okf 1.8.0 → 1.10.0

data/lib/okf/skill/SKILL.md

@@ -5,14 +5,15 @@ description: >-
   directory of markdown files with YAML frontmatter that humans and agents read
   from one source. Use when capturing knowledge into a bundle (a service, schema,
   metric, decision, runbook: "document this in OKF", "capture this as a concept"),
+  converting existing docs into one ("migrate/OKFy our docs into a bundle"),
   retrieving from one without reading it whole ("what do we know about X?", "where
-  is X documented?", "search the bundle"), updating one after code or docs change
-  ("update the knowledge bundle"), checking its conformance or curation quality
-  ("validate/lint the bundle"), serving or rendering it as a graph, or working in a
-  repo that already carries an OKF bundle — a `.okf/` directory or a root `index.md`
-  carrying `okf_version`.
+  is X documented?", "search the bundle"), updating one after code or docs
+  change ("update the knowledge bundle"), checking its conformance or curation
+  quality ("validate/lint the bundle"), serving or rendering it as a graph, or
+  working in a repo that already carries an OKF bundle — a `.okf/` directory or a
+  root `index.md` carrying `okf_version`.
 user-invocable: true
-argument-hint: "[search|produce|maintain|consume|<okf-cli-verb>] [dir] [--flags]"
+argument-hint: "[search|produce|migrate|maintain|refine|consume|curate|doctor|<okf-cli-verb>] [dir|@slug] [--flags]"
 allowed-tools: Read Write Edit Grep Glob Bash
 ---
 
@@ -66,12 +67,16 @@ earn your keep as the expert, not the executable.
 
 ## The CLI is your eyes — you are the judgment
 
-Guard once, then trust it — the `okf` executable answers every mechanical question
-deterministically, and its read views show everything the browser UI does:
-
-```bash
-command -v okf >/dev/null || echo "okf CLI missing — install: gem install okf (or from a checkout: cd gem && bundle exec rake install)"
-```
+The `okf` executable answers every mechanical question deterministically, and its
+read views show everything the browser UI does. **Don't probe for it — just run
+the verb.** A proactive `command -v okf` before every task spends a whole tool
+round proving what the next command reveals for free; the CLI's own failure is a
+cheaper, truer signal. (The two deliberate exceptions are [menu](playbooks/menu.md)
+and [doctor](playbooks/doctor.md) — both decide *whether to install*, so they check
+first.) The one distinction to hold: a shell `okf: command not
+found` is the *only* thing that means "install it" (→ [doctor](playbooks/doctor.md));
+every line that starts `error:` is okf *answering* — a bundle or usage result to
+read and act on, never a missing toolchain to send to doctor.
 
 Don't memorize the surface — `okf --help` maps every verb, `okf <verb> --help` its
 flags. The division of labour is the whole game:
@@ -98,7 +103,7 @@ shapes, the tag-curation views, the server's trust boundary.
 ## Orient before you touch anything
 
 Picking up a bundle you don't already know — to consume or maintain — run `okf
-index <dir>` (the §6 map: every directory's index body, rollups, and listings) and
+index <dir|@slug>` (the §6 map: every directory's index body, rollups, and listings) and
 read `log.md` (the §7 baseline of what changed last) **before** greping or opening
 leaves. It is the cheapest high-signal context, and the only reliable way to catch
 enumeration drift: **grep cannot find an index entry that is missing** — you can't
@@ -118,14 +123,22 @@ when you need chapter and verse.
 
 **No subcommand?** Infer intent: "document this / capture X" → `produce`;
 "convert / migrate / OKFy these existing docs into a bundle" → `migrate`; "the
-code changed, update the docs" → `maintain`; "what do we know about X / where
-is X documented" → `search`; a repo already carrying a bundle plus a task
-needing its knowledge → `consume`; "check / graph / preview it" → run the
-matching CLI verb and interpret the result. When genuinely ambiguous, ask.
-
-**Which directory?** Use the path given. Otherwise default to `.okf/` at the repo
-root, but first detect whether the project already keeps its bundle elsewhere
-(e.g. `docs/`) and prefer that. Commit the bundle alongside the code it describes.
+code changed, update the docs" → `maintain`; "restructure / rebalance the
+bundle / is the structure right / get more out of it" → `refine`; "what do we
+know about X / where is X documented" → `search`; a repo already carrying a
+bundle plus a task needing its knowledge → `consume`; "check / graph / preview
+it" → run the matching CLI verb and interpret the result. When genuinely
+ambiguous, ask.
+
+**Which target?** A leading `@` is a *registry ref*, not a path: `@slug` names a
+bundle registered with `okf registry set`, bare `@` the default — route it
+straight to `okf <verb> @slug` and skip the directory hunt (`okf search` spans
+several: `@a @b`, or `@all`). A plain path is used as given. Given no target and a
+cwd that carries no bundle, `okf registry list` is the next move, not a hunt
+across sibling directories. Producing a *new* bundle with no path? Default to
+`.okf/` at the repo root, but first detect whether the project already keeps its
+bundle elsewhere (e.g. `docs/`) and prefer that; commit it alongside the code it
+describes.
 
