adf-codec/todo-history.md

Todo history

The done todo.md items in full, as they were written. todo.md keeps a one-line summary of each.

Milestones

• 0 — Scaffold. package.json per §6, tsconfig.json, .npmrc (save-exact=true), the Docker tooling, renovate.json (§9), and .gitea/workflows/ci.yml gating branches: runs-on: docker-host, actions pinned to semver tags.

• 1a — The directive grammar (spec/flavour.md): inline/block/leaf directive forms, attributes, escaping, nesting, canonical form, the opaque-carry spelling, the raw-HTML input policy.

• 1b — Block node syntaxes in spec/flavour.md: panel, expand/nestedExpand, the media family, the pipe-vs-directive table rule and the directive table form, task and decision lists, layout, extensions, syncBlock.

• 1c — Inline node syntaxes and marks in spec/flavour.md: mention, emoji, status, date, inlineCard, mediaInline; underline, subsup, textColor, border; the spelling for text nodes whose whitespace CommonMark cannot hold (literal newlines, leading or trailing spaces) — escape-based, never literal, since pipe cells trim and pad. At mediaInline, check real payloads for external-URL support — if it exists, revisit the media section's mid-text-image error and its "no slot" ground.

• 1d — Corpus start (§10): checked-in fixtures per spec'd node, in corpus/, one directory per contract kind (corpus/README.md). Settled (the maintainer, 2026-08-26): the nodes CommonMark spells get directive sections of their own, rather than riding the opaque carry. Per @atlaskit/adf-schema 57.1.0 every block node it spells — blockquote, bulletList, codeBlock, heading, listItem, orderedList, paragraph, rule — carries a localId with no spelling, codeBlock also hideLineNumbers, uniqueId and wrap, blockquote also marks, and hardBreak text and localId; 2f gives each a place, and the plain spelling stays wherever the attributes are absent. That hands the two collision sites in corpus/unspellable/ the second spelling they lacked, so each takes the directive form as a media with an empty alt already does (spec/flavour.md, the CommonMark image): a codeBlock whose info string is empty, and an orderedList starting at 1. Also blocked: the link rule covers destination spaces only, so two shapes have no spelling and are refused meanwhile — href https://example.com/a)b and title He said "hi", both in corpus/unspellable/. Two defensible spellings each — angle brackets or a backslash escape, and for titles '…' or (…) besides — so §8 leaves the pick here. Also blocked: block separation is unstated for a CommonMark block beside a directive block in a container body — an expand whose content is paragraph "A" then a panel (panelType warning) holding "B" spells A and :::panel warning either on consecutive lines or with a blank line between. Two defensible spellings, so §8 leaves the pick here; the answer governs every unknown node type too, the block carry counting as a CommonMark block since its spelling is a fenced code block. unspelled-block-separation refuses the pair meanwhile, an empty paragraph's ::paragraph beside a CommonMark block included — and, since a mediaSingle's spelling now follows whether CommonMark can spell its URL, two sibling images differing only by an & land in the same refusal.

  • 1d1 — The CommonMark subset: blockquote, bulletList, codeBlock, heading, orderedList, paragraph, rule, listItem, hardBreak, text, code spans, and the code, em, link, strike and strong marks — one mark per text node; nesting is 1d3's.
  • 1d2 — Block nodes: panel, expand/nestedExpand, the media family and the CommonMark image shape, both table forms, task and decision lists, layout, extensions, syncBlock — with the reserved marks attribute and the fence lengths nesting forces.
  • 1d3 — Inline nodes and marks: date, emoji, inlineCard, mediaInline, mention, status; border, subsup, textColor, underline; the content slot's text attribute and the :text{text="…"} whitespace spelling.

• 2 — adfToMarkdown. First real code. Each sub-item turns one corpus directory green; the two that have no fixtures yet write them in the same chunk, tests first (§10).

  • 2a — The runner and the CommonMark subset. The corpus runner: walk corpus/round-trip/, assert adfToMarkdown emits each .md byte for byte. Decide here where §10's coverage check lives, and gate that every corpus/**/*.json re-serializes to itself under the library's own canonical serializer — one implementation, keys sorted, two spellings: two-space indent for the corpus files and the block carry's body, compact for the inline carry. commonmark-subset/ green.
  • 2b — Block nodes. block-nodes/ green. A nested list that cannot interrupt the block above it is refused meanwhile, not spelled: the maintainer's answer on tight-versus-blank separation turns that refusal into an emission. The test is broader than the name it carries — interruptsParagraph reads the next list alone, so a list after a block no paragraph continues, a code block say, is refused too — and the same answer narrows it. Block separation becomes separationBetween(previous, next, container) here — a boolean cannot hold the third case spec/flavour.md states for two directive blocks in a container body, and the maintainer's answer on a CommonMark block beside a directive block (1d) drops into the same seam. Give the emitter's refusals a corpus home while the directories grow: corpus/unspellable/, a .json beside the ConvertErrorCode it must return, the emitter half of corpus/errors/.
  • 2c — Inline nodes and marks. inline-nodes/ green. InlineSegment splits into its two axes — escapability (backslash, bracketed, none) and the emphasis role. A lone surrogate in a text node emits verbatim and becomes U+FFFD on any UTF-8 encode, a §2 break plain text still holds open — attribute values already escape it. The pipe form's fallback reads the emitted segments rather than naming the nodes whose attribute values spell a pipe as syntax, so 2e's \u007c narrows it in one place.
  • 2d — The opaque carry (§3). Fixtures and emitter together, into corpus/round-trip/opaque-carry/: an unknown node in both positions, the reserved adf info string, and the codeBlock whose language is adf — carried whole ahead of the attribute fallback 2e owes, since the reservation leaves that node no other spelling whatever 1d decides for its localId.
  • 2e — Carve-outs and combinations. Fixtures and emitter together, into corpus/round-trip/combinations/.
    • 2e1 — The carve-outs and the claimed line. The three carve-outs and their escapes, and a paragraph line inside a container body shaped like a closing fence (:::, ::: x). Guard fenceNestingFault's bare-run pop here too — a run shorter than the open fence is a fault, not a close — which today's emitter cannot reach.
    • 2e2 — Mark runs and the runs a carry breaks. The longest-run rule, attributes included, and a mark spelling that cannot open where it sits (un**-real**istic; the spec owes the carry a trigger). One mark vocabulary lands here, before 2e3 changes the attribute spelling: emphasisSpellings, linkAttributes and the code/link names join the mark table (3a parted it across adf/mark-attributes.ts and markdown/mark-spellings.ts), which holds four of the nine marks while the rest are branch literals in the emitter — and the parser (3) needs every name to make :em[x] the named error spec/flavour.md promises.
    • 2e3 — Attribute canonicalization and the quoted value's escape. Settled (the maintainer, 2026-08-26): a quoted attribute value escapes `, &, < and | as \u0060, \u0026, \u003c and \u007c, in every directive, block and inline alike — the constructs those four open all bind at or before a directive does, and nothing else reaches into {attrs}. Emitted attributes being inert leaves 3 free to keep CommonMark's own precedence between a directive and a code span, and collapsed escaping: 'attribute' into none. The escaper's link-opener scan skips emitted syntax to match: a ]( inside a directive escapes no text [.
    • 2e4 — The carry's fallback triggers. spec/flavour.md carries a node its section cannot spell — an attrs key no section lists, a value that is not the section's type, an arg slot holding no bare token, marks no nesting spells — where the emitter still refuses, which leaves the refusals a container's own spelling owns. The flanking trigger 2e2 added to that list is the odd one out: unspellableMark finds it after assembly and names a mark type against the line's path, so the failing run needs identifying before the carry can replace the refusal mark-inside-word pinned.
    • 2e5 — Combined documents and the collision property. Documents combining nodes rather than isolating one, and the gate's collision property: no two corpus documents may emit the same bytes — one spelling for two documents is a round-trip break no parser can undo, and it is provable without one. Settled (the maintainer, 2026-08-27): the approximation this item inherited — flanking exact, CommonMark's matching unmodelled — had two round-trip breaks reachable by hand, so the emitter now models the matching. process_emphasis runs over the runs the emitter wrote (emphasis-matching.ts) and a pair it hands to another delimiter rides the carry, which is what the multiple-of-3 rule did to the em in un*a**b*****c**istic. A delimiter run in text now escapes wherever CommonMark could open or close with it, not only open: one that could only close stole the spelling around it (un*a* b*istic), and escaping both ways keeps every delimiter the emitter did not write out of the matching. The canonical form gained a backslash where a run only closes — \*not emphasis\*, and 2e1's carve-out-strike a third and fourth.
  • 2f — The attributes CommonMark cannot hold. 1d's settled answer: the block nodes CommonMark spells — blockquote, bulletList, codeBlock, heading, listItem, orderedList, paragraph, rule — get directive sections in spec/flavour.md carrying localId, codeBlock's hideLineNumbers, uniqueId and wrap, and blockquote's marks, while hardBreak's text and localId join the inline directive it already has. The plain spelling stays wherever those attributes are absent, so only a node that carries one takes the directive form — which is what keeps a real payload readable rather than a page of carried JSON. Fixtures and emitter together, and the three documents the answer settles leave corpus/unspellable/ as round-trip pairs: block-local-id, code-block-empty-language, ordered-list-start-one. Settled (the maintainer, 2026-08-27): the codeBlock directive's body is one fenced code block, the language staying on the fence line so every renderer still highlights it; a language no info string holds — empty, a backtick, edge whitespace, an entity reference or the reserved adf — rides the language attribute with the fence bare, which retires 2d's carry for the reserved name along with the premise that left it no other spelling. The plain spelling gives way wherever it cannot render what the node carries rather than only where it has no place for it, so a heading level absent or outside 1-6 and an order whose markers would run past 999999999 take the directive form too, and ambiguous-attribute-spelling, unspellable-code-block-language, unspellable-list-marker, unspelled-block-marks and unsupported-heading-level leave ConvertErrorCode; content and placement refusals stay, which leaves the directive form spelling an empty list or a non-listItem child that the plain form refuses. order is the first marker, so order: 1 keeps the plain 1. — what a real payload carries — and a list carrying no order has no number to take and takes the directive form. 2f raises what 1d's unspelled block separation costs: a single localId on a paragraph beside a plain one now refuses every container body that is a directive's — a panel, an expand, a table cell — where before 2f the attribute refused the document anyway.

