Nested + foldable Obsidian callouts

[srid]

The remaining two items in #465 are now in. > [!type]+ and > [!type]- give you foldable callouts (initially expanded and collapsed, respectively); nested callouts work by indenting an inner blockquote with > >. Closes the last two unchecked boxes on the issue.

Foldable callouts render as native <details>/<summary> so the toggle works without JavaScript, and the chevron rotates via a CSS [open] selector. Nested callouts mostly came for free — the recursive pandocSplice already re-applies the block renderer to inner blocks, so the parser only needed to keep the inner BlockQuote intact in Callout.body. The visible "feature" there is really the docs section plus a CSS rule that drops the trailing margin on the last inner callout so it hugs the parent body.

The header parser is split into parseCalloutHeader :: Text -> Maybe (CalloutType, Maybe FoldState) returning both pieces; parseCalloutType is preserved as a back-compat shim for the existing tests. A defensive absorbFoldSuffix merges adjacent Str "+"/Str "-" tokens onto the header in case Pandoc tokenises [!tip]+ as two Str nodes; the comment spells out why we deliberately do not absorb across Space.

Reviewed by /hickey and /lowy. One comment-clarification fix landed (commit 007c722c) and one CSS dead-selector fix from /code-police's fact-check pass (e3497610). Two structural items deferred — see the Hickey/Lowy Analysis PR comment for the disposition table.

Deferred: the three-branch template duplicates the <details> wrapper between expanded/collapsed (Heist doesn't expose conditional attribute presence cleanly without losing access to template-level color/icon binds), and the splice/template split for fold state isn't compiler-enforced for exhaustiveness. Both are quality-of-life refactors — neither is structurally wrong.

Try it locally

nix run github:srid/emanote/dizzy-king -- -L docs run

Then visit /callout to see the new Foldable callouts and Nested callouts sections render.

[srid]

Hickey/Lowy Analysis

#	Lens	Finding	Disposition
1	Hickey	Unvalidated callout type can crash via lookupHtmlTemplateMust	No-op — predates this PR; an allowlist would break the documented custom-callout extension point (#573)
2	Hickey	absorbFoldSuffix doesn't handle Space, Str "-" form	No-op — absorbing across Space would mis-parse > [!tip] - (a tip with a hyphen-prefixed title) as foldable; comment now spells this out
3	Hickey	Three template branches duplicate ~90% of the wrapper structure	Deferred — Heist doesn't support conditional attribute presence in ${...} interpolation, and a Haskell-side wrapper splice loses access to template-level color/icon binds
4	Hickey	FoldState exhaustiveness isn't enforced across Haskell+template	Deferred — same shape as #3; would require consolidating the three-way dispatch into a single Haskell case
5	Lowy	absorbFoldSuffix encoded an undocumented Pandoc tokenisation assumption	Fixed in this PR (007c722)
6	Lowy	Maybe FoldState slightly indirect vs. flat ADT	No-op — correct in context; flattening is a stylistic preference, not a boundary issue
7	Lowy	callout:fold-state data attribute as a customiser hook	No-op — Lowy endorsed this as the right seam; data-callout-fold is a stable string handle that degrades gracefully if a new state is added
8	Lowy	Nested-callout CSS selector tightly coupled to DOM shape	Fixed in this PR (e349761) — dead second selector removed; the surviving one matches both foldable and non-foldable parents
9	Lowy	Custom callout types extension point unaffected by changes	No-op — type templates just <apply template="_base">, so the new conditionals inherit automatically

Hickey rationale

The three-branch template — <div> for non-foldable, <details open> for foldable+expanded, <details> for foldable+collapsed — is structurally justified: these are three distinct HTML elements with different semantics. The duplication cost is the price of branching at the template layer rather than at the Haskell layer; the alternative (always-<details> plus CSS to disable the toggle) would complect the DOM with a presentational workaround and obscure the accessibility intent. The shared content (callout-icon-svg, callout-fold-marker, callout-body-block) is already factored via Heist <bind>. Pushing dedup further is a quality-of-life cleanup rather than a structural correction.

Lowy rationale

Three volatility axes are encapsulated independently: parsing (parseCalloutHeader + absorbFoldSuffix), render orchestration (the three conditional Heist splices), and presentation (CSS in styles.tpl). Each can change without ripple — if Obsidian extends the fold-suffix grammar, only the parser changes; if a new fold state is added, only the orchestration plus a CSS rule. The custom-callout extension point (#573) is unaffected because individual type templates delegate to _base.tpl without knowing about fold state. The callout:fold-state data attribute is the right customiser seam — a stable string handle that downstream CSS/JS can hook into without touching Haskell internals.

[srid]

/do results

Step	Status	Duration	Verification
sync	✓	0s	git fetch ok; forge=github
research	✓	6m 10s	Mapped callout implementation; identified 4 files to change
branch	✓	16s	On feature branch dizzy-king
implement	✓	9m 20s	Parser + Callout.foldState + 3-branch template + CSS; nested verified end-to-end via emanote gen
check	✓	5s	cabal build all clean
docs	✓	27s	callout.md + CHANGELOG.md updated
fmt	✓	30s	cabal-fmt, fourmolu, hlint, nixpkgs-fmt all clean
commit	✓	31s	887a62b (6 files, +219/-26) pushed
hickey+lowy	✓	6m 15s	Lowy 1 fixed (007c722); Hickey 3+4 deferred; rest no-op
police	✓	17m 31s	Fact-check: dead CSS selector fixed (e349761). Elegance: whenSplice → Heist.Splices.ifISplice (0031e5b)
test	✓	30s	cabal test: 32 examples 0 failures (incl. 9 new callout parser tests)
create-pr	✓	1m 50s	Draft PR #652 + Hickey/Lowy disposition table comment
ci	✓	4m 5s	e2e (live) 4m20s, e2e (static) 4m11s, flake-parts-docs 1m49s — all green on HEAD 0031e5b0
done	✓	0s	Workflow complete
Total		48m 3s

Optimization suggestions

• police dominates at 36% of total (17m 31s). The bulk was the /simplify Phase 2 spawning three reviewer agents in parallel, then code-search inside each — for a diff this small, a single inline pass would have been faster than the multi-agent fan-out.
• research (6m 10s) and hickey+lowy (6m 15s) ran similar lengths. Combined they're 26%. Pre-reading Emanote/Pandoc/Renderer/Callout.hs and the _base.tpl callout template before invoking /do would shave the research step substantially.
• implement (9m 20s) included one fmt-induced rebuild and an end-to-end render verification via emanote gen. The render check was worth the time (caught zero issues here, but would have caught template-XML errors immediately).
• CI (4m 5s) was the unavoidable e2e wait. For follow-up commits to this PR, --from ci-only skips everything else and just re-runs CI against the new HEAD.

Workflow completed at 2026-04-25T00:25:33Z.