Folding the 24 `state:` commits into the code they described gave every later
commit a new hash, so the verified-against line named 8a2ef08 — a commit only
the backup refs still hold. It is 8900aab in this history. Content unchanged;
the old sha resolved only because backup/pre-fold and origin/main still exist,
which is exactly the kind of advisory that would rot into a lie.
Adopts ideas/token-conservation.md, parked 2026-07-28, plus the disciplines
the human added: read the compressed form first, discover by mechanism,
shrink output at the source, never pay twice for the same bytes.
docs/context-economy.md owns all of it and leads with a floor, because
every cheap failure mode is also a token saving — skipping the owning doc,
guessing a signature, reporting from a diff, thinning a test — and each has
already cost this repo a defect. Frugality is for presentation and
discovery, never for the artifact or the evidence.
Mechanical, not remembered:
- scripts/surface.sh generates docs/surface.md — every top-level
declaration with its line, 261 lines standing for 3757 of source. The
pre-commit hook regenerates and stages it, so it cannot be stale, and
verify.sh compares independently for a clone that never set
core.hooksPath. The hook refuses a commit with unstaged .go changes,
since what it generated describes the working tree, not the commit.
- verify.sh --quiet: 48 lines of gate output become 1. The hook uses it.
- CLAUDE_LOC_MAX=150, the only budget billed per turn rather than per read.
Both new gates were watched failing before being kept: a doctored
surface.md, and CLAUDE_LOC_MAX temporarily set to 5.
state.md's inventory loses its LOC column. It had already drifted on six
files (content.go 381→450, render.go 447→454, web.go 206→217, check
216→223, watch 129→137, chrome 105→110) which is what a number written in
two places does; the generated file owns sizes now, the table owns purpose.
The subagent question is recorded there as the one open decision, with the
case for and against written out in the idea file.
The loop gains a seventh step. Work was being committed all session on a
verbal instruction, which means it was a property of one conversation
rather than of the harness: a fresh session would have left everything in
the working tree and reported success.
Why automatic: a commit is one `git revert` from undone, and work that
only ever existed in the working tree is not recoverable — this session
already lost uncommitted work to a `git checkout` while signing was
broken. Why announced: the safety argument covers the loss, not the
surprise, so a session that did not ask for this gets one prominent line
before the first commit and the off switch, once.
Ownership split so nothing is stated twice: conventions.md "Git" owns
what one commit contains — now one *revertible* unit rather than one
feature, which is the human's correction to the "never bundle two passes"
reading — and CLAUDE.md §4 owns when a commit happens. ADR-0052 records
both, including that no gate can check this: no script can see whether a
sentence was said.
Docs 6 files, +72/-6 lines. No code, no counters moved.
The human asked for a demo extensive enough to review by hand, tracked as files,
and kept current the way docs are. Generated filler cannot be reviewed — you cannot
read a Go function and see what a reader sees — so the generator is deleted and
`examples/demo-site/` is an ordinary site: 33 Markdown files, six pictures,
site.yaml, a template override and static files.
My reason for generating it was ADR-0011, and I had that rule wrong: it puts the
*author's site root* outside this repository, not fixtures. conventions.md already
keeps golden files in testdata/, and this is the same category one size up. Keeping
both a generator and files would have been two sources of truth, so one had to go.
Kept true by two gates rather than by good intentions. A table-driven test in
internal/web serves the directory through the real handler with one case per
feature — thirty-five of them, each naming what it proves — and verify.sh runs
`khosra check` over it and fails on anything fatal. Adding a feature now means
adding it to the demo and adding its case, and the build says so if you don't.
Extensive on purpose: eleven dated posts so a section actually paginates, a
four-chapter series so first/last are not the neighbours, a gallery with three
JPEGs and an SVG so both the resampled and the untouched paths show, a Bengali-only
bundle, a titleless status, a draft, a future date, an include, a nested extras
tree, and a typography page that demonstrates what the engine will not do.
Two expectations of mine were wrong and the demo corrected them: the site's own
list template overrides *tag* listings too, so there are no group headings there —
which turns out to be the better demonstration of ADR-0046, since the theme picking
the flat shape is exactly the point. And template literal text is not escaped, so
an apostrophe stays an apostrophe.
The demo writes a whole site root that exercises every feature: two languages with
a fallback, a series with ordered chapters, a gallery, a figure, an include,
extras, tags across sections, a slug with an alias, an undated page, a draft, a
template override, static files and site.yaml. It generates its filler rather than
copying stored files, because nothing in this repository is content (ADR-0011) — and
that makes it a test of the engine rather than a fixture: anything khosra can do
that the demo cannot express is a gap.
Two things found by generating and then serving it, which is the whole point:
`khosra check` reported the demo's own series as mixing ordered and unordered
members. It was right — the chapter bodies *described* `order: 10` while the
frontmatter never carried it. The checker caught its own author.