 **Target isn't a bundle?** When a verb points at a directory that holds markdown
 but no root `index.md` carrying `okf_version` — `validate` failing wholesale on
@@ -147,15 +160,17 @@ Read the referenced playbook before executing — it *is* the procedure.
 | `produce`  | Author | create or extend a bundle | [playbooks/produce.md](playbooks/produce.md) |
 | `migrate`  | Author | convert existing docs in place: frontmatter + reserved files, bodies verbatim | [playbooks/migrate.md](playbooks/migrate.md) |
 | `maintain` | Author | sync the bundle's content with reality after a change | [playbooks/maintain.md](playbooks/maintain.md) |
+| `refine`   | Author | optimize the bundle's structure: evidence-driven, cohesion-first; proposes, never auto-applies | [playbooks/refine.md](playbooks/refine.md) |
 | `consume`  | Use    | use the bundle as context for a task | [playbooks/consume.md](playbooks/consume.md) |
 | `curate`   | Curate | structural upkeep as it stands: validate + lint + loose | [playbooks/curate.md](playbooks/curate.md) |
 | `doctor`   | Setup  | install and verify the CLI, then doctor the bundle | [playbooks/doctor.md](playbooks/doctor.md) |
-| `<okf-cli-verb>` | Read | validate, lint, loose, index, catalog, files, tags, types, stats, graph, server, render, registry, skill | `okf <verb> --help` + [reference/cli.md](reference/cli.md) |
+| `<okf-cli-verb>` | Read | validate, lint, loose, index, catalog, files, tags, types, stats, graph, server, render, registry, skill — **plus any verb an installed extension adds** (`okf help` is authoritative, this list is not) | `okf <verb> --help` + [reference/cli.md](reference/cli.md) |
 
-Two boundaries worth keeping sharp: `curate` is structural upkeep only — when
-the *content* no longer matches reality, that is `maintain` — and `doctor` is
-the one playbook that does not assume the CLI is installed. In Claude Code with
-the okf plugin, `/okf:gem` routes these same verbs.
+Three boundaries worth keeping sharp: `curate` is structural upkeep only — when
+the *content* no longer matches reality, that is `maintain`, and when the
+content is right but the *shape* underserves retrieval, that is `refine` — and
+`doctor` is the one playbook that does not assume the CLI is installed. In
+Claude Code with the okf plugin, `/okf:gem` routes these same verbs.
 
 ## The lifecycle is a flywheel, not phases
 

data/lib/okf/skill/playbooks/consume.md

@@ -1,8 +1,10 @@
 # Playbook: consume — use a bundle as context
 
-1. **Orient first** (the [SKILL.md](../SKILL.md) reflex): `okf index <dir>` maps the
-   whole bundle in one pass — every directory's index body, rollups, and listings —
-   and `log.md` gives recent history. Then follow links only into the concepts the
+1. **Orient first** (the [SKILL.md](../SKILL.md) reflex): `okf index <dir|@slug>` maps
+   the whole bundle in one pass — every directory's index body, rollups, and listings —
+   and `log.md` gives recent history. Address a registered bundle by `@slug` (bare
+   `@` = the default); if the cwd carries no bundle, `okf registry list` finds one
+   instead of a directory hunt. Then follow links only into the concepts the
    task needs. For a *pointed question* rather than broad context, switch to the
    [search playbook](search.md): map → finder (`okf search`) → only the winning
    bodies. For a large bundle, `okf graph --json --minimal` gives the whole link

data/lib/okf/skill/playbooks/curate.md

@@ -9,7 +9,9 @@ reachability, backlog, completeness, hygiene. It is not `maintain`, the
 skill's workflow for when the project changed and the bundle's *content*
 must catch up with reality; reach for that one when what is written stopped
 being true. Curating can surface semantic staleness, and when it does,
-switch to `maintain` for those concepts.
+switch to `maintain` for those concepts. And it never moves knowledge: when
+the content is right but the shape underserves retrieval — fat areas,
+mis-homed hubs, an uncurated tag layer — that is [refine](refine.md).
 