• 3 — markdownToAdf (0.1.0). Each sub-item lands the fixtures its own code reads, and the runner grows a parse half as they do: readers for corpus/normalization/ (setext, indented code, loose lists, */+ bullets, entity references, soft wraps — one-way, the markdown not canonical) and corpus/errors/ (a markdown input per named error, the code in a .error beside it) with the first fixture each. commonmark-subset/ cannot be the first to green — ::paragraph and :hardBreak{} sit in it — so 3b through 3f answer to their own tests and the one-way fixtures they land, and 3g is where the first directory reads back. The raw-HTML element mapping is empty until milestone 6, so at 0.1.0 every raw-HTML construct in input — a block, an inline tag, a comment, a processing instruction — is a named error. Input is where unbounded nesting actually arrives, so §11's 500 binds all three of the emitter's guards here: block depth at 3c and again at 3f's container fences, inline and mark depth at 3f and 3i, a carried value's JSON at 3j, where isJsonValue already bounds it.

  • 3a — The hierarchy. Mechanical, ahead of the first parser file: src/adf/ and src/markdown/ (html/ arrives with its first file, 6-7), the grammar module shared inside markdown/, and emphasis-matching.ts beside it — the parser reuses it whole, delimiterFlags and matchEmphasis taking CommonMark's own run vocabulary rather than the emitter's, so no second process_emphasis exists to drift from the first. block-directives.ts and inline-directives.ts each part by file, a node table milestones 6-7 need in adf/ beside a markdown spelling that belongs in markdown/. directive-attributes.ts cannot: vocabularyPairs walks the vocabulary and spells the value in one pass, the type check living inside spellAttributeValue, so the check comes out as its own predicate and goes to adf/ with the walk while the spelling stays in markdown/, isBareToken with it — only spelling calls it. markSpellings is the one table whose keys part rather than its file, so key the markdown half off the ADF half's type: a mark named in one and not the other is then a compile error instead of a false refusal. spellDestination, spellTitle and balanced leave markdown-inline.ts here too — CommonMark destination spelling emitLink and tryImageLine share, and the six concerns that file carries are one fewer for it. AttributeKind and AttributeVocabulary follow the walk into adf/, the vocabulary a string-typed attribute grammar needs and HTML will want too, not a markdown spelling. Settled (the maintainer, 2026-08-27): markdown/ parts here as well, into emit/ and parse/ with the shared set at the root — the grammar module, emphasis matching, the tables' markdown halves — and parse/ arriving with 3b's first file, the rule html/ already follows. And the node tables, a second copy of spec/flavour.md's prose whose mistyped attribute name degrades into a false refusal no test catches, get their guard: a test reads the spec's node sections, takes each name (type) list and asserts it equals the table, leaving the spec the source a human writes with no build step and no generated file. It is built at 3g, where a wrong entry starts refusing documents.
  • 3b — The leaf blocks. The line walk that opens and closes a block, ahead of any inline parsing: paragraph, ATX and setext heading, thematic break, fenced and indented code block, the HTML block whose lines it swallows whether or not the construct then errors, the link reference definitions a closing paragraph gives up, and the blank lines between them. The openers are commonmark-grammar.ts's — one table answers both directions, or the emitter under-escapes a line the parser reads as a block and §2 breaks in silence — the HTML block's start conditions excepted, which are new here since the emitter writes none. Block-level claiming lands here too: a colon run or an unescaped leading | is claimed, the parse behind it 3f's and 3h's, a claim with nothing yet to parse it the named error the claim promises meanwhile. The runner's parse half comes with it, and the first normalization/ fixtures, holding inline-trivial content so 3d and 3e add beside them rather than editing them.
  • 3c — The container blocks. Blockquote, bullet and ordered list: the continuation a marker's width sets, lazy continuation, and the tightness ADF does not record — > repeated being two bytes a level, so this is the cheapest way to reach §11's 500. 3b's leaf readers scan the physical line themselves, so a container re-cuts the walk rather than adding to it: the open containers' prefix comes off the line first and the readers take one line at a time, LeafBlock renamed with the union they join and blockNode's chain gaining their branches. Settled (the maintainer, 2026-08-27): a claimed line ends lazy continuation, so a closing fence on the line after a blockquote's open paragraph closes its container instead of continuing the paragraph CommonMark would fold it into. Claiming at block level is already absolute, and this binds input alone — 2e1's closing-fence-line orders the emitter's blockquote away from the edge either way — so spec/flavour.md's claiming paragraph gains the case here. And 2b's tight-versus-blank, one answer for both directions: the tight spelling stays wherever it parses back, a blank line going in only where the nested list would be swallowed — interruptsParagraph inverted from a refusal into the separation it names, and spec/flavour.md's "none between a nested list and a CommonMark block above it" gaining that exception. Every fixture spelled tight today keeps its bytes, and nested-list-tight becomes the round-trip pair nested-list-separation.
  • 3d — Inline text. The inline scanner over a block's content: backslash escapes, entity references decoding to their characters, code spans and the literal they hold — directive syntax and ~~ included — CommonMark's own hard breaks, a trailing backslash and two trailing spaces alike, a soft line break as one space, the fenced info string's own decoding the block walk leaves raw, and the raw inline tag, comment and processing instruction refused by name, recognized by the commonmark-grammar.ts predicates the emitter already escapes against, under 3b's one-table rule. Settled (the maintainer, 2026-08-30): entity references decode against HTML5's whole named table, checked in packed (§5) — a curated subset leaves 3k an exception class and a cutoff line nobody can defend. And the escape superset the emitter reads for raw HTML tightens into one precise CommonMark inline reader both directions share, the email autolink parting off as 3e's own predicate: refusing on the superset would refuse 1 <b 2, a fourth carve-out §4 and the README do not list. holdsEntityReference reads the table for the same reason, so &notareference; is emitted bare.
  • 3e — Emphasis and links. _, * and ~~ runs through matchEmphasis to the em, strong and strike marks; links inline and reference, 3b's definitions resolved here, autolinks, and the image gap's named errors — a titled image, and one amid other text. spec/flavour.md does not yet pin ~'s can_open/can_close, which is transcription rather than a decision: delimiterFlags already gives it CommonMark flanking, as for *, and §8 fixed that the moment the emitter shipped. Settled (the maintainer, 2026-08-27): 1d's deferred pair takes the backslash inside the delimiters it already has — [a](https://example.com/a\)b) and [a](/url "He said \"hi\""). <…> stays reserved for the destination holding a space, where nothing else works, so each construct keeps one spelling and a destination holding both composes. link-destination-parenthesis and link-title-quote become round-trip pairs. Settled (the maintainer, 2026-08-31): the destination escapes only the parenthesis it leaves unbalanced, so /wiki/Foo_(bar) keeps its bytes, and the backslash stays refused in both the destination and the title — which leaves [a](/a\b) a §2 hole 3k's exception list answers, as <http://x?a=1&amp;b=2> is, autolinks decoding neither escapes nor references. The image gap mints unmappable-image, mirroring unmappable-html — a construct in input no ADF node carries. And the CommonMark image shape lands here rather than at 3h: once […](…) reads, a lone ![alt](url) would otherwise misparse as text plus a link, so 3h keeps the rest of the media family and loses only that line. Settled (the maintainer, 2026-08-31, on the review): an empty link text — [](/u) — leaves the brackets the text they are rather than minting a refusal or dropping the destination, giving the label back the way an unresolved pair does, so the shortcut behind [][r] still reads. A description holding an image flattens to that image's own alt, which is what alt text means and what keeps the documented gap to mid-text and titled images; a break of either kind inside one reads as a space. And a destination or title whose entity reference decodes to a control character — [a](/x
y) — joins 3k's exception list beside the two above: the reader takes cmark's reading, the emitter has no spelling for it.
  • 3f — The directive grammar. The three forms — inline :name[content]{attrs}, container :::name arg {attrs}, leaf ::name arg {attrs} — the attribute grammar with its quoting and escapes, the fence-length and nesting rules, and the malformed list spec/flavour.md spells, each a named error. corpus.test.ts's fenceNestingFault stays a second reading of the fence rule over emitted bytes: the double entry is the check. Settled (the maintainer, 2026-08-27): the code span, the entity and raw HTML bind first in input, as 2e3 already assumed of the emitted side — a raw `, &, < or | inside {attrs} breaks the directive and is a named error, the author writing the \u0060 the emitter writes. One precedence covers both directions, and CommonMark's own ordering stays untouched. Settled (the maintainer, 2026-09-01): a directive whose name reads back to no node takes its own code, unknown-directive-name — a well-formed spelling the vocabulary does not hold is not a malformed one, and §8's "erroring input gaining meaning later is MINOR" is what a consumer switches the two apart for. And input reads canonical spacing only: one space parting the name, the argument, {attrs} and each attribute pair, no padding inside the braces, trailing whitespace on a directive block line tolerated — §8 makes loosening a MINOR, so strict is the reversible direction. directive-attributes.ts becomes directive-syntax.ts with the readers in it: the whole directive grammar, both directions, beside the escaping regexes and the spellings it must not drift from. And the 500-level guards compose here for the first time — a recursive reader stacked on the block walk — so nesting-depth-composed pins both axes now rather than waiting for 3i's third. A closing fence closes the innermost open container however long its run, which spec/flavour.md's closing-fence sentence now says: a run reaching past the innermost leaves the fence it did not close a named error, which §2 prefers to closing more than the author wrote.
