okf-wiki/CHANGELOG.md
Kjell Tore Guttormsen 0fd86a23a3
okf-wiki 1.1.1
First public release, published as one snapshot. What changed between
versions is in CHANGELOG.md.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-14 21:03:20 +01:00

8.9 KiB

Changelog

[Unreleased]

[1.1.1] - 2026-11-14

The repair list is written from a reading by a command, the reader can take fewer claims per call, and the repository is ready for its first public release.

Added

  • krepair list: the repair list from a reading's verdicts. It lists every claim read that does not have exactly one verdict, supported, as claims score counts it; given the ids of a sample, it lists only the claims of the sample. Before, no command wrote the list the repair task reads.
  • claims calls --size N: at most N claims per reader call (1 to 10, default 10). A claim read among six to ten neighbours from its page was judged more mildly than the same claim read nearly alone: 12 of 269 accepted claims turned in the sample of 14.11.
  • MIT licence (LICENSE).

Changed

  • The skill assembles a repair into a copy of the collection, from one answers file. Assembled into an empty folder, every page got a new generated time and the diff reported 379 pages without a repair; and one fix builds the whole collection to measure itself (2.5 minutes on 555 pages), so 178 answer files one at a time would take hours.
  • README rewritten for the first public release.

[1.1] - 2026-11-14

A folder of mixed PDF and Office files is built into a collection with the same code as the documentation, both collections are looked up through one read-only server, and the reader measures a changed build on a random sample.

Added

  • okf-wiki serve --bundle NAME WIKI [SOURCES] ...: a read-only MCP server over stdio for lookups across built collections, with four tools (collections, search, page, source). It calls no model and writes nothing. search matches a page's title, description and text; a word in the title counts twice and rarer words weigh more.
  • sources reread ROOT --out ROOT2 reads a folder again as the same collection, and sources watch ROOT --into DIR reports each new, changed or gone document by its original path with the pages built on it, once.
  • kmerge propose: candidate pairs of concepts built twice under two names, found from the page titles; --judged leaves out pairs judged different before. kmerge controls draws the control pairs a build has, so a small folder no longer stops the step.
  • A second calibration set for the reader, from Norwegian and English folder documents. A reading counts only when the reader's model has a green run on both sets.
  • claims sample --of-all: 300 random claims among every claim, with nothing added, for the reading after a changed build.
  • The upstream watch follows the announced new home of the Open Knowledge Format.
  • sources folder: a folder of Markdown and text documents becomes a frozen copy with one command and no hand-written description. Subfolders, capitals and non-ASCII names get a safe name and keep their original path and file:/// URL; Windows line breaks, Windows-1252/Latin-1 and a byte order mark are written as plain UTF-8 and recorded; every file skipped is named with its reason.
  • The repository is a Claude Code plugin: the skill and the helpers reach a session started in any folder, and bin/okf-wiki runs the tools from there.
  • PDF, Word, Excel and PowerPoint in sources folder, read with their exact place: page, paragraph, sheet and cell, slide and notes. Every footnote names the place, the auditor checks it against the frozen copy, and the HTML view opens the original file, a PDF at its page. Hidden Office content is left out and reported. Office files need nothing installed; PDF needs pypdf==6.14.2.
  • python3 -m okfwiki.readers locate, the reader the file-type gate measures: 128 of 128 in its working set.
  • sources folder records each document's language (lang, Norwegian or English, from its common words), and the HTML view states it on every page instead of always en.
  • gates/calibration.passed.json, the record of the reader's green calibration runs (models and sha256 only), and reader collect --gate-passed without a file reads it, so a fresh clone can read without the calibration set.
  • x-aliases: a concept's other names stand on its page, end its index line as (also: A; B), are shown in the HTML view and are searched by the lookup server as its title is. The standard has no field for another name.
  • kmerge apply --operator FILE: verdicts a person set in place of the judge's, listed in the report.
  • The operator's list WORK/homeless-approved.json takes sources beside pages: a document with no readable content, so no page cites it and it has no main page, that he approves is reported by run-after and the guard instead of stopping the build. The auditor still reports it, and it never gets a forced fragment or main page.

