adf-codec/todo.md

Todo

The plan. Design questions are settled in AGENTS.md; remaining spec detail is settled at its own milestone. A done item shrinks to its title here; its full text moves to todo-history.md.

Next session

Start a session with: Read AGENTS.md and todo.md, then do what todo.md's "Next session" says.

1. git fetch origin first and read this file at origin/main, then branch off it, not the worktree left behind: a checkout behind the remote reads a merged item as unchecked.
2. The first unchecked item in shipping order, per AGENTS.md §15 — or, where that item has no release, the planning chunk §15 describes.
3. In flight: nothing.
4. Before stopping, rewrite this section: the in-flight line, and the prompt itself wherever the session found it wrong or short.

Milestones

Shipping order: 3h, 3i, 3j, 5a, 5b, 5c, 5d, 5 → 0.1.0 (shipped 2026-09-05); 3k, 11, 4, 12, 13, 4b, 4c, 14, 15, 16, 18, 4d, 28, 17, 29, 19, 20, 21, 22, 23, 24, 25, 30, 26, 27, 10, 6, 7, 5f, 5g → 0.2.0; 8, 9 → TBD; 5e last. The numbering is the order the work was planned in, not the order it ships. Everything known and shaped ships in one release rather than a string of them: nothing waits on a version, and no consumer is served by the churn (the maintainer, 2026-09-18). So 0.2.0 completes §1's three formats, and 0.2.1 and 0.3.0 are gone. 8 and 9 stay out as the two goals nothing has shaped yet. 0.2.0's order is settled (the maintainer, 2026-09-13, extended 2026-09-18): 11 makes the tables 4 generates from answer to Atlassian's schema, 4 proves 12, 13 spells 11's gaps in 12's grammar, and 12 rewrites code 4b and 4c change; then 14 moves the files 15, 16 and 10 edit and HTML is written against that layout, 4d marks the gate legs before 17 adds one, 17 puts the size ratchet under the largest body of new code, and 5f and 5g read last because 7 is what changes the bundle size and the tagline. 19 to 27 come from a comprehension panel — nine readers across four experience levels, none of them able to see this file, reporting what defeated them and whether the project's shape fits in a head (2026-09-20). They read ahead of 6, 7 and 10 because every one of them is cheaper before the HTML format lands than after: 19 and 20 because HTML has no answer without them, 21 to 24 because HTML doubles the importers and the file count they touch, and 25 to 27 because they are what the panel says the next reader pays for. 29 and 30 come from 17's prose pass (2026-09-20). 29 reads first because every goal is what a later ask is settled against, 19's included; 30 sits beside 25, the other chunk rereading AGENTS.md.

• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 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.
• 0 — Scaffold.
• 1a — The directive grammar.
• 1b — Block node syntaxes.
• 1c — Inline node syntaxes and marks.
• 1d — Corpus start.
  • 1d1 — The CommonMark subset.
  • 1d2 — Block nodes.
  • 1d3 — Inline nodes and marks.
• 2 — adfToMarkdown.
  • 2a — The runner and the CommonMark subset.
  • 2b — Block nodes.
  • 2c — Inline nodes and marks.
  • 2d — The opaque carry.
  • 2e — Carve-outs and combinations.
    • 2e1 — The carve-outs and the claimed line.
    • 2e2 — Mark runs and the runs a carry breaks.
    • 2e3 — Attribute canonicalization and the quoted value's escape.
    • 2e4 — The carry's fallback triggers.
    • 2e5 — Combined documents and the collision property.
  • 2f — The attributes CommonMark cannot hold.
• 3 — markdownToAdf.
  • 3a — The hierarchy.
  • 3b — The leaf blocks.
  • 3c — The container blocks.
  • 3d — Inline text.
  • 3e — Emphasis and links.
  • 3f — The directive grammar.
  • 3g — The node tables read backwards.
  • 3h — The block nodes.
  • 3i — The inline nodes and the marks.
  • 3j — The carry and the combinations.
  • 3k — The CommonMark spec suite.
• 4 — Round-trip property tests.
  • 4.1 — Editor-normal and the node accessors.
  • 4.2 — The ADF property.
  • 4.3 — The markdown property.
  • 4.4 — The real payloads.
• 4b — The block walk's retry (0.2.0).
• 4c — The scanning rule's remaining sites (0.2.0).
• 4d — What the gate says while it runs (0.2.0).
• 5 — Ship 0.1.0.
• 5e — The publish token's deadline. 0.1.0 published only once the npm token carried Bypass 2FA: the account requiring no 2FA on writes was not enough, and npm answered EOTP until the token itself bypassed. npm retires bypass-2FA tokens for direct publishing around January 2027, leaving them npm stage publish, which a maintainer approves with 2FA; its replacement — trusted publishing over OIDC — supports GitHub-hosted Actions, GitLab.com's shared runners and CircleCI's cloud, self-hosted runners planned without a date. So the release path has an expiry date and no drop-in successor yet. Revisit: whether npm has added Gitea or self-hosted OIDC, and otherwise whether the release moves to the staged publish — which fits badly with publish-on-merge, and is the trade to weigh rather than discover on a red release run. Settled (the maintainer, 2026-09-13): last of the known work, clear of 0.2.0, placed there knowing the cutoff may land before 0.2.0 ships.
• 5f — Publish the bundle size (0.2.0). Measure the shipped artifact and put the number in the README, kept honest by the release pipeline rather than by a human re-reading it. The quantity is what a consumer downloads and loads: the tarball npm pack produces, its unpacked dist, and the built JavaScript minified + gzipped — the figure the competitors advertise (marklassian's "12kb") and the only apple-to-apple one, since ours ships tsc's unminified output and no minifier yet (decide here whether to minify for the build or report the unminified gzip). A publish/pipeline leg measures it and fails when the README figure drifts, so the number can't rot; the figure lands in README §The package beside the "no runtime dependencies" claim. Measured today, unminified: tarball 60.4 kB, unpacked 221.5 kB, JS gzipped 45.6 kB.
• 5g — Reweight the README for the reader (0.2.0). It opens with the pre-launch rationale — Atlassian's REST APIs, pf-editor-service/convert being decommissioned, a link to JRACLOUD-77436 — where a shipped package should answer what it is, what it does and for whom first, then the shortest runnable example. Settled (the maintainer, 2026-09-13): the background goes entirely, no endpoint, ticket or "why" note left. The top follows the package-README order: an npm version badge and the Gitea Actions badge, a tagline that is also package.json's description, a feature list and a one-line table of contents, then install and the shortest runnable example; a table of everything exported sits near the bottom. The HTML directions were to stay an aside until a later release shipped them; 7 now ships in this one and reads ahead of this item, so the README documents HTML as it documents markdown, the tagline and description naming both (the maintainer, 2026-09-13, revised 2026-09-18).
• 5a — Rename to @larvit/adf-codec.
• 5b — The consumer's error surface.
  • 5b1 — The error's source position.
  • 5b2 — The error messages.
  • 5b3 — The code list and the flavour's gaps.
  • 5b4 — The README's consumer surface.
• 5c — The build and the release pipeline.
• 5d — The browser leg.
• 6 — The HTML dialect spec (0.2.0). Element-by-element mapping, the data-* fidelity scheme, the opaque-carry form, and the documented foreign-element set htmlToAdf accepts — the set markdownToAdf shares (spec/flavour.md §Raw HTML in input; 29). Settled (the maintainer, 2026-09-20), three answers that shape the set: - A container ADF has no node for unwraps to its children, its own attributes dropped, so <div align="center">text</div> keeps text and loses the box and the alignment ADF cannot hold. AGENTS.md §3's "error result naming the element" stops being the whole rule when this lands, and says so there. - <details><summary>Title</summary>…</details> is an expand, the summary its title; one inside another is a nestedExpand, as 10 already spells for the lossy pair. An empty <details> is still refused — expand requires content, so there is nothing to build. - A comment stays an error result. Neither schema holds a comment node: across 84 and 98 definitions the only "comment" in either file is annotationType: "inlineComment" on the annotation mark, which carries an id and no text, the words living behind an Atlassian API. placeholder is the editor's own visible hint, and extension demands an extensionKey naming a vendor app. Nothing can hold the words, so nothing accepts them. Open, and the reason the first answer is not yet a rule: whether unwrapping is the default for any unmapped element holding children, or only for a named transparent set. <script> and <style> hold text that is not prose, so the default leaks code into the document as visible words.
• 7 — HTML, the third format (0.2.0). adfToHtml, htmlToAdf, the composed markdownToHtml / htmlToMarkdown. CommonMark spec suite runs against markdownToHtml from here (§10). The README's tagline and package.json's description regain HTML (5g).
• 8 — CLI. A later goal, shaped around the personas once the library exists.
• 9 — The online sandbox. A web page with two textboxes converting back and forth between ADF and markdown, powered by the library's browser build.