  • 3g — The node tables read backwards. commonmark-subset/ reads back, the first directory to. A parsed directive becomes its node: the name to the type and an unknown one to a named error, the arg to the attribute it names, each value to the type its section assigns, the body to content, the reserved marks key to the marks array. 3a's drift guard is built here if the answer there was yes. 3f leaves two here: Read<T> moves to src/result.ts once a second reader takes it, and the reserved adf name in block position needs an error of its own — 3f reports it as unknown-directive-name, which §8 makes the signal that a later MINOR may give the name meaning, and adf never will. Settled (the maintainer, 2026-09-01): the reserved adf name in block position is a malformed-directive — the grammar section states the reservation, so it is that spelling the name breaks — and a well-formed directive the tables refuse is unsupported-node-shape, the emitter's code for the same mismatch read the other way; AGENTS.md §8 carries the split. And input reads canonical {attrs} alone, keys in order and every value spelled as the emitter spells it, the error naming the spelling to write instead: §8 makes loosening a MINOR, so strict is the reversible direction, as 3f already settled for spacing. Settled (the maintainer, 2026-09-01, on the review): 2f's plain-versus-directive choice is read back here rather than at 3h — a directive spelling a node CommonMark holds is refused, so ::rule and :::blockquote are errors while ::rule {localId=…} is not. The parser asks spellsCommonMark, the emitter's own choice, rather than restating the per-node conditions: a copy would refuse the list whose first item reads back as a thematic break, which the emitter does spell as a directive, and §2 breaks in silence. Two refusals land here for a later chunk to lift, on the same rule: the inline [content] slot, which 3i opens for emoji, mention and status, and the codeBlock content model's fenced body, 3h's. Read<T> stays where 3f left it — the node reader knows its path and returns Result, so no second reader took it. The drift guard earned itself on the way in: the spec's text attribute was missing from three inline table entries, which the content slot spells and the vocabulary walk already passes over.
  • 3h — The block nodes. block-nodes/ reads back: the codeBlock directive's fenced body and the language attribute a bare fence leaves it; the media family's composition; and both table forms, the pipe table's cell split and its named errors. fenceInfo is a rule both directions answer alike and moves to the markdown/ root with the language attribute. Settled (the maintainer, 2026-08-27): 1d's last pick, the one container-block-separation holds — a CommonMark block and a directive block sit adjacent in a container body with no blank line between them. That reduces the three cases to one rule, separation only where its absence would merge the blocks: the ::: fence is separation already, and 3c's claim ends the lazy continuation that would otherwise swallow it. The fixture becomes a round-trip pair, and with nested-list-separation and 3e's pair that empties corpus/unspellable/: this chunk settles the directory's own guard in corpus.test.ts too, and unspelled-block-separation, which loses its only cause here. The emitter's other refusals survive on causes no fixture in that directory covers, so 3k's one-list pass is where they get fixtures or the directory goes. Settled (the maintainer, 2026-09-01): losing that cause closed one of the shapes input accepted and emit refused, not the last. Two adjacent lists of a kind are what adfToMarkdown refuses and one - spelling cannot hold apart, and the walk reached them two ways — a marker change, which CommonMark opens a second list on, and an empty last item, whose blank line pops the container the list's identity hung from. The parser opens no list beside one of its own kind instead, the way it already drops the blank lines between items; 3k owes the CommonMark suite an exception where the reference HTML holds two <ul>. The normalization/ arm emits each document and reads it back from here, so the population that class lives in is checked rather than read. The README's canonical-fixpoint sentence still claims more than the parser keeps — 3e's three shapes — which stays milestone 5's to narrow.
  • 3i — The inline nodes and the marks. inline-nodes/ reads back: the content slot's text attribute and the error a slot holding anything but one unmarked text node is; the :text{text="…"} whitespace spelling; the four directive marks and their nesting order, outermost first; and :em[x] as the error spec/flavour.md promises. Editor-normal's merging half lands here, text-whitespace being the first fixture that forces it, and 4's toEditorNormal is built on it. 3g's shape leaves three: readInlineDirectiveNode takes the name, the attributes and the slot's parsed text rather than the span, since inline-content.ts already imports it and parsing the slot inside it is a cycle; the four directive marks get parse/directive-marks.ts that inline-content.ts tries ahead of the node reader, as mark-spellings.ts sits apart from emit/inline-directive-spelling.ts; and the five markdown-spelled mark names in inline directive position take unsupported-node-shape rather than a code of their own — §8 already answers a well-formed directive the node tables refuse, and the message names the spelling to use (*x*), while unknown-directive-name's "a later MINOR may give the name meaning" stays the wrong signal, as it was for adf. corpus/errors/directive-content-slot goes when the slot opens. The marks a spelling wraps answer the same question 3g settled for a block's form: only the nesting the emitter writes parses back. Settled (the maintainer, 2026-09-01): :text reads back what the emitter writes and nothing else — one run of spaces and tabs, or one run of newlines. A mixed run, and text CommonMark carries plainly, are named errors, as 3g refuses the directive form of a node CommonMark spells. The reader takes the slot's parsed nodes rather than its text, so the rule refusing anything but one unmarked text node sits beside the node tables that own the slot. Only text, read ahead of the slot, names a refusal before the slot's own: a doubly-broken span reports what its content holds, :date[<div>]{timestamp=1} being unmappable-html rather than date takes no content, which the maintainer pinned with an assertion rather than reordering the readers. directive-content-slot stays with the fixtures, its cause now a marked slot rather than a slot at all. The slot's own whitespace answers the rule the spelling does: :text{text="\n"} and 
 alike reach a slot the emitter refuses a line ending in, so one function answers both directions. Settled (the maintainer, 2026-09-01): a name the other position spells names that spelling rather than reading as unknown — :::em and ::date take unsupported-node-shape naming the inline form, :paragraph[a] the block one — leaving unknown-directive-name for a name no table holds, which is the meaning §8 gives it. The two readers lean on the tables being disjoint, so that is a test beside the spec drift guard now. The same read found the hole the other way: attemptLine refused a line edged with a vertical tab or a form feed, where CommonMark strips spaces and tabs alone, so valid CommonMark parsed to a document adfToMarkdown then refused. The edges that check covered are carried before the line is assembled, so narrowing it to spaces and tabs left it no cause and it goes with them.
  • 3j — The carry and the combinations. opaque-carry/ and combinations/ read back: the adf fence and :adf{json="…"} restoring a deep-equal node, invalid JSON in either a named error, a carry inside a mark spelling another, and the three carve-outs' escapes reading as the literal text they hold. 3g refuses the adf fence rather than reading a codeBlock from it; the refusal goes when the carry reads it. 3i left the slot parse contextless, so the refusal a carry inside a mark spelling earns needs a channel — a reader context in place of parseInline's strip flag, or a return arm from the slot — and directiveNodes takes its fourth reader beside it. index.ts gains markdownToAdf here, and the README's status line with it: this is the last parser chunk, so parsingDirectories becomes emittingDirectories and the whole corpus round-trips both ways — 0.1.0's proof, which 4 widens rather than replaces.
  • 3k — The CommonMark spec suite (0.2.0). Checked in at corpus/commonmark-spec/, pinned to the version it ships — the one commonmark-grammar.ts names for its start conditions — corpus/README.md gaining the kind. Settled (the maintainer, 2026-08-27): three checks an example must pass, the reference HTML each ships read as corpus data — which adds no format and no direction (§1). §2's canonical fixpoint: a named error, or markdown that parses and emits to itself byte for byte. That HTML's text, tags stripped and entities decoded, against the parsed document's concatenated text. And a count of the dozen elements the CommonMark subset covers against the marks and nodes they map to — counting distinct mark types per text node, since 3e collapses a spelling nested inside its own kind and *(*a*)* is two <em> against one em. The fixpoint alone is self-consistency a parser returning the empty document passes, and the text alone one dropping every emphasis; the counts close both. The exception list stays the maintainer's, and one entry is owed already: 3h continues a list across the marker change CommonMark splits on, so an example the reference HTML gives two <ul> counts one bulletList. One outcome is no exception and must not be filed as one: a fixable §2 hole — valid CommonMark parsing to a document adfToMarkdown refuses — which is what corpus/unspellable/ held until 3c, 3e and 3h landed their answers and emptied it. The permanent ones — a link destination or title no escape spells, a paragraph opening with a code span — are the exceptions, named by AGENTS.md §2.

