chore: Re-organized todos.

To reduce overwhelm.
This commit is contained in:
Spencer Brower
2026-08-06 10:43:19 -04:00
parent d1179f12ec
commit e019554557
4 changed files with 235 additions and 311 deletions
+38
View File
@@ -0,0 +1,38 @@
## Code Review
A human should manually review every file in the project. AI cannot complete
these tasks.
- [ ] Review every file in thor
- [ ] Review assets.odin
- [ ] Review content.odin
- [ ] Review defaults.odin
- [ ] Review feed.odin
- [ ] Review frontmatter.odin
- [ ] Review main.odin
- [ ] Review minify.odin
- [ ] Review `markdown/`
- [x] Review alerts.odin
- [x] Review alerts_test.odin
- [x] Review emoji.odin
- [x] Review emoji_test.odin
- [ ] Review footnotes.odin
- [ ] Review footnotes_test.odin
- [ ] Review deflists.odin
- [ ] Review deflists_test.odin
- [ ] Review highlight.odin
- [ ] Review markdown.odin
- [x] Review sectionate.odin
- [x] Review sectionate_test.odin
- [ ] Review suggest.odin
- [ ] Review suggest_test.odin
- [ ] Review `diagnostics/`
- [ ] Review diagnostics.odin
- [ ] Review diagnostics_test.odin
- [x] Review opengraph.odin
- [ ] Review render.odin
- [x] Review site.odin
- [ ] Review treesitter/treesitter.odin
- [ ] Review vfs.odin
- [ ] Review procs
- [ ] markdown.transform_alert
+8 -51
View File
@@ -1,54 +1,11 @@
# Hugo vs Thor: Why Thor Doesn't Need Shortcode Context Isolation
# Hugo Notes
## Hugo's Shortcode Isolation
from the [Hugo docs](https://gohugo.io/quick-reference/glossary/#default-sort-order)
Hugo strictly separates shortcode context from layout template context. Shortcodes run during markdown rendering, before layouts. They get a limited context (`.Page`, `.Site`, `.Params`), not the full layout rendering state.
default sort order
### Why Hugo does this
1. **Circular dependencies** — Hugo shortcodes output into `.Content`. Layouts read `.Content` and compute derived properties (`.TableOfContents`, `.WordCount`, `.ReadingTime`). If shortcodes could access these computed properties, you'd get circular dependencies (shortcode output → computed property → shortcode reads it).
2. **Multiple layouts/themes** — Hugo supports user-selectable themes and multiple layouts per section type. A shortcode must produce identical output regardless of which layout renders it. Isolating context guarantees portability.
3. **Shared mutable page object** — Hugo builds a rich `Page` object that both shortcodes and layouts access. Context isolation prevents shortcodes from mutating layout-visible state.
4. **Security** — Shortcodes live in content files (potentially user-authored). Full template access would let content authors inject arbitrary logic, access sensitive config, or break site structure.
5. **Caching** — Hugo caches shortcode output independently. Predictable context = predictable cache behavior.
## Why this doesn't apply to thor
- **No computed page properties** — The content Mustache pass runs on raw markdown (before cmark). The layout Mustache pass runs on finished HTML (after cmark). There's no shared `.Content` object or derived properties. The two passes are at different pipeline stages with no shared mutable state.
- **Single layout set** — Thor has one set of templates (`base.html` + page templates). No theme switching, no multiple layout variants per section. Portability across layouts isn't a concern.
- **Explicit data passing** — Each `mustache.render` call receives its own data struct. Content pass gets frontmatter + params. Layout pass gets page data + site config. No shared mutable object between them.
- **Single author** — Content is authored by the site owner. No untrusted user-generated content.
- **No caching** — Thor rebuilds from scratch every time. No cache invalidation concerns.
