okf 1.13.0 → 2.1.0

data/lib/okf/skill/reference/authoring.md

@@ -1,26 +1,12 @@
 # Authoring OKF well — the craft
 
-The spec ([SPEC.md](SPEC.md)) tells you what is *legal*. This file is what is
-*good* — the modelling judgment that turns a pile of conformant files into
-knowledge worth consuming. Read it before `produce` or `maintain`, and keep the §9
-conformance rules in mind (parseable frontmatter, a non-empty `type`, and
-well-formed reserved files — the hard rules in [SKILL.md](../SKILL.md)); everything
-else is guidance a consumer must tolerate.
-
-## What each SPEC section governs
-
-Consult the right section on demand instead of re-reading all of [SPEC.md](SPEC.md):
-
-| § | Governs | Reach for it when |
-|---|---------|-------------------|
-| §3 | bundle structure, reserved filenames | laying out directories |
-| §4 | concept documents & frontmatter | writing or validating a concept |
-| §5 / §5.3 | cross-links; **broken links are tolerated** | linking; judging a "broken" link |
-| §6 | index files & progressive disclosure | orienting; writing or synthesizing an index |
-| §7 | log files | recording history |
-| §8 | citations & provenance | any external or empirical claim |
-| §9 | conformance — the hard gate | what `validate` may and may not reject |
-| §11 | versioning (`okf_version`) | the root index's one allowed field |
+The spec ([SPEC.md](SPEC.md)) tells you what is *legal*, and
+[spec-map.md](spec-map.md) says which § settles which question. This file is
+what is *good* — the modelling judgment that turns a pile of conformant files into
+knowledge worth consuming. Read it before `produce` or `maintain`, and keep the
+§11 conformance rules in mind (parseable frontmatter, a non-empty `type`, and
+well-formed reserved files — the hard rules in [SKILL.md](../SKILL.md));
+everything else is guidance a consumer must tolerate.
 
 ## Modelling principles
 
@@ -44,7 +30,9 @@ responsibility. Keep a **small, consistent, descriptive** vocabulary per bundle
 (`Service`, `Dataset`, `Metric`, `Decision`, `Playbook`, `Runbook`, …). Reusing
 types across files is what makes the graph legible; inventing a new type per file
 makes `type` meaningless. Before adding a new type, check what the bundle already
-uses.
+uses. One type the spec *does* name: `Attested Computation` (§10) is a contract,
+not a suggestion — use it exactly when the concept carries a sanctioned
+computation, and use the exact spelling.
 
 ### Tags are the connective axis — curate them like a vocabulary <!-- rule:okf-tag-vocabulary -->
 `type` says what a concept *is*; the directory says where it *lives*; `tags` are
@@ -62,15 +50,53 @@ Lay out directories by what the knowledge is *about* (`services/`, `datasets/`,
 `decisions/`), not by concept type. The directory tree is itself knowledge — it
 shows a reader how the system decomposes, and it usually mirrors the shape of the
 codebase or the org. A `types/`-first layout scatters related concepts and buries
-the domain.
+the domain. Non-concept assets a concept points at — run instructions, attester
+code, computation files — live under `references/` by convention (§6.3): a naming
+convention, not a requirement, and the tools never require it.
+
+### Hidden files are outside the bundle <!-- rule:okf-no-hidden-concepts -->
+The reader excludes dot-prefixed files and every path under a dot-prefixed
+directory — the Unix hidden-file convention, kept deliberately: a project root
+often carries an installed skill (`.claude/`), templates (`.github/`) or other
+markdown that is not knowledge, and reading a directory must not pull those in
+as concepts. §3's taxonomy is read as covering the visible tree. The practical
+rule: never author a concept under a hidden directory — no verb will ever see
+it, and nothing will warn you.
+
+### A path-valued field pointing at `references/` needs its leading `/` <!-- check:broken_attestation_ref -->
+§6.2 gives `resource`, `sources[].resource`, `computation`, `executor.resource`
+and `attester.resource` the same three forms as a link: a URL, a bundle-relative
+path **beginning with `/`**, or a path relative to the concept. So a concept at
+`metrics/revenue.md` writing `attester: { resource: references/attesters/rev.py }`
+is naming `metrics/references/attesters/rev.py`, not the bundle's `references/`
+tree — the trap being that §6.3's own example, and the SPEC's Appendix, spell
+these paths bare from a nested concept. Write `/references/…` and the field
+resolves from the bundle root wherever the concept sits, which is the same reason
+[links](#links-are-untyped-on-purpose) prefer the absolute form. `lint` catches
+the `.md` cases (`broken_attestation_ref`, `broken_source`); a `.sql` or `.py`
+target is not a concept, so no lint check sees it — `okf references` is the
+surface that does: it lists the `references/` tree from disk with every pointer
+that misses, and names the leading-`/` fix when that is the miss.
+
+### A frontmatter `id` renames the concept, not its home <!-- rule:okf-id-extension -->
+§2 defines the Concept ID as the file's path with `.md` removed, full stop.
+This gem additionally honors a frontmatter `id:` as an override — an okf
+extension, not spec, so a bundle leaning on it is trading portability for the
+alias. If you pin one, know the recorded split: links still resolve by path,
+so edges land correctly; the identity views (catalog, hubs, search, `--dir`)
+follow the id, because the edges do; the physical views (`index`, `dirs`,
+stats' `by_dir`) keep the file where it lives, because an index is a physical
+listing. A pinned id that disagrees with the path therefore makes the two
+families answer differently about where the concept is — prefer the default,
+and rename the file when a concept needs a new name.
 