• 4 — Round-trip property tests (0.2.0), widening 3j's corpus round-trip past the documents a human wrote — the thing that proves 2 and 3 beyond them. Settled (the maintainer, 2026-09-13): fast-check generates and shrinks. The gate runs a fixed seed, the properties together adding about five seconds per engine; an environment variable raises the runs and randomizes the seed for local digging, and a counterexample found becomes a round-trip fixture. The generators draw from the node tables — each node's content model and attribute vocabulary as adf/ records them, which 11 holds to Atlassian's schema — and misplace a share of nodes so the carry (§3) is exercised; no JSON Schema walker enters the tests. toEditorNormal stays internal. 2e5's collision test goes, since a collision already fails the round-trip on the same fixtures; the fixture-duplicate test stays.

  • 4.1 — Editor-normal and the node accessors. toEditorNormal(doc) in src/adf/editor-normal.ts, on 3i's merging: adjacent text nodes carrying identical marks and no attributes merged, an empty attrs, marks or content the absent key, -0 read as 0 (§2); the round-trip tests compare the parser's output through it, and serializeCanonicalJson beneath it walks iteratively. nodeContent/nodeAttrs/nodeMarks replace the 46 inline ?? []/?? {} reads in src/ (23 content, 12 marks, 11 attrs) and the attrs?.[key] reads, and the branch floor rises to the integer floor of what the suite then measures. Settled (the maintainer, 2026-09-14): a text node carrying attributes never merges — 0.1.0 merged a carried one into its neighbour on read-back — and the fix lands here, as does the iterative serializer.
  • 4.2 — The ADF property. fast-check joins devDependencies, AGENTS.md §5 naming what it earns — shrinking a failing document to the nodes that break it — and §10 the properties beside the corpus. A generated editor-normal document either refuses in adfToMarkdown with a ConvertError or reads back through markdownToAdf to an equal document, and nothing throws, under Node, Deno and Bun alike. 2e5's collision test is deleted. Settled (the maintainer, 2026-09-14): about half the block positions draw attribute-less CommonMark shapes — single-type lists, headings, blockquotes, pipe-table-shaped tables — where the escaping lives. A round-trip break the property finds is fixed inside 4.2, one commit per break with its round-trip fixture seen red first, and 4.2 lands when a deep run of about 10,000 per engine passes clean; a break needing design goes to the maintainer. The first two, both shipped in 0.1.0: an empty href with a title spelled [a]( ""), which reads back as the href "", and a [ or ] in a link's destination or title inside a directive mark, refused as the emitter's own output or, with ], losing the link. Later, settled the same day: an autolink whose href holds a backtick takes the [text](url) form inside a directive's content; and a V8 fault the deep runs hit — once JSON.parse has read a key holding an escaped backslash, a later escaped quote or newline key comes back as that backslash, on Node and Deno but not Bun — is accepted rather than worked around, since the library only refuses such a document, so the generators' JSON keys avoid those characters; the fault is reported upstream (https://issues.chromium.org/issues/521080746, nodejs/node#63785), where the maintainer added to both on 2026-09-14. The review found one more break of the same class, fixed the same way: a would-be inline directive in a link target inside a directive's content.
  • 4.3 — The markdown property. Generated markdown through markdownToAdf never throws, and the runs fit the budget; where it parses and adfToMarkdown spells the result, that spelling parses and emits to itself byte for byte (§2). Settled (the maintainer, 2026-09-15): the generators and run parameters 4.2 and 4.3 share live in one test-only module in src/, kept out of the build and coverage, with 10c's properties as its third user. Under the gate seed the property asserts floors on the runs reaching the fixpoint and on the directive-shaped ones. It lands when hunts of several hundred thousand runs per engine pass clean, since the breaks hit once per ~150,000 runs, past a 10,000-run bar. Two breaks, both shipped in 0.1.0, are fixed inside it. A backtick string an escape formed closed an earlier bare run's code span, since CommonMark reads no escape inside one: an escaped backtick alone or, as the review found, one joined to the bare run after it. A backtick run now escapes whole, and a bare run escapes wherever a later string of its length forms around an escape in the same inline content: the generalized pass the maintainer chose (2026-09-15). A paragraph's opening read as a link reference definition across a ] the emitter spelled: the emitter now escapes the opening [ exactly when the parser's own definition reader accepts the paragraph, and a link opening it rides the carry until 13b.
  • 4.4 — The real payloads. corpus/real-payloads/ holds ADF Atlassian's editor wrote, each round-tripped ADF→markdown→ADF with no expected markdown. Settled (the maintainer, 2026-09-15): the chunk authors the payloads itself on the maintainer's Atlassian test site — invented content, so nothing needs sanitizing — driving the editor with Playwright, and reads the ADF back over REST. Only the documents are committed; no client or fetch script enters the repo (§7). Later, settled the same day: the mentions keep the test user's real account id (the maintainer, 2026-09-15).

• 4b — The block walk's retry (0.2.0). emitBlock walks a subtree twice wherever readableBlock reads it whole and then gives up — a list item whose first line reads back as a thematic break — and the walk below does the same, so the cost doubles per level: 3.4kB of nested lists takes half a second, depth 20 about eight, depth 24 minutes. It predates 3g on both directions, and 3g's commonMarkSpelling gave it a second entry point. The README's bot and pipeline personas feed markdown nobody typed, so this ships as a hang on a small input; §11's scanning rule is the same argument one shape further in. The retry is what to remove — one walk answering both the readable question and the directive fallback. Memoizing emitBlock is the shortcut, and the node reference is the wrong key: a caller may hold one node object at two positions, where the cached depth and path are another node's. 0.1.0 ships with the retry in it, so a deep document is slow rather than wrong until the patch. adfDocumentFault is the second site to look at: isNodeArray reads every node and attribute value, then nestingFault reads them again, so the emit entry the export persona runs in bulk walks the document twice. Both walks are linear, so this is a constant factor rather than 4b's class change, and the parting is what gives depth its own code (§8) — measure before joining them back. Settled (the maintainer, 2026-09-18): the limit stays 500 readable lists, the walk reporting its headroom (§11). Counting every list twice was rejected for halving the limit, counting the directive form once for doubling the parser's frames per level. Measured (2026-09-18): adfDocumentFault walks a 9 MB document in 52 ms against 314 ms for the emit, so its two walks stay parted.

• 4c — The scanning rule's remaining sites (0.2.0). A trailing-anchored regex re-walks its run from every start position, so an interior whitespace run costs quadratic time rather than linear — 3h measured 80k spaces inside an ATX heading at 11.3s, and 3ms once the walk replaced the regex. The sites the same sweep did not reach: normalizeLabel in link-syntax.ts, whose shortcut-reference input is scan.source.slice(...) rather than the 999-capped readLabel value, and carryEdges in emit/inline-line.ts. A third of another shape joins them: readNestedDirective restarts its depth counter per level, so each parse level re-scans the region below it and nested inline directives cost O(depth × content), bounded by the 500-level guard. A fourth predates 12c: the list-item walk re-scans the rest of a line once per item level — isThematicBreak in containerStart on an opener line, isBlankLine and leadingColumns in continuesContainer on a continuation line, and a blank line continues every open item without consuming input; 30000 nested items take 4.4s at 59 KB (the stability-reviewer, 2026-09-16). §11's scanning rule is the whole argument; the pipeline persona feeds documents nobody typed. A fifth is a throw rather than a cost: adfDocumentFault pushes a node's content with a spread, so past about 125k sibling nodes the guard throws a RangeError where §11 owes a Result (the stability-reviewer and the maintainer, 2026-09-18). Settled (the maintainer, 2026-09-18): the five sites land in one PR rather than split into sub-items, and a behaviour-preserving cost fix is accepted on the suite staying green with no fixture output changed, plus the measurement below — §14 promises no figure, so nothing times the gate. The guard's spread is the one behavioural fix and carries a test. Corrected (2026-09-18): the entry filed two sites in emit/inline-line.ts on 2026-09-01 and the file has changed since — tryImageLine's alternation measures linear (3.4 / 1.9 / 5.3 ms over 10k / 20k / 40k spaces), leaving carryEdges' trailing trim the only one. Measured (2026-09-18), each at the size its filing named: normalizeLabel 1026 ms → 5 ms at 40k interior spaces, carryEdges 1024 ms → 7 ms (its heading path 978 ms → 5 ms), readNestedDirective 434 ms → 10 ms at 397 kB and 200 levels, the list-item walk 4196 ms → 39 ms at 30000 items, and the document guard a RangeError → 42 ms at 200k siblings under one node. The list-item walk's mixed-marker shape, which the fix had to answer too, reads 38 ms where the tail scan alone would have left it quadratic. The sixth site the sweep found went to 18 rather than landing here (the maintainer, 2026-09-18). Left as is (the stability-reviewer, 2026-09-18): of the list-item walk's three re-scans only containerStart's is fixed. continuesContainer's pair costs the same either way — 400 levels at 627 kB read 469 ms before and 448 ms after, linear in the line count and only mildly superlinear in a depth the 500-level guard bounds — so it is measured and left rather than made an item. Widened (the systems-architect, 2026-09-18): the guard's spread was a class rather than a site, and two more threw out of the public API — readIndentedCodeLine releasing the blank lines an indented code block held (200k of them at 200 kB), and emitRun joining a mark run's segments (200k nodes under one mark). Both fixed here with the same loop and a test each, and §11 gained the rule so the spelling cannot walk back in.