## Thor's approach
Thor runs Mustache on content **before** cmark as a pre-processing step, then runs Mustache on layout templates **after** cmark as a post-processing step. These are two independent `mustache.render` calls:
```
content (markdown)
→ mustache.render(content, content_data, partials) ← content pass
→ cmark markdown_to_html
→ inject_sidenotes / inject_alerts / highlight_code
→ mustache.render_in_layout(template, page_data, layout, partials) ← layout pass
→ final HTML
```
The content pass can use Mustache variables, partials (`{{> ./file}}`), sections, and conditionals freely. The layout pass wraps the rendered content in the page chrome. They share nothing except the data we explicitly choose to pass.
## When this WOULD matter for thor
If thor ever adds:
- Computed page properties (table of contents, reading time, word count)
- Multiple selectable themes/layouts
- User-generated content / multi-author support
- Partial template caching
...then context isolation between the content and layout passes would become relevant. Until then, it's unnecessary complexity.
: The default sort order for page collections, used when no other criteria are set, follows this priority:
1. weight (ascending)
2. date (descending)
3. linkTitle falling back to title (ascending)
4. logical path (ascending)
+129
View File
@@ -0,0 +1,129 @@
## v0.2.0
- [ ] [aliases](https://gohugo.io/methods/page/aliases/#redirects)?
- [ ] consider sites with date based urls.
- [ ] Create a json schema file for `thor.json`.
- [ ] Content-hash fingerprinting for CSS and JS cache busting
- [ ] Mount content in VFS
- [ ] `-watch` flag
- [x] basic poll loop
- [ ] filesystem poll loop
- [ ] event based
- [ ] if no `html` tag detected in output, re-render output with base template
(or whatever template is next in the chain). (Hugo seems to do this)
- [ ] Markdown
- [ ] Block attributes extension — hello-world.md
- [ ] add tables extension
- [ ] enable template level rendering of TOCs
- [ ] Write css for toc sidebar and figure out where to put it.
- [ ] Add [hugo style configuration](https://gohugo.io/configuration/markup/#table-of-contents)
- [ ] implicit page titles (Set when missing?)
- guess from h1 tag or first h2 tag?
- [ ] Syntax highlighting
- [ ] grammars
- [ ] load grammars dynamically
- [ ] Search in multiple places
- [ ] Download missing grammars.
- [ ] Durable highlight paths: read `GRAPHS_PATH`/`QUERIES_PATH` from env vars set by the flake instead of hardcoded nix store hashes, so they survive `nix flake update` and let the grammar/query version-mismatch detector fire automatically.
- [ ] Theme selector
- use http://github.com/helix-editor/helix/tree/master/runtime/themes) as a
guide
- [ ] Unit tests for highlighting helpers: `capture_name_to_css`, `escape_html`, `unescape_html`, `extract_query_token`, `helix_version_from_path`.
- [ ] `<pre><code>` blocks need to set background to theme background,
regardless of prefers-dark. (or use a different theme)
- [ ] Avoid `json.Value` / `json.Object` where possible.
- [ ] come up with scrapers / scrape sources to harvest site data
- we'll use this to help us sculpt defaults.
- [ ] merge `render_{section,home_html,page_html}` procs.
- [ ] try to combine render_page_html and render_home_html?
- [ ] make `parse` an overload of `parse_text/parse_inline` and `parse_file`, or something.
### Menus
- [ ] Nav items should be active when the current page is selected.
- [ ] configure opt-out of automatic sections being added to menu.
- [ ] nested menus (i.e. `parent` support)
### Strict Mode
- [ ] opt-in "strict_keys" mode. in this mode, key lookups may not view parent objects.
- [ ] warn/error when unknown key used in mustache.
- [ ] enforce heading structure.
- [ ] Either frontmatter.title set, or 1 h1 tag at top, not both
- [ ] No skipping.
### Dates
- [ ] display an error when no part of the date appears in the output.
- [ ] Handle 0 and whitespace padding i.e. "_2" -> " 2"
- [ ] Do we *need* mustache.Date_Components, or can we use core:time/datetime.DateTime?