 ### `resource` is the bridge to reality <!-- rule:okf-resource-bridge -->
 Set `resource` (a canonical URI) **only** when a concept *is* a real, addressable
 asset — a table (`bigquery://…`), a service repo, a dashboard, an endpoint. Its
 presence is what lets `maintain` find every concept affected by a changed asset in
-one `okf search <dir> <uri>` call. Abstract concepts — a decision, a principle, a metric
-definition — have no resource, and **omitting it is meaningful**, not laziness. Do
-not invent placeholder URIs.
+one `okf search <dir> <uri>` call. Abstract concepts — a decision, a principle, a
+metric definition — have no resource, and **omitting it is meaningful**, not
+laziness. Do not invent placeholder URIs.
 
 ### Links are untyped on purpose <!-- rule:okf-untyped-links -->
 A markdown link asserts only "these two relate." The *kind* of relationship —
@@ -78,15 +104,61 @@ depends-on, supersedes, derived-from, owns — lives in the **prose around the
 link**, never in a made-up typed-edge syntax. Write the sentence that explains the
 relationship and put the link inside it. Prefer absolute bundle-relative targets
 (`/services/auth-api.md`) so links survive file moves. A link to a concept that
-does not exist yet is fine — it is not-yet-written knowledge (§5.3), and `lint`'s
-backlog will surface it as demand.
+does not exist yet is fine — it is not-yet-written knowledge (§6.1), and `lint`'s
+backlog will surface it as demand. One more edge you get for free: a
+`sources[].resource` naming another concept **is** a graph edge (§5.1's lineage),
+so recording provenance is also linking.
 