• 4d — What the gate says while it runs (0.2.0). ci.sh runs nine legs and announces none of them, so five minutes of a Gitea run read as silence and a hang cannot be told from a slow pull — the maintainer hit exactly this on the 0.1.0 release. Three causes, each its own fix. The legs need markers: plainpages' ci.sh prints a step() header per leg and this one prints nothing, so name the leg and the image before each. The longest leg is the quietest: test_output=$(… npm test 2>&1) buffers the whole Node run to replay it after, because the zero-test guard greps the count — stream it and grep a copy (tee), rather than trading the output for the guard. And two legs are silenced outright, npm pack and the tarball install, whose >/dev/null predates the offline install that made them quick and quiet. publish.sh owes the same: today it says nothing between reading private and the registry answering, which is where its npm ci and rebuild sit — the seconds §9 accepts rather than promoting the gate's dist, and unmeasured until the log shows them. Per-leg timing is what turns "slow or hung" from a guess into a reading; the browser leg's own 5.4–7.9s against a 17s warm gate is the number that made it obviously cheap. Measured (2026-09-20): ten legs, not the nine counted above, each naming the image it runs in where it runs in one, on a 28.4 s warm gate — install 1.5 s, typecheck 1.2 s, Node tests 4.9 s, Deno 6.0 s, Bun 4.5 s, build 1.0 s, pack and install 1.6 s, consumer typecheck 1.0 s, engines floor 0.4 s, browser 5.6 s. The browser leg lands in the 5.4–7.9 s the item quotes, and the markers cost nothing measurable: 28.9 s before against 28.4 s after. publish.sh reads its fields in 2.6 s and the registry in 1.4 s; its npm ci and rebuild are the gate's own 1.5 s and 1.0 s, so the seconds §9 accepts for rebuilding rather than promoting the gate's dist are about 2.5. Four things the writing turned up, three of them bash scoping a rule differently than it reads. The markers print to stderr, so a leg whose value is read — publish.sh asking npmjs — stays capturable. leg's locals carry its own name because bash scopes them into whatever the leg runs: unprefixed, name was swallowed by the leg reading package.json. leg returns its command's status the way with_firefox already did, because the bare call dropped a non-zero one wherever set -e is suspended, which also gets the elapsed time printed for the leg that failed. And the || that captures that status suspends set -e for everything the leg calls, so a function a leg runs chains its statements with && or every statement but the last runs unchecked: read_package_fields read on past a failed read, and push_tag pushed a tag the tag step had refused to write, both of which aborted before this chunk (the stability-reviewer, 2026-09-20). §10 carries the rule so the next leg cannot reintroduce it, and EPOCHREALTIME is guarded at source so an older bash names itself rather than dying as an unbound variable on the first leg.

• 5a — Rename to @larvit/adf-codec (0.1.0). Before the first publish, the name being the published identity: package.json name and repository, the Gitea repo and its remote, the README title, §6's published-as line, the checkout directory. Settled (the maintainer, 2026-09-01): ADF's own A is "Atlassian", and "converter" is the one-way lossy tool §2 exists to replace, where a codec is both directions. It names the hub, not the formats around it.

• 5b — The consumer's error surface (0.1.0). A product-owner read of the public surface found the error result legible to the library and opaque to the consumer holding it, and the README documenting no part of it. The sub-items are that read's answers, and they land before 5 because §8 freezes the code list at 0.1.0 and 5b4's table is what reads the list before the freeze closes it.

  • 5b1 — The error's source position. A parse error names an ADF path into a document the caller does not hold yet — unmappable-html at ["content", 5] for a <span> on line 12 — and no coordinate into the markdown string it passed in. ConvertError gains an optional position the parser carries to every parse-side mint, and the README's published shape gains it. Settled (the maintainer, 2026-09-03): the position is the parse side's alone — an emitter has no source string to point into, so emit-side errors keep path unchanged. The representation and where the position is captured are implementation judgment. The block walk mints it and the node walk attaches it as results return — at blockNodes, and at the inline body a directive holds — so the innermost block wins and the emitter's own refusals, which the parser re-enters for the CommonMark spelling, get an input coordinate too. markdownToAdf wraps the walk once more, which is what turns the wide Result<T> into the Result<T, ParseError> its signature promises rather than guarding anything: the depth guard under it cannot fire at depth 0. A paragraph names the line its kept text starts on, never a link reference definition it gave up. Line endings stay as the input spells them, so an offset indexes the string the caller passed rather than a normalized copy of it. §8 records the framings the review settled beside it: unsupported-node-shape stays one code across the two directions, unmappable-html names the version rather than the element, and a direction that reads a source returns the narrowed error type.
  • 5b2 — The error messages. Most state the rule and leave the violation to be inferred — a text node holds text for a node holding none — so rule: violation becomes house style across the sites that do. not-an-adf-document gives one sentence of eight words to null, a string, a missing version, a type that is not doc and a REST envelope around the document; naming the check that failed makes the highest-frequency integrator mistake self-diagnosing without the library naming a REST shape (§7). The three carve-out claim messages name the escape that unclaims the line — \|, \~~, \::: — which today only spec/flavour.md holds. unmappable-html reads as §8 now frames it: this version converts no raw HTML, never a permanent judgment on the element. Thirty-odd sites gained the violation clause and §8 gained the house style. isAdfDocument parts into adfDocumentFault, the guard reading it, so the first failing check is the message — the wrapper mistake names the key it found. Two carve-outs claim a line, not three: a matched ~~ pair spells strike silently, so nothing refuses it and no message names \~~. The escape lands on the refusals a prose line hits, in the form that was claimed — \::: on the directive line's, \| on the pipe table's, \: on the inline directive's, the attribute-pair and unknown-name faults taking whichever form read them. a pipe table row holds 1 cells gained its plural.
  • 5b3 — The code list and the flavour's gaps. A second read of the surface, this one on the fifteen names §8 freezes at 0.1.0: two pairs of them are one cause each, and one names a state the flavour leaves no way out of. unspellable-character and the text half of unspellable-whitespace are one refusal — a character CommonMark rewrites, the message naming it — and merge, unspellable-whitespace keeping the code for its other cause, the content slot no inline directive spans. unspellable-link-destination and unspellable-link-title become unspellable-link, the message naming the attribute. unspellable-adjacent-lists goes entirely: two adjacent bulletList nodes are valid ADF a site writes, and refusing them leaves the viewer persona a document it cannot render at all, so the flavour gains the separator that spells the pair apart, both directions, spec/flavour.md and fixtures. Thirteen codes stand — unspellable-whitespace keeps its own. The bare pipe table — a | b over --- | ---, GFM's shape without the leading pipes — is the one input that loses structure silently, reading back as a paragraph of prose; it becomes a malformed-pipe-table naming the form a row takes. That code keeps its name for the alignment colon: the flavour's own delimiter row is - runs, so the grammar is what refuses, and §8 records it rather than answering it again each review. The ninth adfDocumentFault branch names no node and carries the document's own path, the one branch the other eight outshine; §8 records why the guard stays a boolean. ::listBreak is the separator's spelling: the grammar's leaf form, and a second reserved name beside adf — every other name is an ADF node or mark type, and this one builds none. It reads only between two adjacent lists of one type, and takes the separation any directive block takes where it sits, so a directive container holds it with no blank line. A hard break is the one spelling that can put a bare delimiter row under a row of its own, so the emitter escapes that line's first character rather than refusing the document. spec/flavour.md had two directive blocks inside a container taking no blank line; the rule both directions keep is that a pair holding one takes none.
  • 5b4 — The README's consumer surface. §8 invites an exhaustive switch on code and no code name appears in the README, so it gains a table — code, when it fires, what the consumer does — grouped by direction, over the thirteen names 5b3 settled. Four things a reader who has not opened the code cannot know: raw HTML is core CommonMark and every construct in input is an error until 0.3.0, which the guarantees' "three carve-outs and one gap" denies and which is the bot and LLM personas' most common failure; adfToHtml, htmlToAdf, markdownToHtml and htmlToMarkdown sit unmarked in the code block people copy from, as do the two HTML guarantee bullets, and take a 0.3.0 mark or leave the block; adfToMarkdown is partial on valid ADF — a text node holding a carriage return, a link destination no canonical escape spells — which the viewer persona needs told along with what to do about it; and GFM past tables and strikethrough is literal text, task lists taking :::taskList. One sentence for the LLM persona: code is stable across minors, message is free text. The type-level surface freezes at the same moment and gets the same read: what index.ts exports and what it withholds, ParseError against ConvertError where a direction reads a source, and ConvertFault staying internal — the README table names the shapes a consumer switches on, so the two audits are one. The direction grouping is read off the call sites rather than the code prefixes, which do not partition by direction: the parser asks the emitter which CommonMark spelling a node takes (§11), so six codes reach a markdownToAdf caller as well as an adfToMarkdown one. The trailing pipe of a pipe-table row is optional in input, not required; the leading one is what every row must carry.