And `/` was a **404**. ADR-0008 leaves the root engine-owned, which is right, but
"engine-owned" was never given an answer, so a visitor to the site's own address got
nothing. The root now lists every bundle, newest first, paginated like any other
listing, and 404s only when nothing is published. A hand-written home page stays a
separate decision, recorded as such.
Verified end to end: 23 files written, 12 bundles, 12 derivatives, `check` clean,
and every URL the demo promises answers — including the alias redirecting, the draft
hidden, and the front page rendering through the site's *own* template override.
The theme was supposed to need no engine work. It needed four things, which is
exactly what the audit said would happen if the contract had gaps rather than the
theme (ADR-0046) — a reader could not reach, from any page: another section, the
tags on the page they were reading, the extras beside it, or the same page in
another language.
So the contract grew three fields, all additive: `.Sections` for navigation,
`.Tags` with each term's listing URL, and `.ExtrasURL`, empty when a bundle has
none so a theme never links a 404. Sections arrive through `Renderer.Navigation`, a
callback, because sections change when content does and a copy would go stale — the
nav updates on a rebuild along with everything else.
One flaw only visible by looking at a rendered page: the language switcher pointed
at the canonical host, because `.Alternates` went absolute for hreflang. Those are
two needs, so an Alternate now carries `.URL` (absolute, for machines) and `.Path`
(relative, for a link a person clicks).
The theme now demonstrates every field it is given, including `.First`/`.Last`,
which existed and were never rendered. Still no JavaScript, still one stylesheet.
A test that asserted a page had no sequence nav was matching the inlined
stylesheet rather than the markup, and now matches the element. That is the third
time a loose assertion has passed for the wrong reason.
Polling lives in internal/ext/watch, per the human's call to keep core under its
ceiling rather than raise it a second time — which is what ADR-0041 said a second
raise would mean. It is a poller, deletable without trace, and core stayed at
2671/2800.
A settled change calls the same `rebuilder` that startup calls, because a reload
path that differs from the startup path is a reload path that drifts. The index is
an atomic.Pointer swapped whole, so a request reads the site that was current when
it arrived instead of one being rebuilt underneath it — the alternative, mutating in
place, is a data race with every in-flight request.
Names, sizes and modification times, not contents: reading every file to detect a
change costs more than the rebuild it triggers. Editor droppings are excluded,
because saving in vim writes a swap file, a backup and the number 4913, and each
would otherwise look like a change. A change must hold still for a moment first,
since one save is often several operations.
Verified against the running binary: a page 404s, the file appears, and five seconds
later it serves — one "site root changed" in the log. Then three droppings written
at once produced no rebuild at all.
Two warnings fired and were fixed rather than silenced: `runServe` gave up the
rebuild closure to `rebuilder`, and the fingerprint walk gave up its body to
`record`, where three exclusions read as a list instead of as nesting.
The Dockerfile ships the binary alone. The site root arrives as a volume and is
never copied in — it is somebody's content repository with its own history
(ADR-0011), so the image is the same for every site.
Re-adopts the parked extras entry as ADR-0047: `extras/` inside a bundle is skipped
by the scanner entirely, so a `.md` in there is an asset with no identity and no URL
of its own. The engine enumerates the tree, sorts it by path, classifies by
extension, renders markdown and text, and offers everything else as bytes. One route
with two behaviours — `…/extras/{path}` selects an entry, `?raw` returns the file.
Almost everything it needed already existed, which is the sign the model was right:
the scanner had a directory exclusion, `Assets()` knew which bundles own a
directory, and `Lookup` already decided visibility — so a draft hides its extras
with no new check. A test proves that, including `?raw`.
Two deviations from the parked shape, both because the shape was written before the
code. The directory name is fixed rather than a cascade key, since nothing reads a
section-level setting yet. And an entry is resolved against the *enumeration* rather
than the filesystem: not being in the listing is a stronger answer than os.Root
refusing a path, and cheaper.
Selecting is a link and a full page. No JavaScript is involved, and a
sidebar-and-pane layout is the theme's business — which is the layer rule applied
before writing the feature rather than after.
Three size warnings fired as a result and were fixed by splitting at seams, not by
sharding: render.go gave up its type declarations to view.go, which is the theme
contract in Go and nothing else; serve() split into a dispatcher and serveBundle;
resolve() gave up its language-prefix step to cutLang.
Scaffolds a **directory** bundle — the only shape that can own local files, so the
other kind would hand an author a page their pictures cannot live beside. What it
writes is a draft: title, today's date, `draft: true`. A tool that publishes the
moment it runs publishes by accident, and drafts are honoured now.
This is the first thing that writes into somebody's content directory, so it goes
through os.Root like every read does (ADR-0031), and it never overwrites: an
existing bundle is an error.