-### Provenance is what makes knowledge trustworthy (§8) <!-- check:uncited_external -->
+### Provenance lives in `sources`, attribution in footnotes (§5.1) <!-- check:uncited_external -->
 Any external or empirical claim — a latency number, an approval, a quota, a
-"because X team decided Y" — should carry a citation to its source under a
-`# Citations` heading. Uncited claims are exactly how a bundle decays into folklore
-nobody trusts. `lint`'s provenance category exists to catch missing and broken
-citations; write them as you go so you never have to reconstruct them.
+"because X team decided Y" — should be traceable to a `sources:` entry. Each
+entry carries at least a `resource` (a URL, a bundle-relative path, or a scope
+descriptor like `all queries in project X`), and optionally an `id`, `title`,
+and the credibility signals `author` (an actor), `usage_count` (a liveness
+signal, never a score), and `last_modified` (`YYYY-MM-DD`). To attribute a
+*specific claim*, give the source an `id` and cite it with a markdown footnote
+whose label is that id: `sharded daily.[^ga4-schema]` — keyed, not positional,
+so a reordered list cannot misattribute silently. `lint`'s provenance category
+checks the join in both directions (`unattributed_claim`, `unused_source`).
+Uncited claims are exactly how a bundle decays into folklore nobody trusts.
+("Citations" is the legacy v0.1 spelling: provenance lived in a body
+`# Citations` list, which v0.2 retires — §13.1 keeps it readable, and `lint`'s
+Migration findings tell you what to move where.)
+
+### Say who did it, in §7's spelling <!-- check:unprefixed_actor -->
+Every `by` — `generated.by`, `verified[].by` — takes an actor in one of §7's
+three forms: `<producer>/<version>` (an automated producer, e.g.
+`reference_agent/gemini-2.5-pro`), `human:<id>`, or `process:<id>`. The one
+producer MUST: content a person hand-authored **or confirmed** is marked
+`human:<id>` — §5.3 derives the trust tier from exactly this prefix, so a human
+sign-off written `by: owner` silently reads as *machine-confirmed*, the exact
+downgrade the tier system exists to prevent (`lint`'s `unprefixed_actor` nudges
+it). `sources[].author` is looser on purpose: §7 does not govern it, and the
+SPEC's own examples use `team:<id>` there — any honest attribution works.
+
+### Trust is derived, never stored (§5.2/§5.3) <!-- rule:okf-trust-derived -->
+`generated` records how the current content was produced (`by` REQUIRED within;
+`at` an ISO 8601 datetime); `verified` lists who confirmed it against its
+sources — they stay separate because who *wrote* a concept need not be who
+*checked* it. Consumers derive the tier: no `verified` ⇒ unverified; machine
+actors only ⇒ machine-confirmed; any `human:` verifier ⇒ human-reviewed. Never
+write a tier or a credibility score into frontmatter — it would be subjective,
+unportable, and stale the moment the next verification lands. And never invent
+provenance: a concept whose history you do not know is honestly unverified.
+Prefer **block style** for these mappings (`generated:` with indented
+`by:`/`at:`): the flow spelling (`{ by: human:x, at: …T10:00:00Z }`) is legal
+YAML, but the colons inside its values trip older libyaml parsers, and block
+style is the spelling every parser accepts.
+
+### Lifecycle: `status` and `stale_after` (§5.4/§5.5) <!-- rule:okf-lifecycle -->
+`status` is one of `draft | stable | deprecated`; absent means `stable`, so
+declare it only when it says something. `stale_after` is an absolute
+`YYYY-MM-DD` — the author's own "do not trust me past this date", stale **on**
+the day itself. Use it for knowledge with a known shelf life (a quota, a
+migration window); `lint`'s `expired` check reports the ones whose date has
+passed. It is a declared expiry, distinct from the `--stale-after` *flag*,
+which is a reader-supplied age cutoff — see [cli/checks.md](cli/checks.md),
+rule `okf-two-clocks`.
 
 ### Capture the non-obvious — not what code already says <!-- rule:okf-non-obvious -->
 A bundle that restates function signatures or config keys goes stale the moment
@@ -98,19 +170,61 @@ to it rather than duplicating it.
 
 ### Write for both readers at once <!-- rule:okf-dual-audience -->
 Use structural markdown so an agent can extract deterministically and a human can
-skim: headings, tables, fenced code, lists. Conventional headings a reader expects
-are `# Schema` (field/column tables), `# Examples`, and `# Citations`. Fill
-recommended frontmatter — `title`, `description`, `tags`, `timestamp` (ISO 8601) —
-whenever it aids consumption.
+skim: headings, tables, fenced code, lists. Conventional headings a reader
+expects are `# Schema` (field/column tables), `# Examples`, and — on an Attested
+Computation only — `# Computation` (§10.3). Fill recommended frontmatter —
+`title`, `description`, `tags`, `generated` — whenever it aids consumption.
+
+### Attested computations are contracts (§10) <!-- rule:okf-computation-untouched -->
+A concept of type `Attested Computation` carries a sanctioned computation:
+`runtime` is REQUIRED (it fixes what `parameters` mean), the computation is
+provided **exactly one way** — a body `# Computation` fence *or* a
+`computation:` path, never both — and `executor`/`attester` name how it runs
+and how a run is checked. The one new MUST NOT in v0.2 binds the *consumer*: an
+agent MAY bind values for the declared `parameters` and MUST NOT author or edit
+the computation itself. A "fixed" query silently stops being the one a person
+sanctioned; if it is wrong, tell a human. The
+[attested-computation template](../templates/attested-computation.md) carries
+the whole contract.
 
