okf 1.9.0 → 1.10.0

data/CHANGELOG.md

@@ -1,7 +1,261 @@
 # Changelog
 
+All notable changes to this project are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [1.10.0] - 2026-07-21
+
+### Added
+
+- **The skill gains a `refine` verb** (`playbooks/refine.md`): restructure a
+  bundle to get the most from OKF's capabilities — progressive disclosure, the
+  emergent graph, cross-cutting tags, capture-once-link-many. It is the third
+  authoring boundary: `curate` keeps the structure sound as it stands,
+  `maintain` keeps the content true, `refine` changes where knowledge lives —
+  evidence-first (tag locality, the hub origin test, a fatness alarm),
+  cohesion-over-balance, free levers before file moves, and it *proposes* (a
+  report plus a frozen execution prompt), never auto-applies.
+- **`okf graph --hubs`** — 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`). This is the refine
+  playbook's hub origin test made mechanical: a hub whose inbound majority is
+  foreign to its own area is a move candidate. JSON: `{ bundle, count, hubs:
+  [{ id, area, inbound, by_area }] }`.
+- **The registry has a browser surface.** A meeting with non-technical readers
+  settled what the TUI could not: `okf registry set/del/default/rename` is the
+  right surface for the people who *write* bundles and the wrong one for the
+  people who read them. The graph page's rail grows a **Bundles panel** behind
+  ⚙ — every bundle the server knows about, with its title, `@slug`, folder,
+  concept count and a health verdict — and four routes behind it:
+  `POST /registry/{default,rename,remove,add}`, the only non-GET routes the
+  server has.
+  - **Management is the default, and `--read-only` declines it.** The flag names
+    the restriction rather than the capability, because the audience this was
+    built for should not need a command line to use the page they were pointed
+    at. A loopback bind is writable without a flag; any other address is refused
+    outright, with no flag that opens it — `--bind 0.0.0.0` is how a personal
+    tool becomes a public one, and a write surface does not follow it there.
+  - **Four gates on every write**: is this server writable at all; is there a
+    registry to write to (an ephemeral `okf server ./a ./b` answers `409` rather
+    than leaving the missing controls a mystery); is the verb one of the four (a
+    frozen list — "call whatever method the path names" is how a router becomes
+    an `eval`); and did this come from this page (same-origin *and* a per-boot
+    token, since the token lives in a page another site can get a reader to
+    submit, and Origin alone would trust every tab open on the host). A
+    read-only server hides the controls *and* refuses the request that skipped
+    them: hiding a button is a UI, refusing the request is the boundary.
+  - **A write rebuilds the hub's bundles from disk** before it answers. That is
+    the step easy to skip and impossible to skip safely — a write leaving the
+    running server on the old set is a lie the next click believes.
+  - **`/b/` stops managing and keeps the page.** Both surfaces carried the same
+    four verbs for a while, and two implementations of one contract is the thing
+    that drifts, so the forms came out and the routes stayed. `/b/` answers
+    *which bundles are there* — and remains the empty state a hub with zero
+    bundles still needs — while the panel answers *change this one* where the
+    reader already is. With nothing to post it holds no token either.
+  - **There is no Add on either surface.** A browser cannot hand over a
+    filesystem path — the File System Access API yields an opaque handle, and is
+    Chromium-only besides — so registering stays the agent's act. The route
+    exists for other callers; nothing in the UI reaches it.
+  - **"Workspace" is retired** from the docs and the UI. The things are Bundles
+    and the thing holding them is the registry; a page saying one word while the
+    CLI says another is two products wearing one name.
+- **The hub searches every bundle it hosts.** `GET /search?q=` is the only route
+  in the server that knows about more than one bundle: `Search.across` over one
+  shared index, so BM25 weighs a term against the whole corpus instead of
+  stapling per-bundle lists together. Capped at 50 with the total reported,
+  because a silent cap reads as a complete answer. The engine is named `:index`
+  outright rather than left to route off `fuzzy: true` — that reached the right
+  engine only because nothing else declares the capability, which is correctness
+  by coincidence, and an addon declaring `:fuzzy` would have taken the route
+  silently. A long-lived server also amortizes an index build over every
+  keystroke where a one-shot CLI cannot, and minifts is a port of the browser's
+  own MiniSearch, so a palette hit and an in-page search rank alike.
+  - **The palette's Concepts group comes last**, and not because it matters
+    least: it is the only group that arrives asynchronously, and a group landing
+    above the cursor moves the row under the reader's fingers between the
+    keystroke and the Enter.
+- **The topbar search box says what it filters, and where to go when it finds
+  nothing.** It and ⌘K looked alike and meant different things — the box
+  *filters* what is on screen, the palette *finds* across every bundle a hub
+  hosts — and the box carried neither fact: it emptied the graph in silence and
+  never mentioned the palette, so a reader whose word lived in another bundle
+  got a blank canvas and no way out. Three additions, all inside the box: a
+  **chip** naming the chord (⌘K / Ctrl-K, OS-aware) that opens the palette, a
+  **live count** (`7/8`) that makes an empty result a number which reached zero
+  rather than a view that went blank, and on zero a **panel** naming the bundle
+  and the query — ⏎ hands it to the palette prefilled and already searching, esc
+  clears.
+- **`okf` is extensible.** Any gem that puts `okf/plugin.rb` on its load path can
+  register a verb, and it answers to `okf` — listed in `okf help` under
+  `installed extensions:`, dispatched like a built-in. There is no list of known
+  addons in this gem and no configuration step for the user: installing the gem
+  is the whole installation. It is the same seam `Search.register` opened for
+  search engines, and the same idiom — append-only, idempotent by id, so an addon
+  can never quietly displace a built-in.
+  - **Discovery is lazy**, which is what makes it affordable. A built-in verb
+    resolves against the registry and dispatches without scanning at all; only an
+    unknown verb or `okf help` — which has to know everything by definition —
+    pays the ~11ms `Gem.find_latest_files` costs on the 2.4 floor. A one-shot CLI
+    that will not build a search index for a single query should not pay for
+    discovery to answer a verb it shipped with.
+  - **Extensions must come from gems named `okf-*`**, the convention Jekyll and
+    Vagrant use for the same job: it makes what counts as an okf extension
+    explicit and stops an unrelated gem claiming the `okf/plugin.rb` path by
+    accident. One that is not so named is discovered, skipped, and reported on
+    stderr. A path belonging to *no* gem — a checkout, `ruby -I`, a Gemfile
+    `path:` — stays trusted, because someone put it there deliberately. It is a
+    mild guard too — loading a plugin runs its code — but the naming convention
+    is the reason, not the threat model, which is thin: under Bundler discovery
+    is bundle-scoped anyway, so the Gemfile is already an allowlist.
+  - **The rule holds when it cannot get an answer**, which is a separate promise
+    from the rule itself. A gem name that cannot be read — one corrupt gemspec
+    anywhere on the machine — is refused rather than treated as "belongs to no
+    gem", and the refusal names the exception that caused it. A discovery that
+    fails outright is reported too, since an empty list and no message is
+    indistinguishable from a machine with nothing installed.
+  - **A broken addon is skipped and reported, never fatal** — the same
+    best-effort posture the reader takes with an unparseable file. The note goes
+    to stderr, so a `--json` run's stdout stays a clean machine substrate.
+
+- **The graph page is proven in a real browser.** `test/browser/` drives the
+  page `okf server` and `okf render` share in Chromium, asserting DOM state and
+  computed CSS at real viewport widths, and failing any test where the page
+  threw. Every spec runs twice — once served, once against a `file://` static