Two bugs found by running it rather than by testing it:
A key of `../escape` did not fail. It never left the site root — path.Join collapses
`..` first — but it wrote a real directory *inside* the root and outside content/,
which is not an escape and not a bundle either. Refused outright now, the same guard
the include path needed for the same reason. The test asserts what should be true —
content/ is the only thing this creates — because the weaker assertion I wrote first
would have passed.
`khosra new posts/x -site dir` silently ignored -site, because Go's flag package
stops at the first non-flag argument, and then failed complaining there was no site
root. Parsed in rounds now, so either order works.
main() crossed the function-length warning as a result, so it became a dispatch
table with runServe beside it — the warning was right about the code.
Deleting the widows feature answered one question; the human asked the general one.
So every feature built so far is now audited, with the verdicts as a table in
ADR-0046, and each remaining queue entry carries a layer note before anyone writes
code for it.
The codebase turned out to be otherwise clean, and I checked rather than remembered:
no Go file writes a tag, a class or a style. ADR-0036's rule that all markup comes
from a fragment had already forced that.
One real finding. A tag listing received only `.Groups`, so the engine had decided
that a tag listing *looks* grouped — ADR-0032's own reasoning was "so a busy term
stays readable", which is a readability judgement. It now receives both shapes: the
partition, because a template cannot group for itself, and the flat list, because
choosing between them is markup. `Item` gained `.Section` so a flat listing can still
say where an entry came from.
Two judgement calls recorded rather than left implicit. The typographer stays: turning
`--` into an en dash is a character transformation no stylesheet can express. Chrome
strings stay: translations are data, and the alternative is every theme hardcoding
Bengali month names. The inlined stylesheet is accepted with its cost written down —
bytes per page, no caching — and a trigger for revisiting it.
Two patterns worth reusing came out of this: offer the shape rather than choosing it,
and a split feature is normal — search will be an engine-built index queried by the
browser, not one or the other.
A draft is now not served at all, and neither is a bundle whose date has not
arrived. The filter sits in `Site.Lookup` and `Site.Run`, which is every path to a
bundle — so the files inside an unpublished bundle inherit its status for free,
which is what ADR-0024 asks for and what the asset route was written to allow. A
test asserts the 404 for the bundle *and* its picture, and that nothing leaks into
a listing, a feed or a sitemap.
The clock is read per request rather than at startup, so a scheduled post appears
exactly when its date arrives with nothing to restart and nothing to invalidate.
That first clock read created internal/content/clock.go, which is the only place
`verify.sh` allows `time.Now` — a render that depends on the time is worth being
able to find.
`-dev on` reveals both and reparses the theme before each render. Deliberately not
a bare boolean flag: turning unpublished work into public work should not be one
fumbled argument away. A reload that fails to parse leaves the working template set
in place, so a typo shows an error rather than replacing a good set with a broken one.
The human asked whether widow prevention belonged in the backend at all. It did
not, and it broke two rules already written down: the theme contract says the
engine decides nothing about how something looks, and ADR-0034 says authored body
text is the author's — while this inserted U+00A0 into that text.
The practical harm follows from the layer error rather than from a coding mistake.
The engine cannot see the line box, so joining the last two words is a guess that
can overflow a narrow viewport, and a reader copying the paragraph gets a
non-breaking space in their clipboard. `text-wrap: pretty` and `text-wrap: balance`
in the reference stylesheet know the line box and need no bytes in the content.
108 lines of engine deleted for one CSS declaration. The typographer stays: turning
`--` into an en dash is a text transformation no stylesheet can express, which is
exactly the distinction the new layer test draws.
Also worth recording: this took the Extensions counter from 3 back to 2. A threshold
reached by a feature that should not have existed was never a threshold.
The constitution asks which primitive a feature is. It never asked *where* the
feature belongs, and that let a whole feature get built at the wrong layer.
Four layers, outermost wins: content on disk, engine (facts only the engine can
produce), theme (markup), browser (CSS, then JS). Two rules fall out — the engine
never edits authored text to change how it looks, and a feature needing no engine
fact is not an engine feature. architecture.md carries the table and the test;
CLAUDE.md carries the one-line version, since it is the file always loaded.
The example is named in the docs on purpose. A rule with a scar attached is one an
agent can apply; a rule stated in the abstract gets reasoned around.
The server never fails on a bad bundle — it works around it and logs, which is easy
to miss (ADR-0029). This is the command that looks on purpose, and it exits non-zero
on anything that makes the site wrong rather than merely untidy.
Fatal: content the engine had to drop (unparseable frontmatter, a key two files
claim, an ignored slug or alias) and links that will 404 for a reader. Warnings:
no title, a figure with no alt text, a series where some members declare order and
others do not — that last one because unordered members sort last, so adding order
to one chapter silently moves every chapter that lacks it.