 1. Locate the bundle: the directory you were given, if any; otherwise a
    `.okf/` directory or a root `index.md` whose frontmatter carries

data/lib/okf/skill/playbooks/maintain.md

@@ -1,8 +1,9 @@
 # Playbook: maintain — keep a bundle in sync with reality
 
 Reach for this when the project changed and the bundle's *content* must catch
-up. The modelling craft behind steps 3 and 7 lives in
-[authoring.md](../reference/authoring.md).
+up. Restructuring the bundle itself — moving concepts, adding areas — is
+[refine](refine.md), not maintain. The modelling craft behind steps 3 and 7
+lives in [authoring.md](../reference/authoring.md).
 
 1. **Orient before hunting.** Run `okf index <dir>` (the §6 map — every directory's
    index body, rollups, and listings), read `log.md` (the §7 baseline: what changed
@@ -43,7 +44,7 @@ up. The modelling craft behind steps 3 and 7 lives in
      by design only through its index — leave it. **Terminal-by-design is not a
      defect.** Loose ≠ orphan: an index listing makes a file *reachable* (not an
      orphan) but is not a graph edge, so an indexed file can still float here.
-7. **Curate the tag vocabulary** <!-- rule:okf-tag-vocabulary --> when the pass
+7. **Curate the tag vocabulary** when the pass
    touched tags, or when `okf tags <dir>` shows a long tail of singletons. Run `okf tags <dir> --by area` and
    `--by type` — the grouped view is the analysis; read each group top-down:
    - **twins** — two tags riding the exact same concepts (equal counts sort them

data/lib/okf/skill/playbooks/menu.md

@@ -33,6 +33,11 @@ is the lede.
      the working tree has uncommitted changes to the code the bundle describes
      (`git status`), prefer **`maintain`**: that is exactly the drift it exists
      to close.
+   - **clean, but the shape strains** — one area dwarfing the rest in
+     `okf stats`, tags spread thin across areas in `okf tags --by area`, hubs
+     whose inbound links are mostly foreign in `okf graph --hubs` → offer
+     **`refine`** (evidence-driven restructuring; it proposes before it
+     touches anything).
 4. **Freshness is off by default.** If the bundle carries timestamps, note that a
    plain `lint` said nothing about staleness and `okf lint <root> --stale-after
    90d` is the check that would.