+  render — because the two modes diverge on fetched endpoints vs. a baked
+  `EMBED`, and a pass in one proves nothing about the other. It is opt-in
+  (`rake test:browser`, outside the default task) and non-blocking in CI, since
+  the page boots against a CDN and a slow jsdelivr must not gate a merge. The
+  three fixes below are what writing it turned up: shipped defects invisible to
+  a string assertion over the rendered HTML, each reproduced red and pinned
+  green.
+  - **Coverage is mapped per-contract, not guessed.** `test/browser/COVERAGE.md`
+    enumerates every behavioral contract the page introduced across its history
+    and marks each covered / partial / uncovered against a named spec — **176 of
+    181 net-live (97%)**. The five that remain are each a documented blocker, not
+    a missing test: an absence-proof with no line to break, a node-overlap check
+    no cytoscape layout makes both deterministic *and* mutation-sensitive, a
+    map-visibility observable another contract already owns, a palette scroll
+    whose observable is a tautology, and an unbuilt focus-trap. Reaching the
+    branches the flat 8-concept fixture cannot took four further purpose-built
+    bundles beside the hostile one, each served on its own port and baked to its
+    own static page so the main fixture's count assertions stay put — nested
+    directories, forty-five tags, a five-directory-deep reserved path, and a
+    hundred-node ring that drives the graph past its own fit box. Every new spec
+    is mutation-checked: break the code it covers, confirm it goes red for the
+    predicted reason, restore. The map also caught one of its own stale rows — a
+    note listed as an uncovered gap had in fact been deleted from the page, and is
+    now marked superseded rather than owed.
+  - **The page's CDN libraries are served from a local cache** — a read-through
+    cache keyed on the request URL, so a warm run needs no network and a version
+    bump is a miss rather than a stale hit; `OKF_NO_VENDOR_CACHE=1` bypasses it,
+    which is how you check the pins still resolve. It buys robustness, not
+    speed: measured at one worker, 28.7s without and 29.0s with, because the
+    suite is CPU-bound and Chromium already reused those files across contexts.
+
+### Changed
+
+- **`okf tags --by` rows carry each tag's total.** The grouped view printed only
+  within-group counts, so a tag's spread meant cross-referencing groups by
+  hand; each row now shows `count/total` when they differ (`async  2/3`) and
+  the plain count when the tag is wholly local — locality at a glance, the
+  domain-vs-concern read. The JSON rows gain a `total` key; filters recompute
+  it over the narrowed set.
+
+- **A wrong turn at the hub lands on a directory, not an apology.** The 404 is
+  rebuilt on the app shell, and it reads as what it is: the **asked path is the
+  heading**, set in mono where a dropped slash is legible as a shape, with "not
+  found" demoted to the eyebrow above it, since a reader arrives already knowing
+  they are lost. A near-miss slug is a **row** wearing the same anatomy as the
+  list under it, already marked, with ⏎ pointed at it; rows carry the folder
+  that actually distinguishes `site/.okf` from `minifts/.okf`; and colour marks
+  exceptions only, so a healthy row draws no verdict edge at all. Moving through
+  the list is **Tab's** job — every row is an `<a href>`, and a hand-rolled ↑↓
+  cursor was tried and deleted as a second focus model beside the real one. A
+  query matching no bundle is offered the cross-bundle search that would match
+  it, the same escalation the graph page's box makes.
+
+- **On a touch screen a tap opens a card, not the whole viewport.** At ≤768px
+  the inspector is `grid-template-columns:0 1fr`, so tapping a dot measured the
+  stage at 0px wide: the graph was not covered, it was gone. Exploring on a
+  phone became open → read → close → tap the next dot, and you could never see a
+  concept and its neighbourhood at once, which is the one thing a graph is for.
+  A preview card now rises at the bottom edge over a graph that keeps every
+  pixel and stays live — drag it up for the neighbourhood and the body, tap a
+  row and it swaps in place while the camera walks. Folder and index taps fill
+  the card too; they used to write into an invisible panel, so tree and cluster
+  modes were silently dead on touch. The branch is wider than the chrome's
+  (≤768px, or ≤1024px portrait), because a portrait tablet has the same bug and
+  wants the same gesture.
+
+- **Type chips select instead of deselecting.** Three chip groups carried two
+  grammars: areas and tags were additive — nothing selected means everything, a
+  click narrows, a second click undoes — while types were subtractive, every
+  type showing until you clicked one away. Same component, same panel, opposite
+  meaning, and the catalog's and tags view's own type chips were already
+  additive, so the rule a reader learned in one view was wrong in the next.
+  Types now select, and two of them compound into a union the old model could
+  not express at all — it could say "not the other four", never "Services and
+  Charters". The change is a net deletion.
+
+- **The CLI is one file per verb.** `lib/okf/cli.rb` was 1,794 lines and a
+  15-arm `case`; it is now a registry and a dispatcher, with the verbs under
+  `lib/okf/cli/` and the shared surface on a `Command` base class. Behaviour is
+  unchanged — every existing test passes untouched — but `okf help` is now
+  composed from what the commands say about themselves rather than from a
+  heredoc that had to be remembered separately.
+
+### Fixed
+
+- Selecting a node in cluster mode faded the entire graph rather than
+  emphasising the selection. Dimming set opacity on the unrelated elements, but
+  in cluster mode those include the compound area boxes — and a parent's opacity
+  cascades to the nodes inside it, so dimming the boxes faded the very leaves
+  being highlighted. The highlight was real in each node's own opacity and
+  invisible on screen: measured parent-inclusive, the selection sat at 0.1
+  against an unrelated node's 0.01. `focusNode` now dims the leaves and edges
+  and never the `:parent` boxes, which is a no-op outside cluster mode where
+  there are no parents. After: selection and neighbours at 1, the rest at 0.1.
+
+- A log's "Open in graph" button stayed visible, carrying a stale
+  `onclick` from the last map or concept — the "answers about a different file"
+  symptom, returning through CSS. The code hides it correctly
+  (`#fp-graph.hidden = true`), but `.btn.text{display:inline-flex}` outranks
+  `.btn[hidden]{display:none}` at equal specificity (0,2,0), so the later rule
+  won and the button rendered 143px wide with the attribute present. A
+  `.btn.text[hidden]` rule settles it, the same fix the sibling `.fp-head[hidden]`
+  already carried.
+
+- Leaving the graph for another view and returning redrew it at a
+  fraction of its size, and stayed that way. The cause was misdiagnosed as a
+  resize race for months; tracing `cy.animate`'s caller showed the one animation
+  running was a *fit*. `fitGraph` computes zoom from the container's own width,
+  and the one-shot boot fit (`setTimeout(fitGraph, 400)`) fires on whatever view
+  is up by then — so leaving the graph inside that window fits a hidden 0×0
+  canvas, `(w-2*pad)/bb.w` goes negative, and the zoom clamps to `minZoom`.
+  `fitGraph` now returns early on a zero-size canvas and the graph keeps its last
+  good zoom. The hazard was already known at the other end: the boot fit is not
+  registered at all for a `?view=`/`?select=`/`#hash` deep link, whose comment
+  names this same min-zoom clamp — it was the navigate-away case that went
+  uncovered.
+
+- `okf help`'s map advertised `search … [-e|--fuzzy]`, pairing a shorthand
+  it never expanded with an unrelated flag, so the one hint that an engine is
+  selectable read as though `-e` might *be* the engine switch. The row now says
+  `[--regexp|--fuzzy]`, which is what the command's own banner says; `-e` still
+  works and `search --help` still spells it out.
+- **`okf skill <a> <b>` installed into `<a>` and exited 0.** It hand-rolled
+  its own argument handling instead of using the shared pair every `<dir>` verb
+  goes through, so a second destination was silently dropped — the user named two
+  places and the tool wrote one, saying nothing. It is a usage error now (exit 2),
+  refused before anything is written.
+
 ## [1.9.0] - 2026-07-19
 