• 5c — The build and the release pipeline. Split out of 5, which kept only the maintainer's own acts. The build: tsconfig.build.json gains emit of JS and .d.ts to dist/ (its own allowImportingTsExtensions forces noEmit, so rewriteRelativeImportExtensions lands beside it), plus exports/files in package.json. Publish-on-version-change (§9) as publish.sh, run by a main-only job needing the gate. The ConvertErrorCode freeze (§8) is checkable here: 3h landed the last decision corpus/unspellable/ held and the directory went with it, so what the code list holds from here is permanent. The parser's own code additions are read here as one list before that freeze — nine sessions mint them independently, and one cause wearing two codes is breaking to undo after 0.1.0. That read gets a test rather than an eye — every ConvertErrorCode member named at a production call site, the way spec.test.ts guards the node tables — since unspelled-block-separation outlived its cause until 3h went looking. All thirteen have a call site; the audit's find was the depth one 5 predicted, read wrong in its own text: an attribute value past 500 levels was unsupported-node-shape on parse and not-an-adf-document on emit, the document guard counting the attrs object as a level the parser does not, so a value at exactly 500 parsed into a document the emitter then refused. Depth left the shape predicates on both sides: isJsonValue structural and overNested beside it, adfDocumentFault returning the code with the message and attributeValue the reason it refused, so both directions answer with unsupported-nesting-depth naming the attribute, and isAdfDocument calls a deep document a document as it always did a deep block. engines.node gets its one-line proof too — the built entrypoint imported and round-tripped under a pinned Node 18 image, which cannot run the suite that type stripping wants 22+ for, but proves exactly what the field claims. Beside it, the emitted .d.ts typechecked from a consumer's position: declaration emit leaves the .ts specifiers rewriteRelativeImportExtensions rewrites in the JavaScript, and nothing else in the repo reads them the way an installed consumer would. 3k's exception list landing after the release left the README's canonical-fixpoint sentence claiming more than 0.1.0 keeps — 3e names three shapes that parse and then refuse — so it now says a parse succeeding is no promise of a way back, and names them.

• 5d — The browser leg. §6's browser half is checkable on the emitted dist/index.js a browser can load — the compile gate names no host API, and a real page converting the corpus is the other half. Headless Firefox is that page, settling both at once: the browser proof, and the only SpiderMonkey there is, the gate's three engine legs being two V8s and a JavaScriptCore that is not Safari's. The mechanism is the decision this item opens with: a browser leg wants an image, a driver and a way to carry a verdict back out, none of which the gate's plain docker run per engine has. The answer is with_firefox, which runs the Firefox image beside the node one in a shared network namespace, so the page's server and the driver are each other's 127.0.0.1 and no user-defined network, container name or geckodriver --allow-hosts entry is wanted; its EXIT INT TERM trap bakes in the container id, since the local holding it is gone by the time the trap fires. browser-tests/run.js serves the repo, drives one execute/sync and asserts the results against the corpus with the Node-side assert.deepEqual the corpus runner uses, so the browser page holds no second copy of the comparison. The whole corpus fits: 118 fixtures in 8s warm against a 120s script timeout — no slice was worth choosing. A try around the dynamic import is what turns a broken build into SpiderMonkey's own message rather than an undefined global. Settled (the maintainer, 2026-09-04): selenium/standalone-firefox over the smaller instrumentisto/geckodriver, currency over size — the leg's whole worth is a real SpiderMonkey, which decays the moment the pin stops moving, and the smaller image was four Firefox majors behind with a publisher that may go quiet while Renovate stays silent.

• 11 — Atlassian's ADF schema as the tables' truth (0.2.0). @atlaskit/adf-schema's two JSON Schemas vendored rather than the package installed (AGENTS.md §5), and the node tables gated against them (§10). Settled (the maintainer, 2026-09-13): vendored at spec/adf-schema/ and re-pinned by hand when a need shows; the gate compares attribute names and kinds, never value sets, over full.json and stage-0.json together.

  • 11a — The vendored schema. full.json and stage-0.json, byte-exact from @atlaskit/adf-schema@57.4.9's dist/json-schema/v1/, at spec/adf-schema/, each pinned by its SHA-256 in a test the way spec.json is. The version, the source and the Apache-2.0 attribution sit beside them with the licence text; no gate re-serializes either file.
  • 11b — The gate. For each node and mark type the tables spell, the attribute names and kinds equal the union over every definition in both files whose type enum names it, anyOf/allOf branches included, the argument slot (panelType, state) counting as spelled. Kinds: string; number, integer included; boolean; json for an object, an array or an untyped value; an enum-only attribute takes its values' kind. What the schema holds past the tables is pinned in two exact lists — an entry the schema no longer needs is red, like a difference neither list names: gaps, attributes of a spelled type (57.4.9: link collection id occurrenceKey, rule color style weight, layoutSection columnRuleStyle), emptied by 13; and carried, types the tables do not spell (alignment annotation backgroundColor blockCard bodiedRule breakout dataConsumer embedCard fontSize fragment indentation inlineExtension placeholder), doc and text counting as the grammar's own.

• 12 — The !adf: re-spelling (0.2.0). Replace the colon directive grammar with the namespaced prefix, a breaking change to the emitted contract (shipped 0.1.0, so §8 makes it 0.2.0). Forms: block container !adf:name arg {attrs} … !adf:/name — the / parts open from close, nestable without a fence-length discipline, so the ::::/::::: runs and their length rule go and every container opens the constant !adf:; block leaf !adf:name arg {attrs} with no closer; inline node !adf:name[content]{attrs}; directive marks !adf:border/subsup/textColor/underline [content]{attrs}. Attributes and their escaping stay {key=value}; the literal escape is \!adf:. Leaf vs container is decided by the node's content model rather than syntax — the ::/::: split goes, a simplification the carry makes safe (an unknown block node already rides the fence, not the directive). The carry's reserved name becomes carry, both spellings — the block fence info string carry and the inline !adf:carry{json="…"} — named for what it does: it carries a node verbatim, never "unknown-node", since a known node no section spells where it stands rides it too. No ConvertErrorCode is added, removed or renamed, and the round-trip guarantee and the carry both hold through it. Mechanical surface: the grammar in spec/flavour.md, src/adf/block-directives.ts + inline-directives.ts, src/markdown/'s directive-syntax.ts, opaque-carry.ts and the emit/ + parse/ readers, every corpus fixture (round-trip, normalization and errors/), the prose reader over spec/flavour.md, the markdown property's generator, and the README's examples. Settled (the maintainer, 2026-09-13): - A line opening !adf:name is a block line when a space or the line's end follows the name, and a paragraph when [ or { does. Claiming stays syntactic and structure comes from the tables: an unknown name is unknown-directive-name at the opener, whatever follows it. - An unescaped !adf: claims on its own anywhere inline: one completing no directive is malformed-directive, the emitter escapes every literal !adf:, and !adf:hardBreak{} keeps its braces. Block and inline share the one \!adf: escape hint. - A closer names the innermost open container, crosses no list-item or blockquote edge, indents as a fence does and carries nothing after the name; anything else is malformed-directive. - A node holding no content whose content model takes some is an empty opener–closer pair, never a leaf. - A spelled node's content model is frozen with its spelling: changing it is MAJOR (§8). - The colon spellings are dropped, not refused: 0.1.0 markdown reads back as prose, adf is no longer a reserved language, and MIGRATION.md tells a consumer to convert stored markdown through 0.1.0's parser and 0.2.0's emitter. - Inputs moving between codes ride the break: a leaf given a body, a container missing its closer and listBreak with a body are malformed-directive, and an empty inline-body container parses. - Split by construct, each sub-item both directions: 55 of 78 round-trip fixtures feed both the emit and the read-back test, so an emit-only chunk cannot land green. - The spec leads the code from 12a to 12d: spec/flavour.md and AGENTS.md §4 spell the !adf: grammar whole, while code and fixtures reach it one form at a time. A reader landing in either without todo.md sees a gap that is the plan, not a defect.

  • 12a — The spec and the decision. spec/flavour.md rewritten to the !adf: grammar and the settled answers above, no colon directive form left in it; AGENTS.md §4's directive bullet and prior-art line, and §8's escape hints and ::adf/::listBreak examples, name the new forms, §8 gaining the frozen content model.
  • 12b — The inline form. Inline nodes, directive marks, text and the inline carry !adf:carry{json=…} spelled and read as !adf:name[content]{attrs}, with the prefix claim and its escape; the round-trip, normalization and errors/ fixtures holding inline forms re-spelled, and the gate green. The content slot of emoji, mention and status refuses a text node carrying attributes as unsupported-node-shape, which it drops silently today (the maintainer, 2026-09-14).
  • 12c — The block form. Openers and !adf:/name closers, leaf vs container by content model, empty pairs, listBreak and the carry fence, spelled and read; the fence-length rule and the corpus test's fence nesting check deleted; the remaining fixtures re-spelled and errors/ re-derived under the shifted codes, and the gate green. 12b's two temporary seams expire here: carryFence folds back into carryName once the fence reads carry, and directiveLineEscape into inlineDirectiveEscape once one escape serves both forms. spellLeafDirective takes the Inline its reader-side regex already carries, and header versus opener settles as one word in spec and code. While readNestedDirective is open, its [content] and {attrs} reads lift out as named steps, and the two charAt-against-! fast paths ahead of claimsDirectivePrefix — in readDirectiveContent and line-escaping's bracketed-link-target arm — either earn a reason or go (the systems-architect, 2026-09-16).
  • 12d — The README, MIGRATION.md and the sweep. The README's examples and error tables follow, MIGRATION.md linked from one README line; docs and fixtures swept for any stale ::/:name spelling.

• 13 — The schema's gap attributes (0.2.0). Spell the attributes 11b pins as gaps, in 12's grammar, and empty the list. Settled (the maintainer, 2026-09-13): a link [text](url "title") cannot hold takes the directive mark !adf:link[text]{attrs} — one carrying collection, id or occurrenceKey, or an href or title no CommonMark escape writes — and a directive link CommonMark could spell is unsupported-node-shape. That leaves unspellable-link no cause, so it leaves ConvertErrorCode in 0.2.0, §8 recording the removal.