 ### Reserved files <!-- rule:okf-reserved-files -->
 `index.md` is a directory listing and carries **no frontmatter** — with one
 exception: the **bundle-root** `index.md` is the only index that may carry
-frontmatter, and it may carry *only* `okf_version: "0.1"` (§11; `validate` §9.3
-flags any other key there). `log.md` is an ISO-dated change history, newest first.
-Never use these names for concepts. Templates:
-[concept](../templates/concept.md), nested [index](../templates/index.md),
-bundle-root [root-index](../templates/root-index.md), [log](../templates/log.md).
+frontmatter, and it may carry *only* `okf_version: "0.2"` (§8 states the
+exception inline; `validate` flags any other key there under §11 condition 3).
+`log.md` is an ISO-dated change history, newest first. Never use these names for
+concepts. Templates: [concept](../templates/concept.md),
+[attested computation](../templates/attested-computation.md),
+nested [index](../templates/index.md), bundle-root
+[root-index](../templates/root-index.md), [log](../templates/log.md).
+
+### The log records what shipped, not how it shipped <!-- rule:okf-log-durable-only -->
+`log.md` carries durable knowledge and shipped behavior — never the process that
+produced them. A bug fixed in a release is an entry; the review rounds that found
+it are not. A capability that shipped is an entry; the iterations it took to
+stabilize it before it shipped are not. When a change taught a lesson, the lesson
+belongs in the concept it is about, stated as a principle, and the log entry
+*points* at that concept instead of re-narrating the rounds — a reader finds it
+where the subject lives, not by reading history.
+
+The bar is what a reader six months out needs: *what changed and why it matters*,
+never *how many passes it took to get there*. Watch for the shape this invites —
+the newest entries sit at the top where every reader lands, so a stretch of work
+stabilized by iteration accretes "round N found M defects" exactly where a durable
+summary belongs.
+
+## Migrating from v0.1 (§13)
+
+A v0.1 bundle is consumable forever — §13.1 sanctions reading `timestamp` as
+`generated.at` and a `# Citations` list as `sources` — so migration is
+convenience, not rescue. Exactly two spellings retire: `timestamp` (move the
+value under `generated: { by: <actor>, at: … }` — the actor is the one thing no
+tool can derive; ask, or write `human:<maintainer>` when the history is human)
+and the body `# Citations` section (lift each item into `sources`, add `id`s
+and `[^id]` footnotes where the body cites, delete the section). `lint`'s two
+Migration findings name the files; the [migrate playbook](../playbooks/migrate.md)
+walks the rewrite, and `okf lint <dir> --only legacy_timestamp,legacy_citations
+--fail-on info` is the mechanical done-check.
 
 ## Playbooks
 
@@ -119,7 +233,7 @@ verb (search, produce, migrate, maintain, consume, curate, doctor), routed by th
 Commands table in [SKILL.md](../SKILL.md). The Closeout below is their shared
 finishing gate.
 
-## Closeout — the finishing gate
+## Closeout — the finishing gate <!-- rule:okf-closeout-gate -->
 
 `produce` step 6 and `maintain` steps 4–7 both land here: before calling an
 authoring task done, walk this once. It is the repo's "turn every task into a check
@@ -129,10 +243,12 @@ grep can't:
 - **Index enumerations** — every `index.md` that lists what you added, renamed, or
   removed is updated; re-run `okf index` and eyeball each listing against reality.
   Easy to skip, expensive to miss — this is the check that was missing.
-- **`log.md`** — a dated entry, newest first.
-- **Timestamps** bumped on the concepts you touched.
-- **`validate`** — zero §9 errors.
-- **`lint`** — cheap findings cleared; pass `--stale-after` when concepts carry
-  timestamps (freshness is off by default).
+- **`log.md`** — a dated entry, newest first, and durable only (rule
+  `okf-log-durable-only` above).
+- **`generated.at`** bumped on the concepts you touched (and `generated.by` says
+  who touched them — you, in §7's spelling).
+- **`validate`** — zero §11 errors.
+- **`lint`** — cheap findings cleared; `expired` reports out of the box, and
+  `--stale-after` adds the reader-side age cutoff when you want one.
 - **`loose` review + tag curation** — the two semantic passes (maintain steps 6–7);
   worth a pass in `produce` too on a non-trivial bundle.