+### Added
+
 - **`okf search` gains an opt-in full-text index engine.** `--engine index` — and
   `--fuzzy`, which implies it — routes to
   [minifts](https://github.com/serradura/minifts), the pure-Ruby port of the same
@@ -14,6 +268,93 @@
     browser's own setting. Search stays exact unless you ask;
   - **parity with the graph page**, which runs the same MiniSearch build, so the
     two rank identically when the index is named.
+
+- **`--engine NAME` picks the engine outright**, for the case a capability flag
+  cannot express: a matching *model* requires nothing, so no flag selects one.
+  Naming an engine that cannot do what was also asked is a usage error naming one
+  that can (`--engine index -e` → *try --engine scan*), and an unknown name lists
+  what is available. `--help` reads the registry, so an addon's engine appears
+  without the CLI knowing it exists.
+
+- **The graph can draw the index layer, under any layout.** The §6 map was
+  visible only inside file-tree mode, where a folder node stood in for a
+  directory's `index.md`. **Show indexes** makes it a layer: each map is a tile
+  edged to the concepts it lists and the maps below it, dressed by the same
+  selector as file-tree mode's folder node, because the two are the same thing
+  twice over — clicking either opens that directory's `index.md`. Both are accent
+  squares with dashed edges into them, so colour separates *kinds* rather than
+  modes: a directory is not a concept and no longer reads as one. Authorship shows
+  as form — solid where an author wrote a map, hollow and dashed where the bundle
+  only implies one — so the toggle reads as curation as much as navigation.
+  - **Moving between the modes lands in one click.** Tearing the layer down ran
+    its own layout while file-tree mode ran `breadthfirst` a beat later, two
+    layouts racing the same canvas; and because the layer is fetched, a promise
+    resolving after a mode change could land inside file-tree mode. A `relayout`
+    flag settles the first, a per-toggle ticket the second.
+  - **File-tree mode disables the toggle** rather than doubling the folders it
+    already draws.
+  - **One label on every file's graph button.** It read "Explore the knowledge
+    graph" on the root index and "Open core/ in graph" on a nested one, which made
+    a single action look like three. The question is the same whatever is open, so
+    the label is too — and it lives in the markup, where it cannot go stale.
+  - **Opening a map from the reader keeps the reader's graph.** It forced
+    file-tree mode, discarding whatever layout was running, and dimmed the canvas
+    to the map's immediate neighbours. It now switches the *layer* on rather than
+    the *mode* and leaves the layout alone. A reader already in file-tree mode
+    stays there.
+  - **Selecting anything emphasises it the same way.** A concept dimmed the graph
+    to its neighbourhood, a map did nothing at all, and a folder node did nothing
+    either — three meanings for one gesture. One `focusNode` now serves all three.
+  - **Drawn, never modelled.** `index.md` is reserved, so these nodes are built
+    from `/index` straight onto the canvas; `NODES`, `/catalog` and the type and
+    tag indexes never learn they exist. Filters pass them over — a map has no type
+    or tags — but a map whose concepts are all filtered away leaves with them.
+
+- **A first-visit note tells a newcomer the index exists.** The `index.md` an
+  author wrote to be read first was reachable only by finding the Indexes tab and
+  clicking a row, so a reader meeting a bundle for the first time met unlabelled
+  dots with no way in. The page still opens on the graph — it is what makes a
+  bundle legible at a glance, at every width — and a dismissible note at the
+  bottom now says what the picture is, how to touch it, and where the index is.
+  **Read the index** goes straight there; the dismissal is remembered.
+  - **It absorbed the old mobile-only tip** rather than stacking a second banner
+    under it, and it is written for a finger throughout, since a phone is where a
+    first-time reader is least oriented.
+  - **The wording follows the device on two gates, not one.** What a reader does
+    follows `(pointer:coarse)` — a touch tablet in landscape is wider than 768px
+    and still taps; a narrow desktop window is narrower and still clicks. What a
+    reader can reach follows `(max-width:768px)`, because that is when the rail
+    collapses behind `☰`. Short viewports tighten; short *and* wide puts the
+    question beside the button, taking a landscape phone from half the screen to
+    under a third.
+  - **A second note points at `☰`** on compact layouts only, anchored under the
+    button it names rather than at the bottom of the screen. It fires on leaving
+    the graph by any route, so dismissing the first note does not cost it, and
+    opening `☰` answers it — but only once it is on screen, since `☰` is the only
+    way off the graph there and the first tap always comes first.
+  - **Deep links are unaffected**, and `?select=`/`#hash` now switch to the graph
+    before selecting, since the page can be standing elsewhere when they are read.
+
+- The graph page's search box grows a full-text index. One MiniSearch index —
+  lazy-loaded from the CDN on first search, pinned to the `7.2.0` the Ruby
+  MiniSearch port tracks so an `okf search --engine index` result and the
+  browser's rank identically — now backs the graph, catalog and files views: ranked, multi-term
+  (`AND`), prefix (as-you-type) and typo-tolerant, over title, id, type, tags and
+  **description** — plus each concept's **body** wherever the page already holds
+  it (`okf render` bakes every body in, so a static file searches bodies offline;
+  the live server keeps bodies lazy, so its index stays metadata-only until a
+  backend body index arrives). The graph could not be searched by a leaf's
+  description before; now it can. The Files view's **Indexes** tab gets its own
+  full-text index too, over each `index.md`/`log.md`'s body — not just its
+  filename. Until an index loads — or if the CDN is unreachable — each view falls
+  back to its own substring filter, so the box is never dead.
+
+- `Esc` clears the graph selection. A dense graph leaves almost no empty canvas
+  to click for deselecting; `Esc` now drops the highlight (and lets the URL hash
+  forget the node) the same way tapping empty canvas does.
+
+### Changed
+
 - **The default search is unchanged** — literal, case-insensitive substring
   matching over the same fields with the same weights as 1.8.0. The index is
   opt-in rather than default because a one-shot CLI builds an index, asks one
@@ -31,12 +372,7 @@
     field length, so a short concept dense in `7`, `2` and `0` can outrank the one
     that actually says `7.2.0`. The default has none of these, because raw-text
     matching has no tokenizer.
-- **`--engine NAME` picks the engine outright**, for the case a capability flag
-  cannot express: a matching *model* requires nothing, so no flag selects one.
-  Naming an engine that cannot do what was also asked is a usage error naming one
-  that can (`--engine index -e` → *try --engine scan*), and an unknown name lists
-  what is available. `--help` reads the registry, so an addon's engine appears
-  without the CLI knowing it exists.
+
 - **Search engines are adapters.** `OKF::Bundle::Search` became a facade over N
   engines instead of one class with a `regexp ? scan : index` branch. The facade
   keeps everything that defines a result — documents, the row and its key order,
@@ -54,6 +390,7 @@
     match sets by design, so neither can be the oracle. Every registered engine
     runs the same contract, with capability-gated blocks for its own semantics,
     and a registered engine with no conformance class fails the suite.
+
 - **Cross-bundle search ranks one corpus under `--engine index`.** BM25 prices a
   term by how rare it is, so ranking each bundle separately and interleaving the
   lists would produce a ranking that looks sorted and compares nothing; the
@@ -62,51 +399,20 @@
   scores lower searched beside other bundles than alone. The default's scores are
   absolute and need no such treatment.
 
