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, which sent this session at 14 after it had shipped.
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, 17, 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 complexity guardrail under the largest body of new code, and 5f and 5g read last because 7 is what changes the bundle size and the tagline.

• 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). 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.
• 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.
• 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). 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 guardrail (0.2.0). Add a per-function complexity check to the gate — branch count or size — so the fits-in-your-head guardrail fails the build rather than waiting for a review to catch it (the systems-architect, 2026-09-16). It reads ahead of 6, 7 and 10 so the largest body of new code is written under it, which is also what decides the threshold: today's worst is readDirectiveContent, 27 lines and about 12 decision points over four concerns in one loop — escape, code span, nested directive, bracket balance — which 4c left half-split and this item either passes or forces apart (the systems-architect and the maintainer, 2026-09-18).
• 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).

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.