Changed

  • Pages judged one concept merge into the page with the shortest name, so the merged page carries the concept's short name; before, the main page won, and 48 of 118 merged pages of collection 1 carried a longer name than a page that went into them. The merged page keeps its own description.
  • A page with a sentence that also stands on the page it merges into is merged all the same: the sentence is cut from the page that moves, so it stands once, and the report lists it. Before, such a page was left out.
  • Any file name is a page; only release-vX.Y.Z.md is a release note by name. .txt sources are built like .md.
  • Page ids spell out letters outside ASCII (Årlig kontroll is aarlig-kontroll, not rlig-kontroll).
  • A footnote on a page built from a folder names the original file (Årlig plan.md, line 3), not its safe name.
  • A spreadsheet row is one line, each cell after its column letter, and a run of rows one block of about 4,000 characters that starts with the run's first row; the footnote names the cell the quote covers.
  • A period after a common abbreviation (f.eks., kl., bl.a., e.g., vs., Dr.) or an initial (Steven L. Rickman) no longer ends a sentence.
  • A folder's document named like a release note (Release-v2.1.0.md) gets the safe name release-v2.1.0-doc.md, so it is built as a document with a main topic, not as a release note.
  • The stop line prices the next pack by the characters of its blocks, not their number: every block records chars, every pack block_chars. A PDF or spreadsheet block of a real folder holds about ten times the characters of a docs block, and the cost per character stays near (6.9 tokens on collection 1, 5.8 and 8.5 on two folder trials), so a pack of PDFs was priced about three times too low. A pack log without characters keeps the rule by blocks.
  • kpack --max-chars C closes a new pack before the source that would take it over C characters of blocks (no cap unless given), and run-before names a source larger than the largest measured to hold in one compiler run (a PDF of 387 767 characters: 342 of 345 fragments accepted, auditor 0 findings).

Fixed

  • A link placed after an abbreviation (U.S. [primary aluminum production](...)) no longer splits its sentence in two, which made the auditor report a cited sentence as uncited and stopped a folder build.
  • kmerge apply leaves out a page whose sentence collides with one on the page kept and lists it, instead of refusing the whole merge.
  • The coverage floor of a document fetched again is scaled by the blocks that can carry a claim, so a page that splits its headings into more blocks is not stopped for citing too little.

Known limitations

  • The reader on 300 random claims of each collection found 17 not supported (5.67 %) in the documentation collection of 3,162 pages and 24 (8.00 %) in the folder collection of 555 pages. The aim is under 3 %; it is not reached.
  • docs/known-limitations.md.

[1.0] - 2026-11-07

First version. It builds one collection, the Claude Code documentation and its release notes, into a wiki in Google's Open Knowledge Format.

Added

  • K, the fragment compiler: code splits sources into blocks, the user's own Claude Code session writes short quoted fragments per pack of five sources through the skill (skills/okf-wiki/SKILL.md), and code assembles the pages.
  • The hard check at writing: every sentence has a footnote to a source that exists and a quote that stands verbatim in it, with a provenance mark (x-provenance: documented, inferred or merged).
  • Links made by the model from meaning (okfwiki.klink), checked and placed by code; a main page for every single-topic source and a home for every other page.
  • The auditor (okfwiki.audit): quotes at their locator, SPEC § 11 conformance, links, orphans and index coverage; the static HTML view (okfwiki.view).
  • The independent reader (agents/okf-wiki-reader.md) with its calibration gate, and claim repair at the fragment (okfwiki.krepair, agents/okf-wiki-repairer.md).
  • The change warning: sources that are new, changed or gone are reported and the pages built on them are marked.

Known limitations

  • docs/known-limitations.md.