Two things this needed rather than invented. `Scan` and `Site` now *return* what they
worked around instead of only logging it: `ScanReport` and `Site.Problems`, with
`Scan` staying the logging wrapper so nothing else changed. And required-fields-per-type
is deliberately absent — `title` is the only field the engine requires today, so
checking more would mean inventing the type declaration that is still parked. Its
trigger stays where it was.
The link checker asks the same questions the resolver asks — bundle, alias, section
listing, file inside a bundle — because a checker that guesses differently from the
server is worse than no checker. Engine-owned paths are skipped: they are generated,
not authored.
It lives in internal/ext/check, so cmd/ stays wiring and the feature stays deletable.
Verified against the evidence site: clean before, and five findings across five fault
classes after I introduced them on purpose.
The entry said to measure first and put the number in the commit, so: a plain page
renders in 14µs, a twelve-picture gallery in 1.23ms. Of that, ~102µs per picture
was reading, hashing and decoding bytes the previous request had already read.
Remembering that one fact — keyed by path, size and modification time — brings the
same gallery to 63µs. 19.5× faster, 21× fewer bytes allocated, twenty-odd lines.
After which nothing is slow enough to justify caching whole pages, so ADR-0044
declines the page cache and leaves the parked validity model parked, now with a
measurement rather than an intuition behind its trigger.
That parked model has five axes and was written before any code existed. The
problem it would have been built for turned out to be one repeated file read.
Benchmarks live in internal/web so they measure through the real handler, which is
also what conventions.md wants before any cache goes in the render path. The
invalidation risk has its own test: an edited picture is a different key, so the
memo cannot serve yesterday's dimensions. Everything runs clean under -race, since
the map is read by concurrent requests.
Membership is a publication date, not a type declaration (ADR-0043). The parked
feed shape said "every type declared primary", which would have made feeds wait on
declared types a third time — but the thing that distinguishes a feed item is
already on disk. Pages and section landings drop out because they have no date,
which is the right reason. The parked idea stays parked with a sharper trigger:
someone wanting a *dated* bundle kept out.
Built with encoding/xml from typed structs, never a template: XML in html/template
is escaping for the wrong grammar, and that is a correctness trap rather than a
matter of taste.
A bug the evidence found, older than feeds: content.URL("", lang) built "//", so a
whole-site feed's id and alternate link were https://khosra.example// — every
entry identity wrong in every reader. The root is "/" now, with a test, and the
hand-built "/" the resolver carried for the same reason can follow later.
Two counters re-scoped rather than incremented, the same way transforms was:
Views now counts *per-bundle selection* — the thing architecture.md means by the
View layer, still at zero consumers. Output formats are not it: HTML, sitemap XML
and Atom are three functions with nothing to share, so an interface over them
would have one member and no leverage.
Effects stays at 1. A feed is generated per request like the sitemap, so it is not
a second Effect and the runner is not yet due — the next thing that writes files
off the request path is.
A pass over the content at startup writes three widths per picture into a cache
outside the site root, named by the source's content hash and the width (ADR-0042).
Idempotent by construction: a rerun stats and skips, an edited picture takes a new
name, and nothing stale can be served under an old one. Restarting the evidence
site made 0 derivatives the second time, as it should.
Ahead of the request rather than during it, because resampling is felt and there is
no page cache yet to hide it. Outside the site root, because the engine reads that
directory and must not leave generated files in somebody's content git — a lost
cache costs one startup pass and no correctness.
Markup now carries the original as src, the derivatives as srcset closed by the
original at its own width, and width/height from the original — which retires most
of the latent row about the output floor; only a gallery's alt is still empty, and
a filename cannot supply that.
Two things the work itself decided:
`Fragment.Items` became `Fragment.Pictures`, ADR-0037's own revisit trigger. Items
had one consumer, so widening it beat adding a second list beside it.
"A browser can show it" and "we can resample it" are different questions, and
conflating them nearly deleted content: an SVG has no decoder here, so a single
predicate would have dropped SVGs from galleries silently. Undecodable and
unsupported pictures are now rendered as they are, without a size or a srcset.
ADR-0040 puts the dependency approval on the record, as hard rule 2 requires: the
standard library decodes JPEG, PNG and GIF but cannot resize, and image/draw
scales only by nearest neighbour, which is visibly wrong on the photographic
downscales this site is made of. Four modules against a cap of six, no transitive
dependencies, and WebP decoding comes along. AVIF still has no decoder anywhere,
so it will pass through untouched.