data/lib/okf/skill/reference/cli/checks.md

@@ -0,0 +1,171 @@
+# Checking a bundle — `validate`, `lint`, `loose`, `references`
+
+Kind: reference. Answers: what makes a bundle non-conformant, what each lint
+check means and the severity it is pinned at, which of the two clocks a flag
+reads, and which dangling pointers each surface can and cannot see.
+
+The shared contract — `@slug` refs, exit codes, `--json`, the filters — is in
+[cli.md](../cli.md) and is not repeated here.
+
+## validate — the hard gate (§11)
+
+Implements the spec's §11 conformance definition exactly (its words are in
+[SPEC.md](../SPEC.md), routed by [spec-map.md](../spec-map.md)):
+
+- **§11 cond. 1** every non-reserved file has a parseable YAML frontmatter block;
+- **§11 cond. 2** every such block has a non-empty `type`;
+- **§11 cond. 3** any `index.md`/`log.md` present follows §8/§9 (a nested
+  `index.md` has no frontmatter, a root `index.md` carries only `okf_version`,
+  `log.md` date headings are ISO `YYYY-MM-DD`).
+
+`ERROR`s are the three conditions above; the bundle is non-conformant until every
+one is fixed. `warn`s are soft — missing recommended fields, non-list tags, an
+unparseable timestamp, **broken cross-links, which §6.1 explicitly tolerates**,
+the shape of every §5/§10 family (`generated` not a mapping, a non-integer
+`usage_count`, a `stale_after` that is not `YYYY-MM-DD`, a missing `runtime` on
+an Attested Computation, …), and an `okf_version` the gem does not know (read
+best-effort under §12; an absent one never warns). Absence of an optional family
+is never a fault — a pure v0.1 bundle validates with zero warnings.
+
+In `--json`, every warning carries `check` (a stable id) and `source` — `spec`
+when the SPEC's own words state the rule, `convention` for a shape this gem asks
+for beyond them (`verified[].by` presence, integer `usage_count`, a per-entry
+`usage_window` mapping, `parameters[].name`, `executor`/`attester` `resource`).
+Gate on `source` when you want only the spec-normative set; errors keep their
+two-key `{ path, message }` shape. Fix warnings when cheap; never block on them.
+Use `--json` in CI.
+
+## lint — curation quality (advisory)
+
+Asks the complementary question to `validate`: not "is this legal OKF?" but "is
+this well-curated, navigable, trustworthy?" — precisely over the things §11
+forbids `validate` from rejecting. It has its own report, never emits
+conformance errors, and **exits `0` even with findings** unless you pass
+`--fail-on warn` (exit 1 on any `warn` finding) or `--fail-on info` (exit 1 on
+any finding at all).
+
+**Severity is API.** Every check has a pinned level — `warn` or `info` — and
+machine consumers gate on it, so the levels below are stable, not advisory. A
+finding you want to gate on that is `info` gets `--fail-on info` (usually with
+`--only`), never a hope that its severity changes. <!-- rule:okf-severity-is-api -->
+
+Eight categories, each backed by individual checks (severity in brackets):
+
+- **Reachability** — `orphan` [warn], `not_in_index` [warn],
+  `disconnected_component` [info], `unlinked` [info]
+- **Backlog** — `missing_concept` [info], `broken_index_entry` [warn]
+- **Completeness** — `stub` [info], `missing_title` [info],
+  `missing_description` [info], `missing_generated` [info] (quiet on either
+  spelling — a legacy `timestamp` still counts as a recorded change)
+- **Freshness** — `expired` [info] (§5.5: past the concept's own declared
+  `stale_after`, on the day itself), `stale` [warn] (older than the
+  reader-supplied `--stale-after` cutoff, keyed on `generated_at`)
+- **Provenance** — `uncited_external` [info] (external body links and no
+  sources, in either spelling), `broken_source` [warn] (an in-bundle `.md`
+  source target that names no concept; URLs and scope descriptors are out of
+  scope, and a non-`.md` asset is out of reach — the reader models concepts,
+  so lint never sees the file; `okf references` is the view that checks those
+  pointers), `unattributed_claim` [warn] (a footnote
+  no `sources[].id` answers — it *misattributes* a claim, which is why it
+  outranks its join-twin), `unused_source` [info] (a keyed source no footnote
+  cites — slack, not a defect), `unprefixed_actor` [info] (a `verified[].by`
+  outside §7's three forms reads as machine-confirmed; a `generated.by`
+  outside them feeds no tier but leaves a reader unable to tell a person
+  from a process; info so it informs, never blocks). A missing `generated.by` is the *validator's* warning —
+  REQUIRED-within is shape, not curation — so lint never double-reports it
+- **Attestation** — `incomplete_computation` [warn] (an Attested Computation
+  providing its computation neither way, or both ways — §10.3 says a
+  `computation:` path is used *instead of* the body fence),
+  `broken_attestation_ref` [warn] (on an `Attested Computation`, a
+  `computation`, `executor.resource` or `attester.resource` naming an
+  in-bundle `.md` that is not there — a contract no consumer can follow; the
+  keys are read only on that type, since §4.1 lets any other concept use them
+  for its own purpose). Its reach is exactly the `.md` files: URLs are out
+  of scope, and a `.sql` or `.py` target is invisible to *every* check here,