## Beyond
- [ ] consider logging the number of times an error occurred. (diagnostics)
- [ ] Need to be careful about diagnostics across module boundaries.
- we don't necessarily want to warn users about theme designers mistakes. (though perhaps we do)
- [ ] consider reporting duplicate weights outside of menus (diagnostic)
- [ ] Do mustache's whitespace rules actually suit us, or should we make our own?
- [ ] create a default `head.html`?
- [ ] improve home link customization.
- [ ] currently an accessibility issue.
- [ ] get rid of the global variables in the `treesitter` package.
- [ ] site-wide Integrity hash
- Allows users to verify their output didn't change after upgrading to a new version
- [ ] Add `-production` flag
- sets `-minify`
- [ ] manually pass `site_allocator` to `load_page`
- [ ] Import/export packages. Hugo, jekyll, WordPress, etc.
- [ ] shortcodes?
- include-code shortcode (`{{< include-code ... >}}`) — i-ported-fd-to-odin
- [ ] follow symlinks in `scan_content`?
- [ ] Add overloads for every md extension? i.e. accept ^strings.Builder.
- [ ] content link checker?
- Checks all links on each page to make sure they are valid.
- [ ] Peruse [GitHub's](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax#alerts)
docs for any juicy nuggets we may have missed.
- [ ] Avoid using `render_inline_md` if possible.
- [ ] don't use bullshit "sub-tokens", add filters and pipes as proper tokens.
- [ ] Create a Location struct that somewhat matches Odin's [Source_Code_Location](https://pkg.odin-lang.org/base/runtime/#Source_Code_Location)?
- Note that odin's version doesn't contain the stack trace.
### Memory Management
- [ ] Leaks in highlighter code.
- [ ] Not sure whether to use temp allocator or site_allocator in opengraph.odin.
- [ ] Not sure whether to use temp allocator or site_allocator in `site_load_content`.
- [ ] Might not need to allocate in `strip_html_tags`
- [ ] Fix `apply_filter`'s `format` case (`mustache/pipes.odin`) boxing `apply_format`'s
`string` result into `any` via bare `return`, which materializes a hidden
header temp in `apply_filter`'s own stack frame. Dangling once the frame
returns; caused the `-o:speed` segfault in `write_value`. Fix: box explicitly
with `any{new_clone(formatted, context.temp_allocator), typeid_of(string)}`.
- [ ] Same pattern in `apply_group_by` (`mustache/pipes.odin`): `return groups, nil`
boxes a freshly-built `[dynamic]Group` as bare `any` — same latent
stack-temp UB, hasn't crashed yet but should get the same treatment.
### Performance
- [ ] See if we can disable bounds checks in `write_indented` and elsewhere.
- [ ] Instead of loading the site fresh each time in watch mode, create a
`reload_site` proc, that just updates changed resources.
- [ ] Only publish referenced assets.
- [ ] Split `load_page` into frontmatter-parse + body-process phases so draft pages can skip the markdown pipeline entirely
- [ ] Use spall to find ways to reduce run time.
- [ ] too many `write_string` calls in `highlight_block`
- [ ] return `src: cstring` from `load_query`.
- [ ] Improve `unescape_html` with simd.
- [ ] generate summary before syntax highlighting.
- [ ] generate summary before markdown to html conversion.
- [ ] mount_recursive is pretty significant
- [ ] thread pool for grammar loading is unbounded.
- [ ] load grammars async.
- [ ] during `load_page`:
- pass each code block to the treesitter queue
- continue working on the page,
- `await` the highlighted code.
- [ ] can markdown extensions run in parallel?
- [ ] enforce MAX_SLUG_LENGTH
- [ ] enforce MAX_CONTEXT_DEPTH
- [ ] ensure struct fields are ordered correctly
+67 -267
View File
@@ -1,277 +1,77 @@
## High priority
# TODOs
- Polish existing features before moving on to new ones.