ADR-0041 raises CORE_LOC_MAX 2000 → 2800. The old figure was costed before any
code existed, for the spine and the render path, and never budgeted what cmd/ will
hold: `check` ~250, `new` ~100, `-dev` ~100, change detection ~100, page cache
~200. On a core already at 1870 that is ~2620 before anything optional, so the gate
was going to fail work nobody would call excess. A costing error corrected, not
discipline loosened.
The ADR and HARNESS.md both say what a *second* raise would mean: that something
belongs in internal/ext/. Two ceilings exist so "core stops growing, ext rises" is
observable, and that stops being true the moment leaves are let into core.
Every figure and gallery shipped so far emitted links a browser could not fetch: a
relative src resolves under the page's URL, and nothing answered there. Found by
fetching the pages' own links rather than by reading their markup — the evidence
runs had been checking that the right src appeared, never that it worked.
A directory bundle's files are now served under its URL. The bundle is looked up
first and the file is read only from the directory that bundle owns, never from a
path assembled out of the request: ADR-0024 requires that no route serve bundle
bytes by path alone, since every byte inside a bundle inherits its publish status.
When drafts arrive at queue 19 the filter belongs beside that lookup and nowhere
else, which is why the ordering is written down in the comment.
A single-file bundle owns nothing: its neighbours belong to the section, and its
slash-terminated URL has nothing beneath it. An author with assets writes a
directory bundle, now stated in content-model.md.
A .md inside a bundle directory is never an asset — it is a bundle with its own URL
or a fragment that was never addressable, and serving either raw would publish
source. http.ServeFileFS handles content type, conditional requests and ranges,
none of which is worth reimplementing here.
The human chose the second option: a route sits beside the key rather than
replacing it. So `slug` renames what a bundle is served at, in every language, and
identity stays derived from the path — which is exactly what keeps ADR-0033 intact,
since series membership is the directory. A series landing page can now be renamed
without orphaning its chapters, and there is a test that says so.
`Site` resolves routes at index time, because only it can see whether every variant
agrees. Disagreement is dropped rather than resolved, as is a slug landing where
another bundle already answers — the same rule colliding keys and contested aliases
already follow. The key a slug moved away from stops answering, so the old address
does not quietly keep working.
Two bugs surfaced doing this, both older than this change:
An alias naming its own bundle's former key was rejected as "an alias that names a
real bundle" — which made rename-plus-alias, the entire point of ADR-0008's alias
mechanism, impossible. The check now asks what a request asks: is anything actually
served there.
Aliases were counted per declaring *file*, so a bundle whose two language variants
both listed the same alias looked like two rival claimants and lost the alias. It is
a set of keys now. This one only appears with translated content, which is why no
fixture had caught it since entry 4 — the real binary did, on the first multilingual
rename.
Earned by a real mistake this session: `web.Handler` gained a parameter, its
caller in cmd/ was updated in the working tree, and `git add internal docs` left
that caller out. verify.sh was green throughout, because every gate looks at the
files on disk rather than at the commit being made. The result was a commit that
did not compile — the kind of thing git bisect trips over for as long as the repo
exists. I rewrote the two local commits rather than adding a fix-up on top.
The check builds a throwaway checkout of the index via `git write-tree`, so it
cannot touch the real index or working tree, and it only runs when something is
staged. Proved both directions: staging a signature change without its caller
fails, a clean tree passes.
Eighth gate defect found by running the harness against real work rather than
reasoning about it — and the first that was a missing gate rather than a wrong one.
Two exact paths a crawler asks for by name, so they are mux entries rather than
resolver cases — no bundle can collide, since a key always sits under a section.
robots.txt at the site root is served verbatim, because a site that ships one has
said something deliberate; otherwise the engine emits the minimum that is true and
points at the sitemap. The sitemap lists every bundle in every language it exists
in, since each variant is separately reachable, with lastmod only where a bundle
has a date. Every URL comes from content.URL like every other path the engine
emits, so a sitemap cannot disagree with what is actually served.
Both need a declared base. Without one the sitemap answers 404 rather than listing
paths no crawler can resolve, and robots omits the Sitemap line rather than
writing a relative one.
write() was setting text/html for every caller, and headers only go out with the
first byte — so a handler setting its own type would have had it silently replaced,
which is how a sitemap gets served as a web page. It now splits into write and
writeAs, and the tests assert the content types rather than only the bodies.
The human asked for YAML site config alongside the -base flag, which is the
trigger the parked settings cascade was waiting for — but only for its top level.
So this adopts the site level and leaves section and bundle resolution parked,
with their trigger unchanged: the first section-level override with a reader.
The parked idea now says so, rather than reading as untouched.
Absolute URLs are the forcing function. A canonical link, an hreflang, an
OpenGraph tag and a sitemap entry all need the site's own origin, which the engine
cannot infer from a request it may be serving behind any proxy.
Precedence stated in the ADR: -base beats the file, so a staging host needs no
edit to content.