+  because the linter reads the concept model and the model carries only
+  markdown — `okf references` is the surface that sees those files and reports
+  a pointer that misses, whatever the extension. Remember §6.2 reads a bare
+  `references/…` as relative to the concept, so from a nested concept it wants
+  the leading `/`
+- **Migration** — `legacy_timestamp` [info], `legacy_citations` [info]: one
+  finding per bundle naming the files still in a retired v0.1 spelling, with
+  the rewrite instructions in the message. Info on purpose — §13 says a v0.1
+  bundle is consumable forever, so `--fail-on warn` must not turn red on one.
+  A migration campaign gates explicitly:
+  `okf lint <dir> --only legacy_timestamp,legacy_citations --fail-on info`,
+  exit 1 until clean.
+- **Hygiene** — `duplicate_title` [info], `unused_reference_def` [info],
+  `undefined_reference` [warn], `self_link` [info], `log_order` [info] (§9
+  reads a log newest-first; disorder is slack, never a §11 error)
+
+`--only` / `--except` filter by the **individual check names above**, not the
+category labels — `okf lint <dir> --only orphan,stub` works; `--only reachability`
+is an error. Two knobs tune specific checks: `--min-body N` sets the `stub` body
+threshold in characters (default 50), and `--stale-after DUR` sets the `stale`
+cutoff — a duration like `90d` or `12w`, or an ISO date like `2026-01-01` (a bare
+number is rejected).
+
+**Two different clocks, one unlucky name.** The `--stale-after` *flag* and the
+`stale_after:` *frontmatter field* are different mechanisms that happen to share
+a spelling. The flag is the **reader's** age cutoff: "flag anything not touched
+since DUR", keyed on `generated_at`, feeding the `stale` check. The field is the
+**author's** declared expiry: "do not trust this past DATE", feeding the
+`expired` check. Never read one as the other, and never show them adjacent
+without the distinction. <!-- rule:okf-two-clocks -->
+
+**The clock is explicit.** `expired` compares against a day the CLI supplies —
+today by default, or `--today YYYY-MM-DD` for a reproducible report (CI wants
+this). The pure library runs no clock check unless handed `today:`, and every
+clock-gated check that was selected but could not run is *named* in
+`stats.skipped_checks` (the human report prints one `skipped:` line) — a gate
+that is sometimes absent and does not confess converts "unchecked" into
+"checked and fine".
+
+The report's stats carry the bundle's posture too: `trust` (the §5.3 tier
+distribution, in the hyphenated wire spelling) and `status` (effective-status
+frequency).
+
+`lint --json` is the structured substrate you consume to reason about the two
+things lint deliberately does **not** compute — contradictions and *semantic*
+staleness — which need understanding of meaning.
+
+## loose — files with no graph connections (by folder)
+
+Lists the **loose** files — concepts with graph **degree 0**: no cross-links in
+*or* out — grouped by folder. It is a focused, folder-organized view over `lint`'s
+`unlinked` check (`okf loose <dir>` ≈ `okf lint <dir> --only unlinked`, regrouped),
+for the "which files float in the graph?" question. Advisory: **exits `0`**; `--json`
+emits `{ bundle, count, loose: [{ id, title, dir }] }`.
+
+**Loose ≠ orphan** — the trap. `lint`'s `orphan` is about *reachability*, and an
+`index.md` listing makes a file reachable, so an indexed file is never an orphan.
+But an index listing is **not a graph edge**: a file can be listed in an index yet
+have no cross-links, so it floats in the graph while `lint` reports it as reachable.
+`loose`/`unlinked` catch exactly that gap. A loose file is not automatically a
+defect — a terminal leaf (a backlog item, a spec reference) can be loose by design;
+`loose` surfaces the set so you can judge intent (see the
+[maintain playbook](../../playbooks/maintain.md)).
+<!-- rule:okf-loose-not-orphan -->
+
+## references — the `references/` inventory (§6.3)
+
+Lists every file under `references/` — including the non-markdown ones no other
+verb can see, since the concept model carries only markdown — with which
+concepts cite each file through the §6.2 path-valued fields (`resource`,
+`sources[].resource`, `computation`, `executor.resource`, `attester.resource`),
+plus every pointer into `references/` that resolves to nothing. Advisory:
+**exits `0`** even with dangling pointers — the findings are the output. JSON:
+`{ bundle, dangling, count, references: [{ path, dir, kind, referenced_by }] }`,
+with `--fields`/`--except` projecting the rows. A file that is itself a concept
+(§6.3 allows both) is marked `kind: "concept"`; body links are the graph's
+business and are not counted here.
+
+**The dangling list is where §6.2's bare-path trap surfaces.** A bare
+`references/attesters/rev.py` written from `metrics/` resolves relative to the
+concept — `metrics/references/attesters/rev.py`, nothing — and when the
+leading-slash spelling would have hit, the entry says so:
+`/references/attesters/rev.py exists — missing leading slash?`. Reach is any
+extension, which is exactly what `broken_source` and `broken_attestation_ref`
+cannot offer (their exemptions above), so run it wherever a bundle carries
+attester code or computation files.
+