  • 13a — rule and layoutSection. rule's color, style and weight and layoutSection's columnRuleStyle join their tables and spec/flavour.md bullets, with round-trip fixtures; their gap entries go.
  • 13b — The directive link. link spelled as above in both directions, with a round-trip fixture per trigger, spec/flavour.md's Marks section following; unspellable-link removed from the code list, its errors/ fixtures and the CommonMark suite's unspellable exceptions it cures re-derived, the README's code table and its "not every document converts back" guarantee following, and MIGRATION.md naming the removed code; the gap list is empty. A round-trip fixture holds the shape 4.2's review left refused until then: an autolink-shaped link under a directive mark whose href holds \!adf:name{. A link opening a paragraph whose opening reads as a link reference definition, which 4.3 leaves riding the carry, takes the directive link too, with its round-trip fixture (the maintainer, 2026-09-15).

• 14 — The CommonMark subset's directory (0.2.0). src/markdown/ holds 16 source files at its root and 10 adds more there. The CommonMark subset moves under src/markdown/commonmark/ — backtick-runs.ts, commonmark-grammar.ts as grammar.ts, emphasis-matching.ts, entity-references.ts with its test, link-reference-definitions.ts and link-syntax.ts — leaving the flavour's own constructs at the root, the split spec/flavour.md draws between the subset and the flavour (the systems-architect and the maintainer, 2026-09-16).

• 15 — The href-less directive link (0.2.0). Refuse !adf:link[text] spelling no href with unsupported-node-shape naming the attribute, so the mark has one spelling: today it parses to a mark the emitter writes back as a carry, while the schema requires href and every other directive mark spells without attributes in both directions alike (the stability-reviewer, 2026-09-16; the maintainer, 2026-09-17).

• 16 — The link wrapping a link (0.2.0). Read [<http://x/>](/v) and [!adf:link[a]{href="/u"}](/v) as [[a](/u)](/v) reads — the inner link wins and the outer brackets stay literal text, CommonMark's rule that no link holds another — rather than dropping the outer link silently as closeLink's applyMark does today, with a normalization fixture per shape (the stability-reviewer, 2026-09-16; the maintainer, 2026-09-17).

• 17 — A machine-enforced size ratchet (0.2.0). Add a per-function line ceiling to the gate, set at today's worst and only ever moving down, so the largest body of new code cannot exceed what is already here (the systems-architect, 2026-09-16; narrowed by the comprehension panel, 2026-09-20). oxlint's eslint/max-lines-per-function measures it — one devDependency, carrying the musl binding the gate's image needs, since TypeScript 7 is the native compiler and exposes no parser to write the check against. No cyclomatic rule: eslint/complexity charges ?. and ?? a point each, the guards §10 already exempts from the branch floor, and its two worst functions, readBlockLine and parseInline, went unnamed by all nine readers while isNodeArray and blockNode were volunteered as among the clearest code here. Length ranks no better: emitList and blockNode are both 26 lines, one the panel's unanimous top four and the other the clearest map of the format in the repo. So the ceiling guards against drift and never drives a refactor — 19 to 27 are where the hard work actually is. Done (2026-09-20): .oxlintrc.json carries the one rule, correctness off so nothing else runs, over the 40 files tsconfig.build.json builds — the tests and property-harness.ts out, five functions over the ceiling with them, the worst 64. The ceiling is 52, parseInline's length and the built set's worst; at 51 the gate reddens on it. skipBlankLines and skipComments are spelled at oxlint 1.83.0's defaults, so a changed default cannot move what 52 counts. The leg runs npm run size-ratchet beside the typecheck at 0.8s, and the lockfile carries every platform binding, so npm ci resolves the musl one inside the image. §10 holds the rule and what each switch guards (the stability-reviewer, 2026-09-20).

• 18 — The subtree the directive spelling asks about (0.2.0). The parser asks commonMarkSpelling at every directive-spelled block and the answer emits the whole subtree below, so a node at depth d is spelled d times: three nested rule-first directive lists cost 18 asks over 10 nodes, and 250 levels parse in 1.2 s at 16.4 kB, 4.9 s at 261 kB with a kilobyte of content per level. The depth guard bounds the levels at about 250, never the content, so this is the pipeline persona's hang on an input nobody typed (§11). Keeping each child's emitted result for its parent's ask is not a straight handover: the same node object is asked at different depths — 4, 3 and 2 for the innermost list of three — because the parser counts a list and its item as two levels where the emitter's readable list counts one (4b), and headroom is that guard's slack. The parts that survive the measurement: the paths agree, text and spelling carry no depth, headroom is affine in it, and the parser asks first at the deepest of them, so a kept result rebases by the difference. Either rebase and record that argument in AGENTS.md, or give both directions one list accounting so a node has one depth and nothing needs rebasing — which reopens 4b. A single post-build walk was rejected: it reports the outer offender where the build reports the inner one (the maintainer, 2026-09-18). Measured (2026-09-19): the parse keeps each node's readable spelling (AGENTS.md §11), and 250 nested directive lists fall from 1.22 s to 0.04 s at 16.1 kB and from 5.20 s to 0.12 s at 261 kB; a panel between every pair of lists 1.00 s to 0.05 s, an opaque carry in every item 1.08 s to 0.03 s. No figure enters the gate (§14), as 4c settled for a behaviour-preserving cost fix; what the gate holds is the refusal. An ordered list past the marker cap gives way, so the emitter spends two levels where the parser spent one: one such list between an ask and a kept entry cancels the credit the directive form starts with, and two put the read below the depth that filled it — which a hit would answer without the depth guards the walk runs. That read re-spells. Three cases hold the arithmetic between them: the two depth-boundary tests already there pin the sign, the two-overflow case pins the guard, and the one-overflow case pins the magnitude, a constant rebase accepting there a document the emitter then refuses (the stability-reviewer, 2026-09-19). The one list accounting was not taken: 4b settled that accounting the day this was filed, and reopening it is an ask rather than a chunk.

• 19 — A home for what both formats read (0.2.0). Settle where a construct both formats need lives, and say so in AGENTS.md §11. Today adf/ may hold no format knowledge and each format directory holds its own shared layer, so there is no third place; the first ADF-shaped but format-touching helper either breaks the layering or becomes a second spelling of one rule, which is the loss §2 exists to stop. Both architects ranked this first and the only item cheaper before the feature than after. Settled (the maintainer, 2026-09-21): adf/ is that place, and the test is the vocabulary the answer is in — a node type, an attribute kind, a content model, never a delimiter, an element name or an escape. A helper that cannot answer that way is the ADF question there and a spelling per format, the seam markAttributes and markSpellings already draw; one that cannot be split is a gap to ask. markdown/ and html/ are peers with no third directory between them, and src/ root keeps the primitives knowing neither ADF nor a format. No code moved: a construct rises on its second consumer, so linkHref (markdown/mark-spellings.ts) and the external-image read (markdown/emit/image.ts), both pure ADF attribute reads, move when html/ reads them (7).

• 20 — The give-way channel is unmistakable (0.2.0). emitBlockquote, emitCodeBlock, emitHeading, emitList, emitParagraph and emitRule return Result<EmittedBlock> | undefined, where undefined gives way to the directive form and an error refuses the document. Give the six the try prefix the repo already uses for tryImage, tryPipeTable and tryPipeCell, or a return type that cannot hold both, so a newcomer meeting an unspellable shape cannot reach for failure and silently narrow what converts. Named as the first thing a new senior would break, and it lands on §1. Done (2026-09-21): both answers, each where it fits. Five of the six reach a refusal the directive form shares — a walk, codeBlockText, emitInlineLine — so moving it out is the second walk 4b removed, and they take the prefix alone. tryRule refuses nothing, so it takes the narrower type as well, string | undefined through readableText, which is what tryImage and tryPipeTable already return. emitLink and listItemLines are the same channel outside the six — the CommonMark link and the tight list item, each tried ahead of a directive form — so they take the prefix too, and the comment naming emitLink's undefined arm goes with the rename. §11's readable-spelling rule carries the prefix and gave up the set it enumerated: the prefix is the set now, which is what stops the next spelling being named by the eight beside it. The same rule's nested-list carve-out went with it, naming a refusal 2b's tight-versus-blank answer had already removed — unspellable-adjacent-lists is gone from result.ts and separationBetween returns a string.

• 28 — emitLine's retry loop cannot spin (0.2.0). emit/inline-line.ts:67 is a for (;;) that re-emits the line until every unspellable node has been carried, and its termination rests on a comment: each pass carries at least one more node, or flips openingLinkAsDirective, which happens once. Ten return success({ carry: … }) sites in that file have to honour it and nothing checks them — a range already inside carried loops forever. The library has no I/O and no timeout, so that is a hung caller rather than an error result, and §1's pipeline persona feeds documents nobody typed. Make the loop hold its own guarantee: refuse a carry that adds no node and return an error. Reads first in 0.2.0 because it is the only known way this library fails without a Result. Found by the comprehension panel, 2026-09-20; the ten sites are confirmed, a document that reaches the spin is not. Done (2026-09-20): the loop's progress is one named state and every pass takes a fallback through takeFallback, which refuses a carry adding no node and an opening link asked for the directive form a second time. Both are unsupported-node-shape under §8's rule that a new cause takes an existing code reading true of it: the emitter has no spelling left for that node arrangement, and a code a consumer can never switch on costs a removal later. Neither refusal is reachable — a carried node takes emitLeaf's carried branch before any run forms, so every range a site names holds an uncarried node, and the directive-spelled opening link leaves the first segment with no node range for escape to ask about — so both are uncovered branches like the repo's other guards, 98.92% to 98.84% against the floor of 98.