data/lib/okf/skill/playbooks/refine.md

@@ -0,0 +1,92 @@
+# Playbook: refine — restructure a bundle to get the most from OKF
+
+Reach for this when the bundle's *content* is right but its *shape* may not be:
+areas grown fat by additive passes, hubs homed by history, a tag layer that
+never became the second index. Refine optimizes the projection — the same
+knowledge, arranged to serve progressive disclosure, the emergent graph,
+cross-cutting tags, and capture-once-link-many. It is not [curate](curate.md)
+(upkeep of the structure as it stands) and not [maintain](maintain.md) (content
+catching up with reality): refine changes where knowledge lives, never what it
+says. Its permitted edits are structural — move a concept, extract a duplicated
+fact to one home, section an index, retag, relink, and write the connective
+sentence a link lives in; summarizing, updating, or correcting a body is
+maintain's job, reached by switching verbs, not by stretching this one.
+
+The frame that governs every move: the directory tree is a **lossy projection
+of the link graph**. A tree gives each concept one parent, so the tree encodes
+only the single dominant decomposition; every genuinely many-to-many
+relationship rides links and tags, never new directories. And cohesion outranks
+balance — a move has semantic cost, so balance is a tiebreaker and a fatness
+alarm, never the objective. <!-- rule:okf-cohesion-over-balance -->
+
+1. **Orient.** `okf index <dir|@slug> --no-body` (areas, fan-out, depth),
+   `log.md` (how the bundle grew), `okf stats` (totals). Additive growth
+   optimizes each pass locally, never the whole — that is the drift this
+   playbook corrects.
+2. **Measure — the CLI is the evidence.** Baseline `validate` / `lint
+   --stale-after` / `loose` first: refine assumes a sound bundle, and hard
+   errors are [curate](curate.md)'s job. Then the two structural reads:
+   - `okf tags <dir> --by area` — each row carries `count/total`, so a tag's
+     **locality** reads directly: a tag wholly inside one area names a *domain*
+     (the directories are right); one spread across areas names a *concern*.
+   - `okf graph <dir> --hubs` — concepts ranked by inbound links, each with
+     the areas those links come from: the **origin test** for every hub.
+3. **Diagnose — you are the judgment.** The measurements are evidence, never
+   verdicts:
+   - **Concerns never become containers.** A directory built around a spread
+     tag ("everything async") prunes nothing — most needs would enter it.
+     The cross-cut stays a tag. <!-- rule:okf-concern-not-container -->
+   - **A directory must prune.** The positive test for any area, existing or
+     proposed: does knowing "it's in there" eliminate a large, even slice? A
+     good node splits its parent into chunks that are nameable, mutually
+     exclusive, and roughly comparable in size. And small is not merge-worthy
+     on its own — a two-concept area that is a genuinely distinct domain
+     stays. <!-- rule:okf-directory-prunes -->
+   - **The hub origin test.** Inbound majority from the hub's own area:
+     well-homed, leave it. A dominant *foreign* area: that area is the better
+     home. Foreign majority with *no* dominant area: a shared primitive — the
+     only admission ticket into a shared-core area (without that test, a
+     `foundation/` rots into a `misc/`). Two comparable strong ties, one of
+     them home: stay and carry the other as a tag — moving trades one
+     imbalance for another. And in a design bundle expect the central
+     decisions to fail this test wholesale: that is centrality, not
+     mis-homing.
+   - **Fatness alarm, not fatness rule.** A fat area (≳20–25 concepts) wants
+     **heading sections inside its `index.md`** first — the same prune as
+     sub-directories, for zero extra hops and no new enumeration to keep
+     sound. Directory nesting pays only at hundreds of concepts, and only
+     where the index's own headings already form separable, nameable
+     sub-groups — fatness alone never justifies depth; the sections that
+     formed are the evidence the split exists.
+   - **Duplication.** Read the area overviews for a fact re-explained in
+     several (drifting tables are the tell); capture-once-link-many says
+     extract it into one concept and link from the rest. Extraction is the one
+     refine move that touches bodies, and it redistributes rather than
+     rewrites: assemble the canonical concept from the copies, keep every
+     copy's unique domain-specific detail (in the extract, or in the one-line
+     note left beside each link), and where the copies *disagree*, which is
+     true is a [maintain](maintain.md) question — verify against reality or
+     flag the conflict in the proposal, never silently pick a winner while
+     merging. <!-- rule:okf-extract-not-rewrite -->
+   - **Vocabulary twins.** The [maintain](maintain.md) tag-curation recipe
+     (twins, echoes, singletons) applies to `type` too — `okf types <dir>`.
+4. **Plan — tier by leverage ÷ churn, free levers first.** Tag curation,
+   index heading-sectioning, and extraction before any file move; a move only
+   when the origin test demands one, and each gated by **do-nothing**: skip it
+   unless its value beats its churn. Record what you *declined* and why — the
+   decline list is what stops the next pass from re-proposing it.
+5. **Propose — never auto-apply.** Refine's output is a short report (the
+   evidence, the tiers, the declines) plus a ready-to-run execution prompt the
+   user can hand back later: a scope line, the governing principles above, the
+   explicit prohibitions, the closeout gate as acceptance. Analysis and
+   execution are separate on purpose — the judgment is spent once, frozen, and
+   then executed without re-derivation. <!-- rule:okf-refine-proposes -->
+6. **Execute only on approval**, then walk the
+   [Closeout gate](../reference/authoring.md#closeout--the-finishing-gate):
+   every touched `index.md` re-enumerated, links absolute bundle-relative so
+   they survived the moves, a dated `log.md` entry carrying the *why*,
+   validate/lint/loose clean, and a before/after of step 2's evidence.
+
+Two traps: never split a cohesive cluster (a deliberately paired mirror flow)
+to hit a size band, and never tag a concept with its own directory's name — a
+group-name echo adds no edge.

data/lib/okf/skill/playbooks/search.md

@@ -6,9 +6,13 @@ can query cheaply is dead weight. The discipline is progressive disclosure
 (spec §6): every step pays a few hundred bytes to decide what the next step
 reads, and full bodies are read last, and only the winners.
 