• 10 — Lossy conversion (0.2.0). Markdown other tools render readably, to and from ADF, keeping the content while dropping what markdown cannot hold — format, design and the richer nodes. Settled (the maintainer, 2026-09-14): two exports composed around the lossless pair, so §1's four conversions stay four. adfToPlainMarkdown(doc) reduces the document ADF→ADF and hands it to adfToMarkdown; plainMarkdownToAdf(markdown) hands the markdown to markdownToAdf and lifts the result ADF→ADF. Both carry markdown conventions, so the reduction sits in src/markdown/emit/, the lift in src/markdown/parse/ and what both read in src/markdown/ (§11). The markdown is the flavour without directives — CommonMark, the pipe table and ~~ — plus the conventions below, chosen for readability from a survey of GitHub, GitLab, Gitea, Obsidian, Pandoc, MkDocs, Docusaurus, Typora, Joplin, Logseq, Bear, Notion, Azure DevOps and Discord, GitHub's renderer confirming each shape. Writing refuses only what the document guard refuses (not-an-adf-document, unsupported-document-version, unsupported-nesting-depth) and degrades every other shape; reading refuses what markdownToAdf refuses. A lifted node carries no localId. The lift also reads other tools' spellings — type words in any case, Obsidian's aliases, [X] — since it reads their output and never writes those spellings. - A panel is an alert: the marker alone on the quote's first line, a blank >, then the body (> [!WARNING]), in GitHub's five words by colour — info NOTE, note IMPORTANT, tip and success TIP, warning WARNING, error CAUTION, custom NOTE. The lift reads those words back (NOTE info, IMPORTANT note, TIP tip, WARNING warning, CAUTION error) and Obsidian's by meaning (hint tip; success, check and done success; attention warning; danger, failure, fail, missing and bug error; any other word info). Text after a marker in its paragraph is the panel's first body paragraph. - An expand or nestedExpand is Obsidian's folded callout, > [!NOTE]- Title, a blank >, then the body. The lift reads a fold sign (- or +) as an expand whatever the word, the rest of the marker's paragraph as its title, and an expand inside an expand as a nestedExpand. - A taskList is a bullet list whose items lead with [x] or [ ] (- [x] Write the spec). The lift reads a list whose every item is so marked back as a taskList — a blockTaskItem where an item holds more than one block, a nested task list moved beside its item — and leaves mixed and ordered lists plain. A decisionList is a plain bullet list. - backgroundColor is ==text==, and the lift gives ==text== the Atlassian editor's default highlight colour. - layoutSection/layoutColumn, bodiedExtension, bodiedSyncBlock, multiBodiedExtension and extensionFrame unwrap to their body blocks in order; the CommonMark blocks keep their spelling, attributes dropped. - mention and status become their text, the mention's @ kept; emoji its text or else its shortName; date its ISO date in UTC (2026-09-13); inlineCard, blockCard and embedCard a link to their url, dropped when they carry only data; a mediaSingle holding an external image stays ![alt](url); media, mediaGroup and mediaInline their alt text or nothing; caption its text as a paragraph; extension, inlineExtension and syncBlock their text attribute or nothing; placeholder nothing; a node no row names, or one standing where no spelling holds it, its blocks or its text. - A table stays a pipe table: the first row becomes the header, a cell's blocks join on one line with spaces, and spans and the cells they cover drop. - code, em, link, strike and strong stay and every other mark drops, keeping its text — subsup too, since ~2~ is a strike on GitHub; a link no CommonMark escape writes becomes its text, and a mark run CommonMark's flanking or matching cannot spell drops its mark. - A newline in text becomes a hard break and edge whitespace is trimmed; carriage returns and null characters are removed; a paragraph line opening with a code span whose backticks would read as a fence loses the code mark; an empty paragraph drops, and adjacent lists of one type merge. - Rejected in the survey: ~sub~ and ^sup^, underline and colour spellings, raw HTML (<details>, <mark>), MkDocs !!! and the ::: admonition family, footnotes, definition lists, wikilinks, embeds, tags, comments, TOC tokens, spoilers, task states past [x]/[ ], and lifting bare URLs, @name, :shortcode: or ISO dates into nodes.
  • 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.
  • 10c — The exports. adfToPlainMarkdown and plainMarkdownToAdf exported with their README sections, and two properties over 4.2's generators: writing refuses only the guard's codes, and markdown adfToPlainMarkdown wrote reads back through plainMarkdownToAdf and writes again byte for byte. AGENTS.md §1 records the pair as composed around the lossless one.
• 11 — Atlassian's ADF schema as the tables' truth.
  • 11a — The vendored schema.
  • 11b — The gate.
• 12 — The !adf: re-spelling.
  • 12a — The spec and the decision.
  • 12b — The inline form.
  • 12c — The block form.
  • 12d — The README, MIGRATION.md and the sweep.
• 13 — The schema's gap attributes (0.2.0).
  • 13a — rule and layoutSection.
  • 13b — The directive link.
• 14 — The CommonMark subset's directory (0.2.0).
• 15 — The href-less directive link (0.2.0).
• 16 — The link wrapping a link (0.2.0).
• 17 — A machine-enforced size ratchet (0.2.0).
• 18 — The subtree the directive spelling asks about (0.2.0).
• 28 — emitLine's retry loop cannot spin (0.2.0).

The ADF inventory to cover

From Atlassian's structure reference — not the whole schema: real payloads also carry taskList/taskItem, decisionList/decisionItem, layoutSection/layoutColumn, blockCard/embedCard, extension/bodiedExtension/inlineExtension and placeholder, none documented there. The documented set is the floor: the floor gets designed syntax, the rest rides the opaque carry (§3) until it does too.

Top-level block	blockquote bodiedSyncBlock bulletList codeBlock expand heading mediaGroup mediaSingle multiBodiedExtension orderedList panel paragraph rule syncBlock table
Child block	blockTaskItem extensionFrame listItem media nestedExpand tableCell tableHeader tableRow
Inline	date emoji hardBreak inlineCard mediaInline mention status text
Marks	border code em link strike strong subsup textColor underline

Plain markdown covers blockquote, bulletList, codeBlock, heading, orderedList, paragraph, rule, listItem, hardBreak, text, and the code, em, link and strong marks; strike is the flavour's ~~ carve-out. Everything else is what the flavour is for.