-- **The graph can draw the index layer, under any layout.** The §6 map was
-  visible only inside file-tree mode, where a folder node stood in for a
-  directory's `index.md`. **Show indexes** makes it a layer: each map is a tile
-  edged to the concepts it lists and the maps below it, dressed by the same
-  selector as file-tree mode's folder node, because the two are the same thing
-  twice over — clicking either opens that directory's `index.md`. Both are accent
-  squares with dashed edges into them, so colour separates *kinds* rather than
-  modes: a directory is not a concept and no longer reads as one. Authorship shows
-  as form — solid where an author wrote a map, hollow and dashed where the bundle
-  only implies one — so the toggle reads as curation as much as navigation.
-  - **Moving between the modes lands in one click.** Tearing the layer down ran
-    its own layout while file-tree mode ran `breadthfirst` a beat later, two
-    layouts racing the same canvas; and because the layer is fetched, a promise
-    resolving after a mode change could land inside file-tree mode. A `relayout`
-    flag settles the first, a per-toggle ticket the second.
-  - **File-tree mode disables the toggle** rather than doubling the folders it
-    already draws.
-  - **One label on every file's graph button.** It read "Explore the knowledge
-    graph" on the root index and "Open core/ in graph" on a nested one, which made
-    a single action look like three. The question is the same whatever is open, so
-    the label is too — and it lives in the markup, where it cannot go stale.
-  - **Opening a map from the reader keeps the reader's graph.** It forced
-    file-tree mode, discarding whatever layout was running, and dimmed the canvas
-    to the map's immediate neighbours. It now switches the *layer* on rather than
-    the *mode* and leaves the layout alone. A reader already in file-tree mode
-    stays there.
-  - **Selecting anything emphasises it the same way.** A concept dimmed the graph
-    to its neighbourhood, a map did nothing at all, and a folder node did nothing
-    either — three meanings for one gesture. One `focusNode` now serves all three.
-  - **Drawn, never modelled.** `index.md` is reserved, so these nodes are built
-    from `/index` straight onto the canvas; `NODES`, `/catalog` and the type and
-    tag indexes never learn they exist. Filters pass them over — a map has no type
-    or tags — but a map whose concepts are all filtered away leaves with them.
 - **Collapsing the root folds the file list away** on phones and tablets, where
   the list is stacked on top of the reader and closing the root otherwise left a
   single row above a column of nothing. Reopening the list undoes that collapse,
   so it is one gesture rather than two states to dig out of — the fold remembers
   *why* it happened, and a list folded because a file was opened comes back
   exactly as it was left.
+
 - **The bundle names its own root.** `(root)` and `/` are what a filesystem calls
   it, not what a reader does. The tree's root row, file-tree mode's root node, the
   index layer's root map and the inspector's directory map now all carry the name
   the page header already shows, `--title` included. `areaOf` keeps its own
   `(root)`: that is the area vocabulary `okf stats --by area` and `tags --by area`
   print, not a UI label.
+
 - **The Indexes tab dissolves into the file tree.** The authored layer lived on a
   second tab as a flat list of paths, which put a directory's own map somewhere
   other than the directory. `index.md` and `log.md` are rows now, at the top of
@@ -129,30 +435,7 @@
   - **Fixed on the way:** the reader header rendered empty — an unlabelled badge
     and a graph button pointing nowhere — whenever no file was open, because
     `.fp-head{display:flex}` outranks the UA sheet's `[hidden]{display:none}`.
-- **A first-visit note tells a newcomer the index exists.** The `index.md` an
-  author wrote to be read first was reachable only by finding the Indexes tab and
-  clicking a row, so a reader meeting a bundle for the first time met unlabelled
-  dots with no way in. The page still opens on the graph — it is what makes a
-  bundle legible at a glance, at every width — and a dismissible note at the
-  bottom now says what the picture is, how to touch it, and where the index is.
-  **Read the index** goes straight there; the dismissal is remembered.
-  - **It absorbed the old mobile-only tip** rather than stacking a second banner
-    under it, and it is written for a finger throughout, since a phone is where a
-    first-time reader is least oriented.
-  - **The wording follows the device on two gates, not one.** What a reader does
-    follows `(pointer:coarse)` — a touch tablet in landscape is wider than 768px
-    and still taps; a narrow desktop window is narrower and still clicks. What a
-    reader can reach follows `(max-width:768px)`, because that is when the rail
-    collapses behind `☰`. Short viewports tighten; short *and* wide puts the
-    question beside the button, taking a landscape phone from half the screen to
-    under a third.
-  - **A second note points at `☰`** on compact layouts only, anchored under the
-    button it names rather than at the bottom of the screen. It fires on leaving
-    the graph by any route, so dismissing the first note does not cost it, and
-    opening `☰` answers it — but only once it is on screen, since `☰` is the only
-    way off the graph there and the first tap always comes first.
-  - **Deep links are unaffected**, and `?select=`/`#hash` now switch to the graph
-    before selecting, since the page can be standing elsewhere when they are read.
+
 - **The file tree nests.** Directories were a sorted list of full paths, which
   made `core` and `core/configurations` read as two unrelated folders and left
   the shape of a bundle invisible. Each row is now one path segment indented by
@@ -164,39 +447,6 @@
     folders standing instead of a single `(root)` row. Unfolding clears the whole
     set, root included, so a root closed by hand is still reversible from there.
 