data/lib/okf/skill/reference/cli/graph.md

@@ -0,0 +1,49 @@
+# `graph` — the raw structure
+
+Kind: reference. Answers: what a full dump costs and how to plan a traversal
+without paying it, plus the two rankings that give `refine` its evidence —
+`--hubs` by concept, `--traffic` by directory.
+
+The shared contract — `@slug` refs, exit codes, `--json` — is in
+[cli.md](../cli.md) and is not repeated here.
+
+Prints the node/edge graph. `--json` emits a machine-readable dump — the
+`bundle`/`slug` head every view carries, then `nodes` (with
+`id`/`type`/`title`/`description`/`tags` **and, by default, every `body`** — the
+part that dominates the bytes on a real bundle) plus `edges` — you can pipe into
+other analysis. A concept with a missing *or blank* `type` indexes under
+`Untyped`: §11 condition 2 rejects both identically, so both land in one bucket. To *plan* a traversal, structure is all you need: `--no-body`
+drops each node's body, and `--minimal` ships only `id`/`title` plus the type/tag
+indexes — the lean shape the `server` page boots from. Reach for the full dump
+only when the task truly consumes every body; for one question, the
+[search verb](search.md) is orders cheaper.
+
+`--hubs` swaps the dump for the **inbound ranking**: every concept with at
+least one inbound link, ranked by inbound degree, each with its links grouped
+by *source top-level dir* (`core/status  ×3   flows 2, billing 1`) — the evidence for
+[refine](../../playbooks/refine.md)'s hub origin test ("is this hub well-homed?").
+A source at the bundle root counts under `(root)`. JSON: `{ bundle, count,
+hubs: [{ id, top_dir, inbound, by_top_dir: { <top_dir>: n } }] }`. Advisory read, exit 0;
+`--minimal`/`--no-body` shape node payloads and change nothing here.
+
+`--traffic` asks the same question one grain coarser: **directories**, not
+concepts. Every concept collapses into the directory it lives in and every link
+between two directories collapses into one weighted arc, so a bundle's wiring
+becomes a table you can read — measured on one 47-concept bundle, 227 links
+collapsed into 50 arcs, of which the fitted cut draws 22. Each row carries the
+directory's traffic split three ways
+(`internal` / `out` / `in`) plus **cohesion**, its internal share of the total:
+the evidence for [refine](../../playbooks/refine.md)'s container test, where
+`--hubs` only ever answered about concepts. Rows lead with the lowest cohesion,
+so the directories with a case to answer come first, and a directory with no
+traffic at all prints `—` rather than a `0%` it did not earn.
+
+`--cut N` is the least arc weight drawn. It defaults to a value **fitted to the
+bundle** — enough arcs for roughly 1.5 per directory, floored at 8 — because a
+fixed weight cannot serve both ends: measured at weight 3 across ten bundles it
+left 2 arcs on one and 136 on another. The JSON says which you got. Cohesion is
+computed over *every* arc and never the drawn ones, so tightening the cut
+changes the picture and never the evidence. JSON: `{ bundle, cut, fitted, dirs:
+[{ dir, parent, count, subtree, internal, out, in, cohesion }], arcs: [{ source,
+target, weight }], total_arcs }` — a fraction of the full dump (2.6 KB against
+27 KB on that bundle), and the shape rather than the contents. Advisory, exit 0.