The last two words of a paragraph or heading are joined by a non-breaking space,
so one word never falls alone onto its own line. Deferred from entry 11 for the
right reason: over rendered HTML this cannot tell prose from an escaped code span,
so it had to wait for a tree transform.
The interesting failure is worth keeping: written against goldmark alone it passed
six tests, and did nothing in the real engine. The typographer splits a text run
wherever it looks for a substitution, so a paragraph ending "hand." arrives as two
text nodes and the last of them holds no space at all. A version that inspects
only the last child therefore finds nothing to join. It now takes the whole
trailing run of text nodes, stopping at a line break or any markup, and there is a
regression test that builds both extensions together — the only configuration that
would have caught it.
The joined text becomes a String node, which carries its own bytes: a segment is
an offset into bytes every node shares, so editing the source in place is not
possible. That path still escapes, and a test says so, since otherwise this
transform would be an injection route.
PhaseMarkup is now empty in extensions.md, and honestly so: everything expected
there turned out to belong either earlier or later.
Found by serving the include evidence: `tools.md` beside a bundle's index was
itself scanned as a bundle, so a file meant only to be included took a URL of its
own, appeared in its section's listing, and turned the including bundle into a
one-member series. The real binary showed the phantom series nav; no test would
have, because every fixture happened to name its partials differently.
The rule mirrors the one directories already have, `_index` excepted since that
names its directory. `{{< include file="_tools.md" >}}` is now the shape to write.
{{< include file="notes.md" >}} renders a file from the bundle as Markdown in
place. The included file is converted by the same goldmark instance that is
rendering the page — handed to the transformer in Extend — so its configuration
can never drift from the page's.
Three properties, each tested:
A name containing ".." is refused. os.Root would stop a path leaving the site
root, but path.Join collapses ".." long before the filesystem sees it, so without
this an include could read a template or a stray dotfile from the site root and
publish it. Verified the test fails without the guard: it took three levels of
".." from content/pages/d to reach the root, and the first version of the test
used two, so it passed either way and proved nothing.
An include inside an included file renders nothing and logs. The nested parse is
marked, so one level is all there is and a file including itself is a log line
rather than a stack overflow (ADR-0038, ADR-0029).
A gallery inside an included file still resolves, because the nested parse carries
the same Origin.
The node gained `content` for output a feature produced itself, plus `isContent`
so a failed include renders nothing instead of falling through to a fragment
lookup and complaining about a template that was never meant to exist.
The human chose non-recursive includes over textual splicing, and this amends the
phase table that promised otherwise: includes turn out to be parse-phase, not
load-phase.
The reason is in the ADR rather than lost in a commit: splicing an included file's
AST into the page is invalid, not just buggy, because goldmark nodes hold offsets
into their own source. Converting separately is what remains, and once conversion
is separate, nesting costs a read and a parse per level with a crash at the end of
any cycle. Forbidding nesting removes that failure mode rather than bounding it.
Names the declined option and its trigger, so textual splicing can arrive with the
Stage pipeline's load phase if composing partials ever becomes a real need.
A feature now learns which bundle is rendering: render.Bundle puts an Origin —
the bundle's directory plus the rooted fs.FS — on the parse context, and
render.OriginFrom reads it back. Available while parsing, not while rendering,
which decides where a feature does its filesystem work: goldmark hands the
context to a block parser and not to a node renderer, so gallery gathers its
filenames at parse time and carries them on the node.
Reads stay inside the site root because Origin passes the fs.FS rather than a
path to join (ADR-0031).
Fragment{Args, Items} lands with it (ADR-0037), so figure's template now reads
.Args.src. Authored arguments and engine-gathered items stay in separate fields:
a src argument beside a src the engine found would otherwise silently pick one.
A gallery is pictures beside the bundle, in filename order, skipping
subdirectories and anything a browser cannot show. Filename order is what makes
the sparse numeric-prefix convention work without numbers in URLs (ADR-0016).
New latent row: the reference theme's images carry no width/height and a
gallery's carry no alt, which is below the output floor conventions.md states.
Nothing can supply either yet — dimensions need the image read, and a filename is
not alt text. Queue 13 computes dimensions and brings structured items with it.
gallery is the second fragment and needs a list of filenames, which the current
map[string]string cannot carry. The theme contract says fields are added but never
renamed, so widening the shape costs one commit today and a contract version once
a theme exists — this is the last cheap moment.
Naming the argument map also stops arguments and gathered data colliding: a call
with a src argument beside a feature-supplied src would otherwise silently pick
one. figure becomes .Args.src.
Since Arc 1 the latent list has carried "raw HTML is omitted only because
goldmark's default omits it". Shortcodes were the trigger, and they came and went
without needing unsafe mode — a call renders a theme template (ADR-0036). So the
right close is not a note saying we got away with it, but a gate: verify.sh now
fails on WithUnsafe in any non-comment line under cmd/ or internal/.