-- The graph page's search box grows a full-text index. One MiniSearch index —
-  lazy-loaded from the CDN on first search, pinned to the `7.2.0` the Ruby
-  MiniSearch port tracks so an `okf search --engine index` result and the
-  browser's rank identically — now backs the graph, catalog and files views: ranked, multi-term
-  (`AND`), prefix (as-you-type) and typo-tolerant, over title, id, type, tags and
-  **description** — plus each concept's **body** wherever the page already holds
-  it (`okf render` bakes every body in, so a static file searches bodies offline;
-  the live server keeps bodies lazy, so its index stays metadata-only until a
-  backend body index arrives). The graph could not be searched by a leaf's
-  description before; now it can. The Files view's **Indexes** tab gets its own
-  full-text index too, over each `index.md`/`log.md`'s body — not just its
-  filename. Until an index loads — or if the CDN is unreachable — each view falls
-  back to its own substring filter, so the box is never dead.
-- The Files tree's folder collapse works during a search. An active search or
-  type/tag filter used to force every folder open, so fold clicks did nothing;
-  folders now honor their collapsed state always (a collapsed group still shows
-  its header, so a match is never hidden). A **fold/unfold-all** control in the
-  Files tab header collapses or expands every visible group at once.
-- Clustering no longer leaves phantom empty boxes. When a filter or a search hid
-  every concept in an area, the cluster's labelled box lingered as an empty
-  rectangle; the box now hides when no child survives and returns when one does —
-  the same rule the fit already used to leave stale boxes out of view, now
-  applied to what is drawn.
-- `Esc` clears the graph selection. A dense graph leaves almost no empty canvas
-  to click for deselecting; `Esc` now drops the highlight (and lets the URL hash
-  forget the node) the same way tapping empty canvas does.
-- A title-less concept now wears one name in every view. `catalog` and the §6
-  index listing fell back a concept with no `title` to its full id — `area/thing`
-  — while the graph node fell back blank-aware to the basename — `thing` — so the
-  same concept answered to two labels across two views of one bundle, and a
-  `title: ""` slipped past the nil-only `||` to catalog as an empty string. Both
-  now fall back the graph's way (`File.basename`, blank-aware), so the label is
-  the same wherever the concept appears.
 - `okf render` stops baking a redundant description map. The static page derived
   its `/node/meta` fragments from a separate `meta` payload that held nothing but
   each concept's description, HTML-escaped — data the embedded `catalog` already
@@ -204,6 +454,7 @@
   same escape the server applies at `/node/meta`), so the `meta` key leaves the
   baked payload and the description lives in one place. Both XSS guards are
   unchanged; `okf server` is untouched.
+
 - The bare not-a-directory error now teaches the registry grammar. A verb given
   a target that is neither a directory nor an `@ref` moved from
   `error: <arg> is not a directory` to
@@ -211,6 +462,7 @@
   so a consumer who typed a query or a bad path meets `@slug` addressing at the
   error instead of hunting for it. (`@all` stays out of the message — it is
   `search`'s alone, and the error seam is shared by every verb.)
+
 - The bundled skill teaches `@slug` as a first-class target and stops probing for
   the CLI. `SKILL.md`'s "Which directory?" is now "Which target?" — a leading `@`
   is a registry ref routed straight to `okf <verb> @slug`, with the fallback
@@ -219,8 +471,32 @@
   probe is gone: run the verb, and treat a shell `command not found` as the only
   signal to install, so the common case pays no guard round.
 
+### Fixed
+
+- The Files tree's folder collapse works during a search. An active search or
+  type/tag filter used to force every folder open, so fold clicks did nothing;
+  folders now honor their collapsed state always (a collapsed group still shows
+  its header, so a match is never hidden). A **fold/unfold-all** control in the
+  Files tab header collapses or expands every visible group at once.
+
+- Clustering no longer leaves phantom empty boxes. When a filter or a search hid
+  every concept in an area, the cluster's labelled box lingered as an empty
+  rectangle; the box now hides when no child survives and returns when one does —
+  the same rule the fit already used to leave stale boxes out of view, now
+  applied to what is drawn.
+
+- A title-less concept now wears one name in every view. `catalog` and the §6
+  index listing fell back a concept with no `title` to its full id — `area/thing`
+  — while the graph node fell back blank-aware to the basename — `thing` — so the
+  same concept answered to two labels across two views of one bundle, and a
+  `title: ""` slipped past the nil-only `||` to catalog as an empty string. Both
+  now fall back the graph's way (`File.basename`, blank-aware), so the label is
+  the same wherever the concept appears.
+
 ## [1.8.0] - 2026-07-17
 