-1. **Guard once**: `command -v okf`. Missing → [doctor](doctor.md). No CLI at
-   all → read the root `index.md`, then each relevant area's `index.md`, by hand.
-2. **Ingest the map and decide where to look.** `okf index <dir> --no-body` is
+1. **Just run it — no presence probe.** Point the finder at a path or an `@slug`
+   (a registered bundle; bare `@` = the default). Only a shell `okf: command not
+   found` means install (→ [doctor](doctor.md)); with no CLI possible at all, read
+   the root `index.md` then each relevant area's `index.md` by hand. No bundle in
+   the cwd? `okf registry list` names the registered ones — address them by
+   `@slug`, don't hunt sibling directories.
+2. **Ingest the map and decide where to look.** `okf index <dir|@slug> --no-body` is
    the skeleton: every directory with its concept count, types, tags, children.
    *You* do the semantic matching here — the question names a meaning, the map
    names areas; connect them by judgment, not string equality. When an area
@@ -17,8 +21,37 @@ reads, and full bodies are read last, and only the winners.
    <!-- rule:okf-search-map-first -->
 3. **Cut across with the finder when the question is lexical.** An exact
    symbol, an error code, a column name, a phrase — things structure won't
-   surface — go to `okf search <dir> <terms>` (terms AND together; `--regexp`
-   for patterns like `err_[a-z]+_409`). Scope it with what the map taught you:
+   surface — go to `okf search <dir> <terms>`. Terms AND together and are matched
+   **literally against raw text**, so an exact query means what it looks like: a
+   phrase, a dotted version (`7.2.0`), an underscored identifier (`customer_id`),
+   a mid-word fragment (`ustomer`) and a word written in `backticks` all match
+   the way you typed them. <!-- rule:okf-search-exact-identifiers -->
+
+   **Match the engine to the shape of the query, not to habit** — the default
+   answers most of them, and the two engines fail in opposite directions:
+   <!-- rule:okf-search-engine-choice -->
+
+   | Your query is | Reach for | Because |
+   |---|---|---|
+   | an identifier, version, path, phrase, or anything in `` `backticks` `` | *nothing — the default* | matched literally; the index shatters all of these |
+   | a mid-word fragment (`ustomer`) | *nothing — the default* | an infix is not a token, so the index cannot reach it |
+   | a pattern (`err_[a-z]+_409`) | `-e` | Ruby regexp over raw text; still the scan |
+   | a partial word (`dedup` → `deduplication`) | *nothing — the default* | substring covers prefixes, and suffixes and infixes too |
+   | a theme, where you want the best match to lead | `--engine index` | BM25+ ranks by relevance, not by summed field weight |
+   | possibly mistyped | `--fuzzy` | edit distance 0.2 × term length — the index's alone |
+   | being reconciled with the browser page | `--engine index` | same MiniSearch build, so the two rank alike |
+
+   The index has exactly **three** things the default lacks: relevance ranking,
+   typo tolerance, and page parity. Its `prefix` capability is not a fourth — a
+   substring match already reaches every prefix, so `prefix` is what the index
+   needs to *catch up*, not a reason to choose it.
+
+   **`--fuzzy` is an engine switch, not a mode.** It routes to the index, so a
+   run that only wanted a typo forgiven also gets token matching, shattered
+   identifiers and unfindable code spans. Fix the spelling and stay on the
+   default when you can. <!-- rule:okf-search-fuzzy-is-a-switch -->
+
+   Scope any of them with what the map taught you:
    `--area billing`, `--type Decision`, `--tag idempotency`, `--in body`.
    Matches rank by where they hit, and the snippet often *is* the answer.
    When the answer may live in another registered bundle, span them — leading
@@ -40,6 +73,13 @@ Anti-patterns, each a real token bill:
 - **Grep before map.** Grep cannot find the entry that is *missing*, and it
   returns line noise where `search` returns ranked concepts. Grep is the
   fallback when the CLI is absent, not the first move.
-- **Mechanical synonym retries.** The finder is exact by design; *you* are the
+- **Mechanical synonym retries.** The finder is exact by default; *you* are the
   fuzzy layer. When terms miss, learn the bundle's vocabulary — `okf tags
-  <dir>`, `okf types <dir>` — and re-ask in its own words.
+  <dir>`, `okf types <dir>` — and re-ask in its own words. `--fuzzy` forgives a
+  *typo*, not a wrong vocabulary, so it is the wrong reach for this.
+- **Flag-shopping a query that found nothing.** Cycling `--fuzzy`, then
+  `--engine index`, then `-e` over the same terms is guessing, and each engine
+  fails differently enough that one of them eventually returns *something* —
+  which is how a wrong answer gets found. Zero matches is usually a vocabulary
+  result, not an engine result: go back to the map and the tag list. Reach for a
+  different engine when you can say which property of the query needs it.

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

@@ -115,8 +115,9 @@ bundle-root [root-index](../templates/root-index.md), [log](../templates/log.md)
 ## Playbooks
 
 The step-by-step playbooks live in [../playbooks/](../playbooks/), one file per
-verb (produce, maintain, consume, curate, doctor), routed by the Commands table
-in [SKILL.md](../SKILL.md). The Closeout below is their shared finishing gate.
+verb (search, produce, migrate, maintain, consume, curate, doctor), routed by the
+Commands table in [SKILL.md](../SKILL.md). The Closeout below is their shared
+finishing gate.
 
 ## Closeout — the finishing gate
 

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

@@ -6,14 +6,13 @@ reimplemented in this skill. They run the deterministic `okf` executable shipped
 the companion gem — the single source of truth for OKF mechanics. Your job is to
 invoke it correctly and interpret the result, not to reason out conformance by hand.
 