Proved both directions before keeping it. The first version also failed on the
comment in render.go that explains the rule — a gate that fires on its own
documentation is one nobody keeps, so it now skips comment lines. That is the
seventh gate defect found by running against real code rather than reasoning
about it.
The latent row is gone rather than reworded, and the untrusted-boundary row says
which half is now mechanism and which half still isn't.
A call is `{{< name key="value" >}}` alone on a line, parsed by a goldmark block
parser into an AST node and rendered by executing a theme template of that name
(ADR-0036). `figure` ships; `include` and `gallery` need the including bundle's
directory, which the parser does not carry yet, so they wait.
The layering did the design work here. internal/render may not import
internal/ext, so render.New takes a callback that receives a Partial and returns
Markdown extensions, and cmd/khosra/wire.go holds the only list of enabled
features. Empty that list and the engine still builds and serves — which is the
property extensions.md says the contract should have.
Raw HTML stays disabled. An author's text reaches a page only as arguments that
html/template escapes in context, which the real binary shows: a hostile alt
becomes <script> and src="javascript:…" becomes #ZgotmplZ. Getting
contextual escaping from the standard library rather than writing it is the whole
reason a fragment renders this instead of the feature.
parseSet became variadic so the fragment set reuses it rather than growing a
second copy of the overlay logic; `Partial` takes map[string]string after the
advisory correctly flagged `any` as generality nothing had asked for.
Two things already recorded settle the shape. The theme contract says the engine
decides nothing about how content looks, "including how media is embedded", so a
<figure> assembled in Go would be the engine dressing content. And invariant 2
wants the trusted/untrusted split as real code, not goldmark's default.
Both are satisfied by the same design: parse the shortcode into an AST node, then
render it by executing a theme template of that name. Raw HTML stays disabled, so
every byte of HTML on a page came from a template the site owns, and an author's
text survives only as arguments that html/template escapes.
Recorded before the code because the syntax authors type is a disk contract, and
because a feature that cannot render itself needs a partial-rendering function
passed in at wiring time — a consequence worth agreeing to in advance.
Found by /invariants: a symlink under static/ pointing outside the site root
answered 500. The guard held — os.Root refused it and no bytes escaped — but the
response confirmed the path was there, where every other miss answers 404. Same
reasoning as a hidden bundle answering 404 rather than 403 (ADR-0024).
serveStatic now stats through the rooted FS first, so a directory, a missing
file, and a refused name are one answer. That also folds the old noListing and
staticFS into one function, since "cannot serve this" was already their shared
job.
The test uses a real temp directory rather than a MapFS, because the guard under
test belongs to os.Root; verified it fails with 500 against the previous code
before keeping it.
Splitting web_test.go at the seam the package already had — resolve_test.go for
what a path means, web_test.go for what happens once it resolves — because it
crossed FILE_LOC_WARN. Same response as content.go at entry 9.
Confirms ADR-0009's "same path" clause against the alternative that was on the
table: a Bengali variant at its own Bengali address. Chosen: one path per bundle,
prefixed per language, so identity stays stable across translations (invariant 3)
and the resolver keeps one lookup direction.
The declined option is named in the ADR so it is not re-proposed as new. Cost is
stated too — a Bengali reader sees a Latin address, and reversing this later needs
an alias for every published path.
No code moves: `slug` is still unread. Recorded now because nothing is published
yet, which is the only cheap moment a permalink decision has.
Localisation stops at the words around the content. No path segment is
translated: reserved segments stay `page` and `tags` in every language, page
numbers stay ASCII, and a variant is reached at the English key under its prefix.
The theme contract gains the matching rule — a URL, key or path segment must
never go through `t`, `num` or `day`.
This is already how the code behaves; every localisation function lives in
internal/render/chrome.go and the content package, which owns every URL, cannot
reach it. Recorded because nothing in the code says so out loud.
state.md now names the five frontmatter keys the parser actually lifts, so
content-model.md's table reads as the accepted format rather than a list of what
runs — `slug` in particular is unread today.
The engine now owns the words it puts on a page that the author did not write
(ADR-0034). `internal/render/chrome.go` holds the phrase table, Gregorian month
names and decimal digits per language, keyed phrase-then-language so both forms
sit side by side and a half-translated row is visible while reading. Templates
reach it through `t`, `num` and `day`, registered before parsing so a site
override's blocks may call them too.
The reference theme stops hardcoding English: "Newer", "Page 2 of 2" and every
date now come from the table, while `datetime` attributes stay ASCII because a
parser reads them.
Authored text gets goldmark's typographer and nothing else — quotes, dashes and
ellipses smoothed, code spans untouched because it works on the parsed tree. It
is a parser option rather than a function over a page, so the transforms counter
does not move; state.md now says why, so the next reader does not miscount.