+### Added
+
 - A persistent bundle registry and a multi-bundle hub. `okf registry`
   (list / set / del / default / rename) keeps a per-user list in a plain JSON
   file at `$OKF_HOME/registry.json` (default `~/.okf`), and `okf server` reads
@@ -230,6 +506,7 @@
   switcher (⌘/Ctrl-K, or the rail button), `/b/` is a browsable index, and an
   unknown slug 404s as a page with a way home. The hub reads its bundles at
   boot — restart after registry changes.
+
 - The registry is **ordered, and the first entry still on disk is the default** —
   the bundle a bare `okf server` opens at `/`. `okf registry default @slug` moves
   that entry to the front, and `okf registry set --default` registers straight to
@@ -239,11 +516,13 @@
   vanished directory is stepped over rather than starred — `registry list`'s `*`
   always names the bundle `/` opens — and `registry default @slug` refuses one,
   just as `registry set` refuses to register a directory that is not there.
+
 - `$OKF_HOME` is the single lever on which registry a command reads: set it and
   every verb follows, from `okf registry list` to an `@slug` on `okf lint`. It
   names exactly one registry, with no fallback to `~/.okf` behind it, and an
   empty value counts as unset rather than planting `registry.json` in the
   current directory.
+
 - `@slug`: wherever a command takes a `<dir>`, `@slug` names a registered
   bundle and bare `@` the registry default — `okf lint @handbook`,
   `okf render @ -o graph.html`. A slug is normalized like registration was
@@ -254,6 +533,34 @@
   (`okf server @a @b`) mounts each bundle under its registered slug, the first
   at `/`, and a registered slug reserves its mount ahead of any plain
   directory that shares the name.
+
+- `okf search` spans bundles: several leading `@slug`s, or `@all` for every
+  registered one. Rankings merge across bundles with every row labeled by its
+  bundle's slug (a `bundles` list and a per-match `slug` key in the JSON).
+  Asking for everything tolerates gaps — `@all` skips a bundle whose directory
+  has vanished, with a note — while naming one insists on it, and `@all @docs`
+  simply dedupes. `all` is reserved *in the registry*, on all three ways in: a
+  directory named `all/` registers as `all-2`, `--as all` is refused, and a row
+  already claiming the name in the registry file — hand-typed, or written before
+  the name was reserved — is read as `all-2`, so the reservation never strands a
+  registry it inherited. An ephemeral `okf server ./all` still mounts at
+  `/b/all/` — no registry, no refs, nothing to reserve.
+
+- The inspector's type and tags are filter handles: clicking one focuses the
+  graph on that facet — the same jump the stats bars make — and clicking it again
+  clears it. The chip lights while its facet is the only filter in play, which is
+  exactly when a second click is an undo, so what you see and what the next click
+  does are the same question. With another filter set it re-focuses instead,
+  rather than throwing away more than the click put there.
+
+- The graph page answers `?` with a sheet of every keyboard shortcut, reachable
+  from a rail button too — a shortcut list you can only open with a shortcut helps
+  whoever needs it least. `/` focuses the current view's search where it has one,
+  skipping the view that only reads; the sheet is written against the key handler
+  it documents, since a shortcut list that has drifted is worse than none.
+
+### Changed
+
 - `@slug` is spelled where it is used, not just where it is explained, and it is
   the one token — `okf help`'s map (`lint <dir|@slug>`) and each command's own
   banner show it the same way, with a note under the map defining it: the slug
@@ -263,25 +570,17 @@
   advertising a bare `<dir>`. The registry-editing verbs — `del`, `default`,
   `rename` — take the slug bare (no `@`) or as an `@slug`; the read verbs need
   the `@`, since a bare word there is a path.
+
 - `okf <command> -h` prints that command's own banner and flags. Help now answers
   on stdout with an exit code like every other command: it was OptionParser's
   officious handler, which printed past the caller's injected streams and ended
   the process with `exit` instead of returning a status.
-- `okf search` spans bundles: several leading `@slug`s, or `@all` for every
-  registered one. Rankings merge across bundles with every row labeled by its
-  bundle's slug (a `bundles` list and a per-match `slug` key in the JSON).
-  Asking for everything tolerates gaps — `@all` skips a bundle whose directory
-  has vanished, with a note — while naming one insists on it, and `@all @docs`
-  simply dedupes. `all` is reserved *in the registry*, on all three ways in: a
-  directory named `all/` registers as `all-2`, `--as all` is refused, and a row
-  already claiming the name in the registry file — hand-typed, or written before
-  the name was reserved — is read as `all-2`, so the reservation never strands a
-  registry it inherited. An ephemeral `okf server ./all` still mounts at
-  `/b/all/` — no registry, no refs, nothing to reserve.
+
 - The registry validates its file's shape, not just its JSON syntax: a
   hand-edited entry missing `path` is a usage error naming the file instead of
   a `TypeError`, and `okf registry --json set <dir>` — a subcommand behind a
   flag — is a usage error rather than silently listing and exiting 0.
+
 - The graph page's ⌘/Ctrl-K palette opens in **every** mode and reaches a view as
   well as a bundle. It was wired only behind a hub, so a standalone
   `okf server ./docs` and every `okf render` page — the two modes most people meet
@@ -291,12 +590,7 @@
   carries the rail's own icon and label, read from the rail so the two cannot
   drift. Where there is no hub there is no bundle to switch to, and views become
   the whole list.
-- The inspector's type and tags are filter handles: clicking one focuses the
-  graph on that facet — the same jump the stats bars make — and clicking it again
-  clears it. The chip lights while its facet is the only filter in play, which is
-  exactly when a second click is an undo, so what you see and what the next click
-  does are the same question. With another filter set it re-focuses instead,
-  rather than throwing away more than the click put there.
+
 - The inspector's *Links to* / *Linked from* rows read as concepts rather than a
   wall of accent-coloured text. Each carries its type's dot — the colour that node
   already wears in the graph beside it — with the type named and the section
@@ -304,15 +598,14 @@
   title is read. The rows share one panel with hairline dividers and a hover fill:
   the container carries the click affordance, which leaves colour free to mean type
   and nothing else.
+
 - The inspector's widen chevron splits the screen instead of taking 70% of it. The
   panel drag-resizes, so the chevron is a preset rather than a maximum, and burying
   the graph to read one concept was the wrong thing to default to.
-- The graph page answers `?` with a sheet of every keyboard shortcut, reachable
-  from a rail button too — a shortcut list you can only open with a shortcut helps
-  whoever needs it least. `/` focuses the current view's search where it has one,
-  skipping the view that only reads; the sheet is written against the key handler
-  it documents, since a shortcut list that has drifted is worse than none.
-- Fixed: a file the reader could not **open** (permissions) threw its errno out
+
+### Fixed
+
+- A file the reader could not **open** (permissions) threw its errno out
   of the read, so a single locked file took the whole bundle down through every
   verb that reads one — `lint`, `validate`, `catalog`, `server`, `registry set`
   — as a backtrace, under an exit code claiming the bundle was non-conformant.
@@ -322,12 +615,13 @@
   The stderr note reads `skipped N unusable file(s)` — it counts two kinds now,
   so it names neither and points at `validate`, which names both.
 