- [ ] Do mustache's whitespace rules actually suit us, or should we make our own?
- [x] `{{>title}}` default partial? `{{site.title}} | {{ page.title }}`
- [ ] implicit titles (Set when missing?)
- [ ] create a default `head.html`.
- [ ] html comments are rendered, except in minify mode.
## General
Remember to polish existing features before moving on to new ones.
- [ ] `build` command alias of default
- [ ] `new site` command to set up new project
- [ ] Some render errors blank the entire page instead of scoping to the failing tag.
- [ ] We must remove all mention of `posts` from the odin code.
At present, "posts" are a user-level construct defined as pages in a
particular collection.
- [ ] ? Clean up the default layouts (no idea what this means)
- [ ] `group_by` currently requires a computed `year` field on the page.
- We should replace this with `{{ pages | group_by (date | "2006") }}` or similar
- [ ] does it make sense for partials to be inside layouts?
- current: `layouts/partials`
- alt1: `layouts`, `partials`
- alt2: `templates/layouts`, `templates/partials`
- [ ] [aliases](https://gohugo.io/methods/page/aliases/#redirects)?
- [ ] Improve diagnostics
- [x] keep track of every error and don't report them more than once.
- [x] show parsed arg when -extension is unrecognized
- [x] also do typo detection?
- [ ] show where the typo occurred (config file path for `apply_extension_config`, full `-ext:value` for `parse_extension_list`)
- [ ] `*` make sure the frontmatter parser has good diagnostics.
- [ ] fix the diagnostics in [DIAGNOSTIC TODOS](./DIAGNOSTIC_TODOS.yaml)
- [ ] only report format errors once.
- [ ] only report missing partiall errors once.
- [ ] All Diagnostics should show:
- [ ] *What* went wrong
- [ ] *where* (in the file)
- [ ] *where* (in the stack trace)
- [ ] *how* you can fix it (if applicable)
- [ ] Create a Location struct that somewhat matches Odin's [Source_Code_Location](https://pkg.odin-lang.org/base/runtime/#Source_Code_Location)?
- Note that odin's version doesn't contain the stack trace.
- [ ] show "stack traces" in template error diagnostics
- [ ] better diagnostics for syntax errors in treesitter.
- [ ] Ensure diagnostics for MAX_CONTEXT_DEPTH are good.
- [ ] improve matching weights message.
- [ ] Test menu diagnostics
- [ ] Honestly, Test **all** diagnostics
- [ ] Need to be careful about diagnostics across module boundaries.
- we don't necessarily want to warn users about theme designers mistakes. (though perhaps we do)
- [ ] consider reporting duplicate weights outside of menus
- [ ] Extend `tag_error` to all render-time errors, not just pipe errors.
- [ ] pipes
- [ ] first
- [ ] last
- [ ] sort_by
## Documentation
- [ ] talk about the context stack (and its limit).
- [ ] highlight the differences in the way menus are handled.
- [ ] explain that html/css/js comments are rendered, except in minify mode.
## Investigate
- [ ] running ./thor/thor still logs the debug message: using config /home/spencer/github.com/sbrow.github.io/thor.json
- wrong cwd?
## Diagnostics
All diagnostics (errors) should show
- *What* went wrong
- *where* (in the file)
- *where* (in the stack trace)
- *how* you can fix it (if applicable)
- [ ] try to make file paths clickable links.
- [ ] make sure the frontmatter parser has good diagnostics.
- [ ] add all known diagnostics to [DIAGNOSTIC TODOS](./DIAGNOSTIC_TODOS.yaml).
- [ ] fix the broken diagnostics in [DIAGNOSTIC TODOS](./DIAGNOSTIC_TODOS.yaml)
- [ ] show *stack trace* for template partials?