• 29 — The README reads raw HTML as refused for good (0.2.0). §Goals 3 says "the three carve-outs and the one gap below are the whole of the exception" and §The guarantees says "Raw HTML in markdown input is an error result", both reading as settled, where spec/flavour.md §Raw HTML in input says the opposite: markdownToAdf routes each construct through the foreign HTML element mapping, and only a construct without one is refused. The spec stands — commonplace markdown is accepted, and every tool writes some HTML (the maintainer, 2026-09-20). So rewrite the two README texts to name the exception that survives 6 and 7, a construct outside the documented element set, and state it in one place, since three already spell this one rule. markdown-to-adf.ts:73 and inline-content.ts:158 are the whole of the refusal and already say "at this version"; 7 is what makes them route. Done (2026-09-20): the rule has one home, the unmappable-html row, naming the element set and what its absence covers at this version; §Goals 3 bounds the exceptions without listing them, the standalone raw-HTML guarantee goes, and the 0.2.0 guarantee says markdown's raw HTML reads the same set. That guarantee's "never a silent drop" went with it: 6 settled that <script> and <style> drop whole, so the claim does not survive 7.

• 30 — AGENTS.md says each thing once (0.2.0). §15's ask protocol — name the class, cite the earlier asks of it, never "A or B?" — is the rule reviewers cite most and has no heading, two thirds down a 50-line section in a file with no index. Give it one. §15 also offers "the gate's seconds" as a stated number that is kept, and no such number is stated anywhere, §14 forbidding the category outright; drop the example. Then the restatements: §15 repeats the one-chunk rule three times and version/NPM_TOKEN twice, §12 says the default is delete twice, and §5's "few, each earning its keep; they never reach a consumer" is npm's own definition of the field. Cut to one copy each, the one carrying the why. Done (2026-09-20): §15 gains three sub-headings — Ask, don't guess; Rules the loop has settled; The continuous loop — so the protocol is one of four entries rather than a paragraph two thirds down. The one-chunk rule keeps the head paragraph, which now carries the chain-of-sessions why the settled bullet held; version/NPM_TOKEN keeps the reserved paragraph, which carries §9's publish-on-bump why, and moves up beside the chunk steps. §12's "every prose comment in a diff is a review question" and §5's devDependencies clause go whole. The gate's seconds is confirmed stated nowhere: docker-runner.sh measures each leg's elapsed time and no number bounds it.

• 21 — The ADF tables carry ADF's nouns (0.2.0). adf/block-directives.ts and adf/inline-directives.ts hold the ADF node tables — paragraph, heading, blockquote and rule among them — under the markdown flavour's word, inside the directory §11 forbids to know a format. Rename to the noun spec/flavour.md uses, types and accessors with them. Before 7 doubles the import sites. Done (2026-09-21): the tables are adf/block-nodes.ts and adf/inline-nodes.ts, blockNodes and inlineNodes — spec/flavour.md's own headings — each row a BlockNodeModel or InlineNodeModel read by blockNodeModel and inlineNodeModel. The row is the node's model, never the node: BlockNode beside AdfNode put two kinds of thing one lookup apart in every signature the rename touched, where §8 already calls the content model the model. markdown-to-adf.ts's own blockNodes/blockNode walk became readBlocks/readBlock, so the table's name means one thing repo-wide. BlockType stays: it names a node type, the parallel of mark-attributes.ts's MarkType. markdown/'s blockDirectiveForm, blockArgument and the spell*Directive* family keep the word, naming the spelling rather than the node.

• 22 — LineContainer sits at the markdown level (0.2.0). Two of the four parse/ → emit/ imports fetch this type from emit/line-escaping.ts, camouflaging the two that are the deliberate spelling consultation. Move it, and name those two in §11 as the whole of that surface, so a reviewer checks the seam with one grep. Done (2026-09-21): the type is markdown/line-container.ts, read by both directions, so parse/ imports emit/ twice — commonMarkSpelling and openingLinkTakesDirective — and §11 says those two are the whole of the seam a ../emit/ grep under parse/ reads.

• 32 — The mark depth adf/ counts is stated in ADF's terms (0.2.0). document.ts's markAttributeNesting is largestNesting - 3, and both the comment above it and AGENTS.md §8 give the reason as markdown's: the block directive spells the whole mark set as one JSON attribute, so the parser reads the value at the bottom of array, mark and attrs. That is a format's spelling deciding a constant inside the directory §11 forbids to know a format. Move the derivation to where that spelling lives, or state the three levels in ADF's own vocabulary. Before 7 gives the constant a second format whose spelling may not spend the same three levels. Done (2026-09-23): adf/ counts a mark's attribute from its value at 500, as a node's; the block directive's marks spelling refuses its own over-deep JSON with the parser's message and the node's path, and §8 says a deeper spelling refuses in its own format.

• 23 — The block-directive fragments are one file (0.2.0). block-directive-arguments.ts, -forms.ts and -marks.ts are three files under 25 lines answering one question. Fold them, and take src/markdown/ — the worst level both architects named, 13 entries with no organising question — down with them. Done (2026-09-23): markdown/block-directive.ts holds the form, the argument, the marks attribute and the listBreak spelling — what the block directive spells of a node that the grammar does not — and src/markdown/ drops from 15 entries to 12.

• 24 — The conformance gates have a directory (0.2.0). Six root tests with no sibling source (adf-property, adf-schema, commonmark-spec, corpus, flavour, markdown-property) plus property-harness.ts are the machinery that makes the docs executable, and they read as leftovers. Give them one, so src/ root shows what it holds. Done (2026-09-23): src/conformance/ holds the six gate tests and the property harness, so src/ root holds the primitives and the entrypoint.

• 25 — AGENTS.md §8 and §11 are findable (0.2.0). Both are single unindexed paragraphs holding the answer to nearly every question the panel had, and six readers and both architects independently reported that finding the sentence cost more than reading the code it governed. Sub-headings or an index at each section's head; delete whatever the code, a type or a test name already says rather than reorganising it. Done (2026-09-23): §8 parts into the code list, which code a cause takes, and message, path and position, as bullets one rule each; §11 into style, bounds, spellings and layout. What src/result.ts, README.md §The errors and ls src already say went, and the level's definition, matchEmphasis's transcription and emitLine's termination moved to one line at the code each defines.

• 26 — The two mutable structures say what they guarantee (0.2.0). escapedIndexes fills the escaped set left to right while the predicates it calls read the half-built set, then escapeClosedRuns walks the same set right to left and adds to it; the order is load-bearing and asserted nowhere, and a refactor to filter/map breaks it silently. Separately, walk.edges is the blockquote and list-item subsequence of walk.stack with each entry's stack index, maintained by hand in six places and stated in none. Put each invariant where it cannot be got wrong — a type, a derived value, a named phase — rather than in a comment. The panel's first and second hardest places. escapeClaims hands each predicate the one fact it reads — whether the character before was escaped — and escapeClosedRuns takes that pass's set read-only and returns the whole. containerStack is the one writer of the stack, edges and the open directives' depths by name, all read-only elsewhere; deriving edges from the stack instead makes every line walk the open directives, quadratic in their nesting.

• 27 — The dead headroom write goes (0.2.0). directiveItems (emit/adf-to-markdown.ts:258) writes item.walk.headroom - 1 onto each PlacedBlock, and nothing on that path reads a block's headroom: joinBlocks and separationBetween read text and spelling, and emitDirectiveBlock takes the level from the Walk. Of the two subtractions three readers flagged as double-counting, this is the one that is dead.

• 10 — Lossy conversion (0.2.0).

  • 10a — The reduction. adfToPlainMarkdown's ADF→ADF reduction, tests first, a test per row above.
  • 10b — The lift. plainMarkdownToAdf's ADF→ADF lift, tests first, a test per row it reads, other tools' spellings included; the editor's default highlight colour looked up and cited.

5 — Ship 0.1.0

• 5 — Ship 0.1.0. Only the maintainer's own acts are left (§15): make the Gitea repo public (§6), create the NPM_TOKEN secret, confirm the Actions token may push tags — the publish succeeds and the tag push then reddens the run, though the next push to main retries the tag alone — and open the bump PR that sets version to 0.1.0 and drops private: true, the guard against any earlier publish. The bump and the drop go in one commit: dropping private alone publishes 0.0.0, which also differs from npm's nothing. 0.1.0 is the markdown round-trip: both markdown directions, the types, isAdfDocument, proved over the checked-in corpus. Settled (the maintainer, 2026-09-01): the round-trip proved over the checked-in corpus is what 0.1.0 ships on, and the open-ended proof work follows it rather than gating it — 3k's spec suite and 4's generators and maintainer-supplied payloads are 0.2.0, 4b's retry 0.1.1. A consumer using the library is worth more than a wider proof nobody has needed yet, and §8's pre-1.0 rules cover what the wider proof then finds.

Shipped 2026-09-05: @larvit/adf-codec@0.1.0 published and v0.1.0 tagged on 8a847de. Publishing needed a bypass-2FA token — the account carrying no write-2FA requirement was not enough, npm demanded an OTP until the token itself bypassed it.