-- Fixed: `okf registry del <path>` could delete the wrong bundle. A path that
+- `okf registry del <path>` could delete the wrong bundle. A path that
   matched no registered directory fell through to a normalized *slug* lookup, so
   `del ./notes` — naming a local directory — removed whichever entry happened to
   be slugged `notes`, wherever it pointed, and reported `removed notes` with exit
   0. An argument with a `/` in it now names a location and only a location.
-- Fixed: the registry read trusted stored slugs verbatim while both write paths
+
+- The registry read trusted stored slugs verbatim while both write paths
   normalized, so a hand-typed `"slug": "My Docs"` listed fine but could not be
   named by `@my-docs`, `registry rename`, or `registry default` — the verbs that
   could repair it were the ones that could not see it. The read normalizes now,
@@ -335,31 +629,37 @@
   quote could reach a slug, and with it a DOM XSS in the server's bundle
   switcher, whose JS escape covered `& < >` but not quotes; the escape now covers
   quotes too, so the page does not depend on a guarantee three layers away.
-- Fixed: `okf lint` bucketed a whitespace-only `type` under its own literal
+
+- `okf lint` bucketed a whitespace-only `type` under its own literal
   heading while `types`/`graph`/`stats` bucketed it as `Untyped`, so two verbs
   reported type inventories for the same bundle that would not reconcile. §9.2
   makes a blank type as non-conformant as a missing one; both sides say so now.
-- Fixed: `okf search <dir> --fields slug` passed the field guard and returned one
+
+- `okf search <dir> --fields slug` passed the field guard and returned one
   empty object per match with exit 0. Only registry mode labels rows with a slug,
   so the two modes now declare the shape each actually emits and a path-named
   search names the fields it does have.
-- Fixed: `okf registry set` reported `(0 concepts)` for a bundle whose files it
+
+- `okf registry set` reported `(0 concepts)` for a bundle whose files it
   could not read; it notes the skipped files like every other reading verb.
   `okf registry rename DOCS handbook` echoed the argv rather than the slug it
   renamed, naming a bundle that never existed. An unwritable `$OKF_HOME` raised
   a bare `Errno::EACCES` at exit 1 instead of a usage error at exit 2.
-- Fixed: a long path in the Files view's Indexes tab pushed its `map`/`log` badge
+
+- A long path in the Files view's Indexes tab pushed its `map`/`log` badge
   past the right edge of the list. The badge was already pinned and unshrinkable;
   the filename beside it was the problem — a bare text node, and an anonymous flex
   item's automatic minimum size is the full width of its text, so it refused to
   give way and drove the badge out instead. It truncates now, with the full path on
   hover. Only a wide enough path in a narrow enough pane reached it, which is why
   it never showed on mobile, where the list runs full width.
-- Fixed: the bundle switcher scrolled its own first row out of view as it opened.
+
+- The bundle switcher scrolled its own first row out of view as it opened.
   It rendered, and scrolled the active row into view, while the dialog was still
   hidden — and a list that is not being displayed measures zero, so the scroll
   landed arbitrarily.
-- Fixed: the `3` (Files) shortcut did nothing once the Indexes tab was open, and
+
+- The `3` (Files) shortcut did nothing once the Indexes tab was open, and
   the palette's Index row would have blanked the page. "Index" is not a view —
   there is no `#view-index`, only the Files view showing its Indexes tab — so
   `setView('files')` from that tab early-returned, and `setView('index')` named a
@@ -370,11 +670,14 @@
 
 ## [1.7.0] - 2026-07-16
 
+### Added
+
 - `okf server`: responses are gzipped when the client accepts it
   (`Rack::Deflater` at the boot seam). Lossless and transparent — the browser
   decompresses automatically — and no new dependency, since `Rack::Deflater`
   ships inside rack. Clients that send no `Accept-Encoding` keep getting
   identity responses. `okf render`'s static HTML is untouched.
+
 - The agent skill gains a `migrate` verb (`playbooks/migrate.md`): convert
   existing documentation into a conformant bundle **in place** — frontmatter
   and reserved files added, bodies kept verbatim (`produce` keeps
@@ -382,11 +685,15 @@
   inference, the menu playbook now leads with it when a target already holds
   markdown docs, and pointing any verb at a directory that is not a bundle now
   suggests `migrate` instead of grinding through the validate errors.
+
+### Changed
+
 - The graph page's link-preview image points at the renamed
   `okfgem.com/og-demo-v3.png`. The site's OG art was refreshed to drop "Live
   Graph" from the package formula (it is `Agent Skill + CLI/Lib + Graph` now
   that `okf render` makes the graph live *or* static), and the filename carries
   the version so social scrapers pick the new art up.
+
 - The plugin's `/okf:gem` command is now a pass-through shim: it hands its
   arguments to the okf skill unchanged, making `SKILL.md` the single router
   for every channel. The routing prose the command used to duplicate had no
@@ -396,6 +703,8 @@
 
 ## [1.6.0] - 2026-07-15
 
+### Added
+
 - New CLI verb: `okf render <dir> [-o FILE]` — the live graph as one static,
   self-contained HTML file, so it hosts where a server can't (GitHub Pages, an
   object store, an attachment). It is the same page `okf server` serves, one
@@ -408,6 +717,7 @@
   through `DOMPurify.sanitize(marked.parse(...))`, so the trust boundary holds;
   the trade-off is weight — each body is inlined, so a big bundle makes a big
   file, and `okf server` stays the choice at scale.
+
 - Official Docker image: `ghcr.io/serradura/okf`, a portable CLI that runs every
   `okf` command (the graph server included) with no Ruby on the host. It is built
   from source and published multi-arch (`linux/amd64`, `linux/arm64`) to the
@@ -417,6 +727,8 @@
 
 ## [1.5.0] - 2026-07-13
 
+### Added
+
 - New CLI verb: `okf search <dir> <term…>` — deterministic ranked retrieval
   over concept metadata *and bodies*, the browser page's search brought to the
   CLI. Terms AND together as case-insensitive substrings, or as Ruby regexps
@@ -426,14 +738,17 @@
   and carry a bounded context snippet, so "which concept covers X?" costs a
   few rows instead of a body read. Advisory read: exit 0 even with no matches.
   Deliberately not fuzzy — the consuming agent is the fuzzy layer.
+
 - The skill learns retrieval as a first-class verb: a new `search` playbook
   (progressive disclosure end to end: ingest `okf index`, decide where to
   look, cut across with `okf search`, read only the winning bodies),
   search-aware routing in SKILL.md and the menu/consume playbooks, and
   `/okf:gem search <query>` first in the Claude Code plugin's routing.
+
 - Retrieval eval in the suite: the progressive path (index skeleton → search →
   one body) must answer a planted question in under 25% of the bytes of the
   full graph dump, so the playbook's economics stay true by construction.
+
 - Graph server: the authored layer joins the UI. The Files view carries two
   tabs — **Files** (the per-directory concept groups, foldable) and
   **Indexes** (the log first, as the chronological index, then every
@@ -447,19 +762,27 @@
   is fetched fresh on every read, so a just-appended entry shows without a
   restart. A reserved file's "Open in graph" jumps to its folder in the file
   tree, map in the inspector. New `/index` and `/log` endpoints back it all.
+
 - Graph server: Mermaid diagrams in concept bodies are click-to-inspect. A
   click (or tap) opens the diagram full screen — drag to pan, wheel or pinch
   to zoom, buttons and double-click reset, Esc closes — powered by
   [Panzoom](https://github.com/timmywil/panzoom), lazy-loaded from the CDN
   exactly like Mermaid itself.
+
+### Changed
+
 - The Claude Code plugin's `/okf:gem` command now weighs the shape of a
   free-form ask: a question about what the bundle knows routes through the
   search playbook and answers from retrieved concepts instead of guessing.
+
 - Skill efficiency audit: every playbook now takes the CLI's lean paths.
   `maintain` hunts affected concepts with `okf search` and pulls edges via
   `graph --json --minimal` instead of the full-body dump, `menu` reads the
   plain-text reports it only scans, and SKILL.md pins the discipline as a
   rule: skeleton first, bodies last.
+
+### Fixed
+
 - Docs: the CLI reference's server section now reflects the DOMPurify
   sanitization that landed in 1.1.0 (it still said bodies render unsanitized),
   and the server page's link-preview image points at the renamed
@@ -467,6 +790,8 @@
 
 ## [1.4.0] - 2026-07-12
 
+### Changed
+
 - Graph server UX round. Selecting a node now makes one camera move instead of
   two (the pan used to race the opening panel and the debounced canvas resize,
   a dizzying double movement; rapid clicks also queued animations — both fixed).
@@ -485,14 +810,19 @@
 
 ## [1.3.0] - 2026-07-12
 
+### Added
+
 - The graph server page now emits link-preview metadata: Open Graph and Twitter
   Card tags with a social image, plus `theme-color` and `color-scheme`, so a
   shared `okf server` URL unfurls as a proper card in chat and social apps.
+
 - Docs: a themed README hero (light and dark), a GitHub social preview image,
   and Website / Live demo / Claude Code plugin links.
 