-## Presence guard
+## When it isn't installed
 
-Check the tool exists before relying on it. If it is missing, the gem is not
-installed — say so and stop; never fabricate a result:
-
-```bash
-command -v okf >/dev/null || echo "okf CLI not found — install it: 'gem install okf' (or from a checkout: 'cd gem && bundle exec rake install')"
-```
+Don't probe for the tool before using it — just run the verb. A shell `okf:
+command not found` is the only thing that means the gem isn't installed: say so
+and stop (`gem install okf`, or from a checkout `cd gem && bundle exec rake
+install`); never fabricate a result. Any line that starts `error:` is the CLI
+*answering* — a bundle or usage result to read, not a missing toolchain.
 
 ## Invocation
 
@@ -21,6 +20,12 @@ The surface is self-describing — `okf --help` maps every verb, `okf <verb> --h
 its flags. Ask the tool for what exists; this file carries only what `--help`
 cannot: each verb's semantics, its traps, and its JSON shape.
 
+**The verb list is open, not closed.** An installed extension gem adds verbs of
+its own, listed under `installed extensions:` in `okf help`. So a verb that
+`--help` shows and this file does not document is **normal, not a documentation
+error** — ask `okf <verb> --help` for it, and expect nothing here about its
+semantics or JSON. Everything below documents the built-ins only.
+
 **`--json` is compact by design.** Every emitting verb prints single-line JSON —
 the token-efficient substrate you consume; `--pretty` (which implies `--json`)
 indents it for a human. The bytes differ, the JSON is identical, so parse either.
@@ -140,16 +145,74 @@ defect — a terminal leaf (a backlog item, a spec reference) can be loose by de
 The browser page's search brought to the CLI and extended to bodies, so "which
 concept covers X?" costs rows, not body reads. `okf search <dir> <term…>`:
 terms AND together — every term must hit at least one searched field, not
-necessarily the same one — as case-insensitive substrings, or as Ruby regular
-expressions with `--regexp`/`-e` (an invalid pattern is a usage error, exit 2).
+necessarily the same one — matched **literally against raw text** by default, or
+as Ruby regular expressions with `--regexp`/`-e` (an invalid pattern is a usage
+error, exit 2). `--fuzzy` forgives typos; pairing it with `-e` is a usage error,
+since a pattern is matched literally rather than by edit distance.
 `--in a,b` restricts the searched fields (title, id, tags, type, description,
 body); the shared `--type/--area/--tag` filters narrow the candidates *first*,
 so a search scoped by what `index` taught you stays surgical.
 
+**The default is exact, so an exact query means what it looks like.** A phrase in
+one argument (`"dedup key"`), a dotted version (`7.2.0`), an underscored
+identifier (`customer_id`), a mid-word fragment (`ustomer`) and a word written in
+`backticks` all match literally. This is what the scan engine buys, and it is the
+default precisely because those queries are the common ones and the alternative
+loses them silently. <!-- rule:okf-search-exact-identifiers -->
+
+**`--engine index` is the other engine, and the one to reach for when ranking
+matters more than exactness.** The engine is normally chosen by what the query
+needs — `--fuzzy` routes to the index, anything else stays on the default scan —
+and nothing is printed about the choice. `--engine NAME` overrides that for the
+case the flags cannot express: a matching *model* requires no capability, so no
+flag selects one. Under the index, terms match whole tokens and their prefixes
+(`dedup` finds `deduplication`), rows rank by BM25+, and it is the engine the
+browser page runs — so name it when reconciling a CLI answer with the page. The
+cost is real: its tokenizer splits on punctuation, so identifiers shatter
+(`customer_id` → `customer` + `id`), an infix finds nothing, and a backtick is
+never split off at all, so a word inside a code span is unfindable — a large
+silent loss, since technical prose is full of them. **Do not count on ranking to
+rescue it** — BM25 normalizes by field length, so a short concept dense in `7`,
+`2` and `0` can outrank the one that actually says `7.2.0`. Naming an engine that
+cannot do what you also asked (`--engine index -e`) is a usage error naming one
+that can. <!-- rule:okf-search-engine-choice -->
+
+**The capabilities, and which engine has them.** An engine is selected by what
+the query *requires*; only a matching model has to be named, because requiring
+nothing is not something a flag can express:
+
+| Flag | Capability | Engine | What it does |
+|---|---|---|---|
+| *(none)* | — | scan | literal substring over raw text; scores by summed field weight |
+| `-e` / `--regexp` | `regexp` | scan | each term is a Ruby regexp, case-insensitive; invalid → exit 2 |
+| `--fuzzy` | `fuzzy` | **index** | edit distance 0.2 × term length — and switches engine |
+| `--engine index` | — | index | whole-token + prefix matching, BM25+ ranking, browser parity |
+| `--engine scan` | — | scan | the default, spelled out |
+
+Two consequences worth holding. **`--fuzzy` is an engine switch, not a mode**: it
+carries the whole index with it, so a run that wanted one typo forgiven also gets
+shattered identifiers and unfindable code spans — fix the spelling and stay on
+the default when you can. And **`-e` moves nothing** now, because the default
+engine already offers `regexp`; it changes how a term is *read* (pattern rather
+than literal), not where it is matched. <!-- rule:okf-search-fuzzy-is-a-switch -->
+
+`prefix` is a capability the index declares but no flag selects — it is always on
+there. **It is not a reason to reach for the index**: a substring match already
+covers every prefix and then some, so `dedup` finds `deduplication` under both
+engines, while `duplication` and `uplicat` find it under the default only. Prefix
+is what the index needs to catch up to raw text, not a capability it adds on top.
+The index's real advantages over the default are exactly three — relevance
+ranking, typo tolerance, and page parity.
+
 **Search spans bundles.** Leading @refs pick several registered bundles
 (`okf search @handbook @notes auth`); **`@all`** is the ref that means every one.
-The per-bundle rankings merge — scores are absolute term weights, so they
-compare across bundles — and each row carries its bundle's slug. This is the
+Rows from different bundles are ranked together and comparable, and each row
+carries its bundle's slug. Under `--engine index` the bundles go into **one
+corpus** — BM25 prices a term by how rare it is, so separately-ranked lists would
+not compare — which makes a score relative to the whole answer: the same concept
+scores lower searched beside others than searched alone. The default scan needs
+no such trick — its score is absolute, so a row is worth the same either way.
+This is the
 cross-bundle retrieval the in-page search does not have: one question, every
 bundle you keep. <!-- rule:okf-search-all -->
 