data/lib/okf/skill/reference/cli/map.md

@@ -0,0 +1,98 @@
+# Orienting — `index` and `dirs`
+
+Kind: reference. Answers: which view to run first on a bundle you do not know,
+how `--dir`, `--depth` and the ancestor chain compose, what a synthesized
+listing means, and how to keep the map from paging the whole bundle.
+
+The shared contract — `@slug` refs, exit codes, `--json`, the filters — is in
+[cli.md](../cli.md) and is not repeated here.
+
+## index — the progressive-disclosure map (§8)
+
+The "orient before you read" view, and the read verb that sees the layer the
+concept views can't: `index.md` files are reserved/structural, so
+`catalog`/`files`/… never show them (in the browser, the Indexes tab and
+folder clicks render this same map). `okf index <dir>` prints one entry per directory
+that holds concepts or carries an `index.md`, root first — the authored index body
+(frontmatter stripped), a `type`/`tag` rollup over the concepts that live directly
+there, its child directories, and the concept listing. Run it first when picking up
+an existing bundle: it is the cheapest high-signal orientation, and it surfaces
+enumeration drift a grep can't (you can't grep for a listing entry that is *missing*).
+
+`--dir PATH` narrows to a directory **and everything below it**, and is
+**repeatable** — `--dir model --dir format` shows both; `root` (or `.`) names the
+bundle root, unless the bundle really has a `root/` directory, which owns the
+word. A `--dir` also brings the **chain from the root down to it**, so a branch is
+never shown adrift of the authored context that says what it is — the root
+`index.md`'s prose first among it. Those rows print with a leading `↑` and carry
+`ancestor: true`; `--no-ancestors` drops them. Ascent and descent are separate
+axes, so `--depth` never bounds the chain: `--dir X --depth 0` is X alone, plus
+how you get to X. A `--dir` that names nothing gains no chain — a lone root row
+would read as a partial answer to a query that matched nothing.
+
+`--depth N` bounds how far below the starting point the map reaches
+(the `--dir` when one is given, else the bundle root), counted **relatively**:
+`--depth 1` is the top of the tree, `--dir X --depth 1` is one branch of it, and
+the pair walks down a level at a time.
+
+**On a bundle of any size the map is unreadable whole** — every directory is a
+section, and even `--no-body` keeps one listing row per *concept* — so narrow
+rather than paging it: `okf dirs` is the orientation, `--dir <branch> --depth 1`
+is the step down into it,
+and `--except body,listing` on top of either is the lean JSON skeleton. Full
+`index` output on a few hundred concepts runs to hundreds of KB; the same map at
+`--depth 1` is a couple of KB.
+
+`--no-body` drops the prose to a
+skeleton (headers, rollups, child pointers). For a directory that has concepts but
+**no `index.md`**, the listing is **synthesized** from the concepts' descriptions
+and tagged `(no index.md)` — §8 explicitly permits synthesizing a map on the fly.
+
+It is a **read view**: advisory, always exit 0. A synthesized directory is a
+*signal* (a map worth writing), never a defect — `index` emits no lint findings and
+never fails a bundle. JSON: `{ bundle, count, directories: [{ dir, index_path,
+present, synthesized, count, types, tags, subdirs, body, listing: [{ id, title,
+description, type, tags }] }] }` — `ancestor` marks a row that is there to place
+the branch rather than to answer about it.
+
+## dirs — the bundle's clusters and their sizes
+
+`okf dirs <dir>` lists every directory the bundle has — the ones holding
+concepts, the ones carrying an `index.md`, and the empty intermediates that only
+exist to connect the tree — with the number of concepts living **directly** in
+each and the number in its **subtree**. A cluster *is* a directory here, so this
+is the view that tells you what `--dir` can be pointed at and how much sits
+behind each choice.
+
+Two numbers, because one cannot answer the question. `count` is direct, so the
+column sums to the bundle's concept total and a dir holding only sub-directories
+reads `0` rather than a hidden rollup. `subtree` is defined as *exactly what
+`--dir <that row>` returns*, so the row and the flag can never disagree — which
+is also why the root's subtree is its own direct count (`.` is a prefix of
+nothing). Without it a truncated listing is all zeroes at the top of a deep tree,
+which is where you most need to know where the mass is. The human table shows the
+second column only where some dir actually nests.
+
+`--dir PATH` (repeatable) narrows to a directory and its subtree, and brings the
+**chain up to the root** with it so the branch is placed rather than shown
+adrift — those rows are marked `↑`, carry `ancestor: true`, and stay out of
+`total` (`--no-ancestors` drops them). `--depth N` keeps only N levels below the
+starting point — the `--dir` when one is given,
+the bundle root otherwise. Relative, not absolute, so `--dir a/b --depth 1`
+reads "a/b and one level under it" without your first working out how deep `a/b`
+is. `--depth 0` is the starting point alone. A `--depth` that is not a whole
+number is a usage error (exit 2).
+
+**This is the first command to run on a bundle you do not know** — the same first
+move [SKILL.md](../../SKILL.md) prescribes. `okf dirs <dir>` is one row per
+directory, so its size tracks the tree rather than the concept count: it tells
+you the shape and where the weight sits, `--depth 1` trims it further on a deep
+bundle, and you then descend with `okf index --dir`, one level at a time.
+
+The root prints `(root)` and stores `.` — the split every grouped view keeps, so
+a table and its `--json` never disagree about which spelling is the data. JSON:
+`{ bundle, total, count, dirs: [{ dir, ancestor, count, subtree, subdirs }] }`,
+root first. `count` is rows printed, chain included; `total` sums the direct
+counts of the rows you actually asked for, which is what keeps a row's `subtree`
+equal to the `total` that `--dir` on that row returns.
+