 ## [1.2.0] - 2026-07-12
 
+### Added
+
 - Claude Code plugin. The repository now doubles as a plugin marketplace:
   `/plugin marketplace add serradura/okf-gem`, then `/plugin install okf@okfgem`.
   The plugin carries the canonical skill (a generated copy; `rake plugin:sync`
@@ -516,12 +846,8 @@
 
 ## [1.1.0] - 2026-07-12
 
-- The graph server now sanitizes every concept body before rendering it. The
-  page runs marked's HTML output through [DOMPurify](https://github.com/cure53/DOMPurify)
-  (loaded from the same CDN as Cytoscape and marked) on the way to the DOM, so a
-  bundle carrying active content in a Markdown body can no longer script the
-  viewer. Inlined graph data was already escaped through `json_for_script`; this
-  closes the other path.
+### Changed
+
 - `require "okf"` now loads the pure library only. The two argv-facing shells —
   `OKF::CLI` and the `OKF::Skill` installer — load on demand, from `exe/okf` or
   an explicit `require "okf/cli"` / `require "okf/skill"`. `optparse` moves with
@@ -529,28 +855,45 @@
   in-memory model and on-disk handles no longer pulls in the command-line
   machinery. The CLI itself is unchanged.
 
+### Security
+
+- The graph server now sanitizes every concept body before rendering it. The
+  page runs marked's HTML output through [DOMPurify](https://github.com/cure53/DOMPurify)
+  (loaded from the same CDN as Cytoscape and marked) on the way to the DOM, so a
+  bundle carrying active content in a Markdown body can no longer script the
+  viewer. Inlined graph data was already escaped through `json_for_script`; this
+  closes the other path.
+
 ## [1.0.0] - 2026-07-12
 
 Initial release.
 
+### Added
+
 - `OKF::Concept` / `OKF::Bundle`: pure in-memory model of an OKF v0.1 bundle,
   buildable straight from data (no disk) with link, citation, and markdown
   round-trip primitives.
+
 - `OKF::Bundle::Validator`: the spec §9 conformance gate (hard errors) with the
   spec's soft guidance reported as warnings — broken cross-links are tolerated,
   as §5.3 requires.
+
 - `OKF::Bundle::Linter`: advisory curation-quality report across reachability,
   backlog, completeness, freshness, provenance, and hygiene, with `--json` as a
   machine substrate.
+
 - `OKF::Bundle::Graph`: the knowledge graph (nodes, edges, type/tag indexes) at
   selectable fidelity.
+
 - On-disk handles: `OKF::Bundle::Folder`, `OKF::Bundle::Reader`,
   `OKF::Bundle::Writer` (atomic, validate-before-publish), and
   `OKF::Concept::File`.
+
 - `OKF::Server::App`: the interactive graph as a mountable Rack app — five views
   (graph, catalog, files, tags, stats) with type/area/tag filtering throughout,
   bodies fetched live from disk — served by a built-in WEBrick runner
   (`okf server`).
+
 - `okf` CLI: `validate`, `lint`, `loose`, and `graph`, plus the read views as
   text — `index`, `catalog`, `files`, `tags`, `types`, `stats` — at full parity
   with the browser: every list view narrows with `--type`/`--area`/`--tag`
@@ -558,6 +901,7 @@ Initial release.
   `tags --by type|area` regroups the tag index per concept dimension with
   within-group counts — the tag-curation view. `server` boots the graph page;
   `skill` installs the companion skill.
+
 - `okf index`: a read view over the progressive-disclosure layer (spec §6) — one
   entry per directory that holds concepts or carries an `index.md`, root first,
   with its authored index body (frontmatter stripped), a type/tag rollup over the
@@ -565,19 +909,35 @@ Initial release.
   directory with concepts but no `index.md` has its listing synthesized (§6 permits
   it) and is flagged. `--area` (repeatable), `--no-body`, and `--json`; advisory,
   always exit 0. Backed by the pure `OKF::Bundle#directory_index`.
+
 - JSON output is **compact by default** across every emitting verb (the
   token-efficient machine substrate, matching the server); `--pretty` indents it
   for reading and implies `--json`. JSON semantics are identical either way — only
   whitespace differs — so any parser is unaffected.
+
 - JSON property projection on the list views: `index`, `catalog`, and `files`
   take `--fields a,b` (emit only these properties) or `--except a,b` (emit all but
   these), so an agent never pays tokens for fields it will not read. The flags are
   mutually exclusive, imply `--json`, match property names case-insensitively, and
   reject an unknown name (exit 2) listing the valid ones; `okf index --no-body` is
   shorthand for dropping the `body` field.
+
 - Bundled companion agent skill (`okf skill <dest>`): SKILL.md carrying the
   judgment (the CLI surface stays self-describing via `--help`) — including the
   orient-before-you-read protocol and the CLI/judgment boundary — the OKF v0.1
   spec, authoring and CLI references (tag-vocabulary curation, the SPEC-section
   map, the closeout gate), and concept/index/log templates.
+
 - Runs on Ruby >= 2.4 with two runtime dependencies: rack and webrick.
+
+[1.10.0]: https://github.com/serradura/okf-gem/compare/v1.9.0...v1.10.0
+[1.9.0]: https://github.com/serradura/okf-gem/compare/v1.8.0...v1.9.0
+[1.8.0]: https://github.com/serradura/okf-gem/compare/v1.7.0...v1.8.0
+[1.7.0]: https://github.com/serradura/okf-gem/compare/v1.6.0...v1.7.0
+[1.6.0]: https://github.com/serradura/okf-gem/compare/v1.5.0...v1.6.0
+[1.5.0]: https://github.com/serradura/okf-gem/compare/v1.4.0...v1.5.0
+[1.4.0]: https://github.com/serradura/okf-gem/compare/v1.3.0...v1.4.0
+[1.3.0]: https://github.com/serradura/okf-gem/compare/v1.2.0...v1.3.0
+[1.2.0]: https://github.com/serradura/okf-gem/compare/v1.1.0...v1.2.0
+[1.1.0]: https://github.com/serradura/okf-gem/compare/v1.0.0...v1.1.0
+[1.0.0]: https://github.com/serradura/okf-gem/releases/tag/v1.0.0