@@ -178,12 +241,14 @@ which has no slug to give. Two sharp edges: every *leading* @-arg is taken as a
 the CLI notes both traps on stderr — and any ref, even one, switches the JSON
 envelope (next paragraph).
 
-Rows rank by **where** they hit — title 5, id 4, tags 3, type/description 2,
-body 1, summed over matched fields — and carry one bounded context snippet from
-the strongest match that needs context (description or body). Deliberately not
-fuzzy: the consuming agent is the fuzzy layer — when terms miss, learn the
-bundle's vocabulary from `tags`/`types` and re-ask in its own words, rather
-than hammering synonyms. Advisory read: **exit 0 even with zero matches**.
+Rows rank by where they hit — title 5, id 4, tags 3, type/description 2, body 1 —
+summed as an absolute score by the default scan, and carried as per-field boost
+into **BM25+** under `--engine index`. Each row carries one bounded context
+snippet from the strongest match that needs context (description or body). Every row still names the fields that hit (`matched`), so a result stays
+citable rather than being a bare relevance number. Exact by default: the
+consuming agent is the fuzzy layer — when terms miss, learn the bundle's
+vocabulary from `tags`/`types` and re-ask in its own words, rather than
+hammering synonyms or reaching for `--fuzzy` before you have looked. Advisory read: **exit 0 even with zero matches**.
 JSON, plain-dir mode: `{ bundle, query, count, matches: [{ id, title, type,
 area, tags, matched, score, snippet }] }`. Registry mode — any leading @ref,
 `@all` among them — swaps the envelope: `{ bundles: [{ slug, dir }, …],
@@ -237,10 +302,14 @@ in/out link degree). Add `--json` to any for a machine substrate.
 - **`tags`** — every tag with the concepts that carry it, ordered by count
   descending. The "what themes dominate" view. JSON: `{ bundle, count, tags: [{ tag,
   count, concepts: [id, …] }] }`. `--by type|area` regroups the list per concept
-  dimension with **within-group** counts (a tag spanning groups appears in each) —
-  the substrate for tag curation; the judgment recipe lives in the
-  [maintain playbook](../playbooks/maintain.md). JSON: `{ bundle, count, by,
-  groups: [{ <dim>, count, tags: […] }] }`.
+  dimension with **within-group** counts (a tag spanning groups appears in each);
+  each row also carries the tag's **total** across the narrowed set, printed
+  `count/total` when they differ — so a tag's locality reads per row (a plain
+  count = wholly local; `2/7` = a cross-cutting spread). The substrate for tag
+  curation and for [refine](../playbooks/refine.md)'s domain-vs-concern read;
+  the judgment recipes live in the [maintain playbook](../playbooks/maintain.md)
+  and the [refine playbook](../playbooks/refine.md). JSON: `{ bundle, count, by,
+  groups: [{ <dim>, count, tags: [{ tag, count, total, concepts }] }] }`.
 - **`types`** — every type with the concepts that carry it, ordered by count
   descending. The "what kinds of knowledge" view. JSON: `{ bundle, count, types:
   [{ type, count, concepts: [id, …] }] }`.
@@ -368,3 +437,11 @@ drops each node's body, and `--minimal` ships only `id`/`title` plus the type/ta
 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--ranked-text-retrieval-metadata--body) 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 area* (`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, area, inbound, by_area: { <area>: n } }] }`. Advisory read, exit 0;
+`--minimal`/`--no-body` shape node payloads and change nothing here.

