and needs type:. The reservation stops at a mirror boundary — inside a vendored subtree those names are upstream's own files.
• A grouping directory gets a README.md and never an index.md. README.md inside a bundle is an error.
• Sub-directory index.md files carry no frontmatter.
• Index bullets mirror the target concept's description. Changing one means changing the other.

Working on the skill itself

The scripts are Scala, run through Mill's single-file mode. Each file declares its own dependencies in a //| YAML header; kb.scala is the entry point and names the others in moduleDeps.

| File | Holds | | ---- | ----- | | kb.scala | kyo-case-app command definitions and CLI plumbing | | KbModel.scala | Pure domain model — bundles, docs, frontmatter, path arithmetic | | KbStore.scala | Loading and parsing: frontmatter via SnakeYAML, bodies via commonmark-java | | KbCheck.scala | The check catalogue | | KbScaffold.scala | Bundle and concept creation, index and log editing | | KbIndex.scala | SQLite schema, index build, and query surface | | KbRefresh.scala | Reconciling derived state — index bullets and the database | | KbSync.scala | Vendoring an upstream repository: manifest, lockfile, frontmatter injection and projection | | KbIntent.scala | Intent model, lifecycle states, kinds and checks | | KbIntentEdit.scala | Creating intent, transitions, generated intent index | | KbDecision.scala | Decision Record model, supersession checks and rendering | | KbTests.scala | The kyo-test suites — run them with mise run kb:test | | KbRender.scala | Text and JSON rendering |

kyo is the standard library here: kyo.Path for paths and file access, kyo.Command for subprocesses, kyo-case-app for the CLI. JDBC sits inside Sync.defer at the edge, as everything else effectful does. Build state lands in out/, which is gitignored.

Compile without running:

.claude/skills/kb/mill compile kb.scala

Run the tests:

mise run kb:test

80 cases across seven suites, using kyo-test — the same framework langkit and kit use. Mill's script mode exposes no test module (mill test <script> does not resolve), so the suite runs through kyo-test's own CLI runner as the script's mainClass, with //| resources: [test-resources] putting the ServiceLoader registry on the classpath.

A new suite must be listed in test-resources/META-INF/services/kyo.test.Test or it will never run. Discovery is ServiceLoader-only and the runner takes no suite arguments, so an omission does not fail — it silently skips, leaving CI green over untested code.

That is too quiet to leave to documentation, so KbMetaSpec asserts that the registry and the suites defined in KbTests.scala agree, and names the offender when they do not. The note above is a courtesy; the check is the guarantee.

Cases named regression pin behaviour that was once wrong. Add one whenever a bug is found, rather than only fixing it.

One caveat worth knowing if you extend the scripts: inside Mill's script sandbox, os.Path values built from an environment variable and those returned by os.list can render identically yet compare unequal. kyo.Path is a value over its segments and does not have that problem — which is why the scripts use it throughout.