- [ ] show where the typo occurred (config file path for `apply_extension_config`, full `-ext:value` for `parse_extension_list`)
- [ ] **all** diagnostics in [DIAGNOSTIC TODOS](./DIAGNOSTIC_TODOS.yaml) must be tested.
- [ ] only report format errors once.
- [ ] only report missing partial errors once.
- [ ] better diagnostics for syntax errors in treesitter.
- [ ] Ensure diagnostics for MAX_CONTEXT_DEPTH are good.
- [ ] improve matching weights message.
- [ ] Test menu diagnostics
- [ ] Extend `tag_error` to all render-time errors, not just pipe errors.
Currently only pipe errors (4 sites in `render_nodes`) get stamped with
the correct template source/path. Other render errors still use the
content template's source/path, which can point at the wrong file.
- [ ] try to make file paths clickable links.
- [ ] centralize diagnostics to one place.
- [ ] consider logging the number of times an error occurred.
- [ ] Load grammars dynamically
- [x] starred must be a param.
- [ ] Documentation
- [ ] talk about the context stack (and its limit).
- [ ] highlight the differences in the way menus are handled.
- [ ] consider sites with date based urls.
- [x] Don't show annoying log output in tests.
- [ ] improve home link customization.
- [ ] currently an accessibility issue.
- [x] support JSON5 in in frontmatter
- [ ] Create a json schema file for `thor.json`.
- [x] cleanup `#partial switch`es.
- [ ] improve json diagnostics.
- i.e. "Missing quotes around string", etc.
- [ ] don't use bullshit "sub-tokens", add filters and pipes as proper tokens.
## Analytics
Debug log stats:
- [ ] final Context_Stack cap
- [ ] highest PIPE args used
- [ ] longest slug length + name that generated it
- [ ] number of pages
- [ ] number of blocks
- [ ] enabled features / extensions
- [ ] etc
## sbrow.github.io
- [ ] Ideas is now in the wrong spot. Date is wrong, and it is showing date
when it shouldn't be.
## Performance
- [ ] See if we can disable bounds checks in `write_indented` and elsewhere.
- [ ] Instead of loading the site fresh each time in watch mode, create a
`reload_site` proc, that just updates changed resources.
- [ ] Only publish referenced assets.
- [ ] Split `load_page` into frontmatter-parse + body-process phases so draft pages can skip the markdown pipeline entirely
- [ ] Use spall to find ways to reduce run time.
- [ ] too many `write_string` calls in `highlight_block`
- [ ] return `src: cstring` from `load_query`.
- [ ] Improve `unescape_html` with simd.
- [ ] generate summary before syntax highlighting.
- [ ] generate summary before markdown to html conversion.
- [ ] mount_recursive is pretty significant
- [ ] thread pool for grammar loading is unbounded.
- [ ] load grammars async.
- [ ] during `load_page`:
- pass each code block to the treesitter queue
- continue working on the page,
- `await` the highlighted code.
- [ ] can markdown extensions run in parallel?
- [ ] enforce MAX_SLUG_LENGTH
- [ ] enforce MAX_CONTEXT_DEPTH
- [ ] ensure struct fields are ordered correctly
## Remove Privileged content
- [ ] `group_by` currently requires a computed `year` field on the page.
- We should replace this with `{{ pages | group_by (date | "2006") }}` or similar
## Memory Management
- [ ] Leaks in highlighter code.
- [ ] Not sure whether to use temp allocator or site_allocator in opengraph.odin.
- [ ] Not sure whether to use temp allocator or site_allocator in `site_load_content`.
- [ ] Might not need to allocate in `strip_html_tags`
- [ ] Fix `apply_filter`'s `format` case (`mustache/pipes.odin`) boxing `apply_format`'s
`string` result into `any` via bare `return`, which materializes a hidden
header temp in `apply_filter`'s own stack frame. Dangling once the frame
returns; caused the `-o:speed` segfault in `write_value`. Fix: box explicitly
with `any{new_clone(formatted, context.temp_allocator), typeid_of(string)}`.