data/lib/okf/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module OKF
-  VERSION = "1.8.0"
+  VERSION = "1.10.0"
 end

data/lib/okf.rb

@@ -40,6 +40,14 @@ module OKF
   require "okf/bundle"
   require "okf/bundle/graph"
   require "okf/bundle/search"
+  # These two lines ARE the engine preference order. Each engine registers itself
+  # at load, `Search.engines` is registration order, and the router walks it after
+  # putting DEFAULT_ENGINE first — so reordering these requires reorders which
+  # engine answers a query two engines could both answer. `loading_test.rb` pins
+  # the result (`[:index, :scan]`) so the coupling cannot drift unnoticed, but the
+  # coupling is here, not there.
+  require "okf/bundle/search/index"
+  require "okf/bundle/search/scan"
   require "okf/bundle/validator"
   require "okf/bundle/validator/result"
   require "okf/bundle/linter"

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: okf
 version: !ruby/object:Gem::Version
-  version: 1.8.0
+  version: 1.10.0
 platform: ruby
 authors:
 - Rodrigo Serradura
@@ -37,6 +37,20 @@ dependencies:
     - - ">="
       - !ruby/object:Gem::Version
         version: '1.4'
+- !ruby/object:Gem::Dependency
+  name: minifts
+  requirement: !ruby/object:Gem::Requirement
+    requirements:
+    - - "~>"
+      - !ruby/object:Gem::Version
+        version: '1.0'
+  type: :runtime
+  prerelease: false
+  version_requirements: !ruby/object:Gem::Requirement
+    requirements:
+    - - "~>"
+      - !ruby/object:Gem::Version
+        version: '1.0'
 description: |
   OKF (Open Knowledge Format) is portable knowledge: Markdown files with YAML
   frontmatter that both humans and agents read from one source. This gem is the
@@ -66,10 +80,28 @@ files:
 - lib/okf/bundle/linter/report.rb
 - lib/okf/bundle/reader.rb
 - lib/okf/bundle/search.rb
+- lib/okf/bundle/search/index.rb
+- lib/okf/bundle/search/scan.rb
 - lib/okf/bundle/validator.rb
 - lib/okf/bundle/validator/result.rb
 - lib/okf/bundle/writer.rb
 - lib/okf/cli.rb
+- lib/okf/cli/catalog.rb
+- lib/okf/cli/command.rb
+- lib/okf/cli/files.rb
+- lib/okf/cli/graph.rb
+- lib/okf/cli/index.rb
+- lib/okf/cli/lint.rb
+- lib/okf/cli/loose.rb
+- lib/okf/cli/registry.rb
+- lib/okf/cli/render.rb
+- lib/okf/cli/search.rb
+- lib/okf/cli/server.rb
+- lib/okf/cli/skill.rb
+- lib/okf/cli/stats.rb
+- lib/okf/cli/tags.rb
+- lib/okf/cli/types.rb
+- lib/okf/cli/validate.rb
 - lib/okf/concept.rb
 - lib/okf/concept/file.rb
 - lib/okf/markdown/citations.rb
@@ -77,10 +109,11 @@ files:
 - lib/okf/markdown/links.rb
 - lib/okf/path.rb
 - lib/okf/registry.rb
+- lib/okf/render/graph.rb
+- lib/okf/render/graph/template.html.erb
 - lib/okf/server/app.rb
-- lib/okf/server/graph.rb
-- lib/okf/server/graph/template.html.erb
 - lib/okf/server/hub.rb
+- lib/okf/server/hub/not_found.rb
 - lib/okf/server/runner.rb
 - lib/okf/skill.rb
 - lib/okf/skill/SKILL.md
@@ -91,6 +124,7 @@ files:
 - lib/okf/skill/playbooks/menu.md
 - lib/okf/skill/playbooks/migrate.md
 - lib/okf/skill/playbooks/produce.md
+- lib/okf/skill/playbooks/refine.md
 - lib/okf/skill/playbooks/search.md
 - lib/okf/skill/reference/APACHE-2.0.txt
 - lib/okf/skill/reference/SPEC.md