- [ ] Same pattern in `apply_group_by` (`mustache/pipes.odin`): `return groups, nil`
boxes a freshly-built `[dynamic]Group` as bare `any` — same latent
stack-temp UB, hasn't crashed yet but should get the same treatment.
## Markdown
- [ ] Add overloads for every extension - accept ^strings.Builder.
- [x] Add conventional (Hugo style) footnotes option.
- [x] Add opt-in deflist support.
- [x] Decide if lambdas actually provide any value.
- [ ] add tables extension
- [x] Table of contents support.
- [ ] enable template level rendering of TOCs
- [ ] Write css for toc sidebar and figure out where to put it.
- [ ] Add [hugo style configuration](https://gohugo.io/configuration/markup/#table-of-contents)
- [ ] Link checker?
- Checks all links on each page to make sure they are valid.
- [ ] Peruse [GitHub's](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax#alerts)
docs for any juicy nuggets we may have missed.
- [ ] Avoid using `render_inline_md` if possible.
## Dates
- [ ] display an error when no part of the date appears in the output.
- [ ] Handle 0 and whitespace padding i.e. "_2" -> " 2"
- [ ] Do we *need* mustache.Date_Components, or can we use core:time/datetime.DateTime?
## General
- [ ] Menus
- [ ] configure opt-out of automatic sections being added to menu.
- [ ] nested menus (i.e. `parent` support)
- [ ] get rid of the global variables in the `treesitter` package.
- [ ] enforce heading structure.
- [ ] Either frontmatter.title set, or 1 h1 tag at top, not both
- [ ] No skipping.
- [ ] Consider using `or_else` when applying default values to structs. i.e.
```odin
package main
X :: struct {
foo: string
}
main :: proc () {
x: X
x.foo = x.foo or_else "bar"
}
```
- [ ] Integrity hash
- Allows users to verify their output didn't change after upgrading to a new version
- [ ] Content-hash fingerprinting for CSS and JS cache busting
- [ ] come up with scrapers / scrape sources to harvest site data
- we'll use this to help us sculpt defaults.
- [ ] merge `render_{section,home_html,page_html}` procs.
- [ ] try to combine render_page_html and render_home_html?
- [ ] Debug log stats. (analytics)
- [ ] final Context_Stack cap
- [ ] highest PIPE args used
- [ ] longest slug length + name that generated it
- [ ] number of pages
- [ ] number of blocks
- [ ] enabled features / extensions
- [ ] etc
- [ ] Avoid `json.Value` / `json.Object` where possible.
- [ ] make `parse` an overload of `parse_text/parse_inline` and `parse_file`, or something.
- [x] Add page params
- [ ] We must remove all mention of `posts` from the odin code.
At present, "posts" are a user-level construct defined as pages in a
particular collection.
- [ ] running ./thor/thor still logs the debug message: using config /home/spencer/github.com/sbrow.github.io/thor.json
- wrong cwd?
- [ ] Clean up the default layouts
- [ ] Menus
- [x] Detailed frontmatter menu form ("menu": {"main": {"weight": 5}})
- [ ] Menu active state (pre-compute is_active based on page.permalink prefix match)
- [x] Page.weight field for general-purpose page ordering (menus, lists, related posts)
- [ ] if no `html` tag detected in output, re-render output with base template
(or whatever template is next in the chain)
- [ ] Add `-production` flag
- sets `-minify`
- [ ] Mustache diagnostics
- [ ] Partial invocation stack in diagnostics: when an error fires inside a partial, show "invoked from" chain through `{{> name}}` calls. Currently warnings inside partials point at the partial (correct file) but don't show the invocation site.
- [x] Error message doesn't show position of faulty pipe name correctly.
- [ ] `render_template` (`render.odin`) blanks the *entire page* to `""` on any
mustache render error and only `log.errorf`s it — a single bad tag/pipe
anywhere on the page silently kills the whole output with no visible
signal outside the terminal log. Should at least be scoped to the
failing tag/section, or surfaced somewhere the person building the
site will actually see it.
```bash
[ERROR] --- [138:render_template()] unknown pipe op 'formats'
--> /home/spencer/github.com/sbrow.github.io/layouts/home.html:1:1
|
1 | {{<base}}
| ^^^^^^^^^
2 | {{$content}}
3 | <main>
```
- [ ] Block attributes on code fences (`{ #ex-1 }`) — hello-world.md
- [ ] include-code shortcode (`{{< include-code ... >}}`) — i-ported-fd-to-odin
- [ ] follow symlinks in `scan_content`?
- [ ] ensure sidenote numbers render in display order and not in declaration order.
- [ ] Nav items should be active when the current page is selected.
- [ ] Theme selector for syntax highlighting.
- use http://github.com/helix-editor/helix/tree/master/runtime/themes) as a
guide
- [ ] grammars
- [ ] Search in multiple places
- [ ] Download missing grammars.
- [ ] Durable highlight paths: read `GRAPHS_PATH`/`QUERIES_PATH` from env vars set by the flake instead of hardcoded nix store hashes, so they survive `nix flake update` and let the grammar/query version-mismatch detector fire automatically.
- [ ] CI
- [ ] Syntax highlighting in production: CI (`nix build` on ubuntu-latest) has no grammar `.so`s and a machine-specific `QUERIES_PATH` nix store hash, so the deployed site renders unhighlighted. Provide grammars + queries as nix build inputs and pass paths to thor at runtime (env vars/flags).
- [ ] Unit tests for highlighting helpers: `capture_name_to_css`, `escape_html`, `unescape_html`, `extract_query_token`, `helix_version_from_path`.
- [ ] `<pre><code>` blocks need to set background to theme background,
regardless of prefers-dark. (or use a different theme)
- [ ] `-watch` flag
- [x] basic poll loop
- [ ] filesystem poll loop
- [ ] event based
- [x] Free cmark HTML output (`body_html`) — cmark allocates via C malloc, not the arena, so it leaks per iteration in watch mode
- [ ] manually pass `site_allocator` to `load_page`
- [ ] Mount content in VFS
- [ ] commands
- [ ] `build` alias of default
- [ ] `new site` set up new project
- [ ] warn/error when unknown key used in mustache.
- [ ] Import/export packages. Hugo, jekyll, WordPress, etc.
- [ ] opt-in "strict_keys" mode. in this mode, key lookups may not view parent objects.
## Notes
from the [Hugo docs](https://gohugo.io/quick-reference/glossary/#default-sort-order)
default sort order
: The default sort order for page collections, used when no other criteria are set, follows this priority:
1. weight (ascending)
2. date (descending)
3. linkTitle falling back to title (ascending)
4. logical path (ascending)
## Code Review
A human should manually review every file in the project. AI cannot complete
these tasks.
- [ ] Review every file in thor
- [ ] Review assets.odin
- [ ] Review content.odin
- [ ] Review defaults.odin
- [ ] Review feed.odin
- [ ] Review frontmatter.odin
- [ ] Review main.odin
- [ ] Review minify.odin
- [ ] Review `markdown/`
- [x] Review alerts.odin
- [x] Review alerts_test.odin
- [x] Review emoji.odin
- [x] Review emoji_test.odin
- [ ] Review footnotes.odin
- [ ] Review footnotes_test.odin
- [ ] Review highlight.odin
- [ ] Review markdown.odin
- [x] Review sectionate.odin
- [x] Review sectionate_test.odin
- [ ] Review suggest.odin
- [ ] Review suggest_test.odin
- [x] Review opengraph.odin
- [ ] Review render.odin
- [x] Review site.odin
- [ ] Review treesitter/treesitter.odin
- [ ] Review vfs.odin
- [ ] Review procs
- [ ] markdown.transform_alert