Deliberately absent: relative dates, which need a validity window that only the
cache entry will have, and body widow prevention, which cannot be done safely by
a pass over rendered HTML.
Draws the line by who wrote the words. Labels, counts, month names and digits
come from an engine table keyed by (key, language) so no template hardcodes
English; authored prose is never localised and never rewritten beyond goldmark's
typographer. Machine-readable output stays ASCII in every locale.
Decided before the code, because the alternative — localising body text — is the
kind of choice that is expensive to walk back once published.
A bundle nested under another bundle is a member of that series (ADR-0033), so
`Site.Sequence` walks up to the nearest bundle ancestor and back down to its
members: ordered by `order` where set, then by name. Members resolve through the
language fallback, so a chapter with no Bengali variant still holds its place in
Bengali reading order instead of breaking prev/next.
One `.Sequence` field carries both shapes a theme needs. A landing page renders
`.Members` as an archive; a chapter renders `.Prev`/`.Next`, which are pointers
into `.Members` so `{{with}}` yields nothing at the ends. `Index == 0` is what
tells the two apart.
`Query` was deliberately not extended. A series ascends where `Run` descends, and
an order knob on `Query` is the config knob rule 6 bans; instead `Site.keys()`
came out so both iterate the index one way, deleting `Run`'s own dedupe map.
`draft` is not honoured: no bundle carries the field and nothing else excludes
drafts, so entry 19 adds it in both places at once. Recorded in content-model.md
rather than left implied.
state.md also corrects six inventory rows that had drifted before this change —
three LOC figures, the test total, `go.mod`, and two lines that were flatly wrong
("Dependencies: none", "goldmark is not yet imported"). The coupling gate proves
state.md changed with the code; it cannot prove the numbers are right.
Answers the question that blocked sequences. A chapter belongs to the series it
is nested under, so the engine reads no `series` field; `order` stays, optional,
and a member without one sorts by name after every member carrying one.
This supersedes ADR-0016's membership clause and keeps its position rules — the
point of that ADR was that position never reaches a URL, which still holds. The
directory already states membership unambiguously; a second statement in
frontmatter can only agree or be a typo that silently orphans a chapter.
Cost, stated in the ADR: a series can never span directories, and membership now
rides on the bundle key, which ADR-0008 makes permanent.
state.md records the open question: ADR-0016 says series membership is a frontmatter
field, and building it revealed structural nesting as a simpler option that
contradicts that ADR. Reversing a recorded decision is a decision, so entry 10 waits.
The scratch queue gains a Resuming section — position, the per-entry rhythm, where
the evidence site lives, and every deferred ADR with its trigger — so a context reset
loses the conversation and not the plan.
The queue said to re-adopt the taxonomy decision when tags landed and I did not,
so ideas/deferred-decisions.md claimed no tag pages existed while /tags/ had been
serving for two commits. ADR-0032 records what the code actually does; only the
feed half stays deferred, and it lands with declared types since feed membership is
part of a type declaration.
Folded from a separate state commit:
state: bump verified-against; narrow the declared-types trigger
Sequences do not need a type declaration — membership and ordering come from
frontmatter — so the trigger for declared types is feeds or check, whichever lands
first.
content.go had passed FILE_LOC_WARN, which conventions.md treats as the moment to
split rather than a number to ignore: flat until the figure, then split, never
pre-partitioned. content.go now parses bundles and builds permalinks; site.go holds
the indexed site — lookup with language fallback, aliases, Query and Run. The tests
follow the same seam.
No behaviour change, and the test-coupling gate was right to demand the tests move:
its exemption covers comments and whitespace, not code relocated between files,
where an edit could hide.
One global namespace (ADR-0018): /tags/{term}/ spans every section and
/{section}/tags/{term}/ narrows it. Listings group by section so one busy term
stays readable, which needed List.Groups alongside Items — list.html renders
whichever is set.
This is Query's second use, so it gained a Tag field rather than being generalised
on speculation: one filter, two callers. Tag slugs lowercase and hyphenate,
preserving script, so "Long Monsoon" and "long monsoon" are one term while Bengali
passes through unchanged. Hand-chosen slugs per term still wait for the type
declaration that owns overrides.
`tags` is reserved at the top level and inside every section, alongside `page` and
the language prefixes. A tag listing redirects to its canonical URL only once it is
known to exist, matching the rule bundles already followed — otherwise a canonical
URL for nothing confirms what is not there.
One stale test expectation fixed rather than worked around: it asserted tags land
in Extra, which stopped being true when tags became a named field.
Evidence: /tags/monsoon/ lists Hello World under posts and First Rain under comics;
/comics/tags/monsoon/ shows one; /tags/monsoon 301s; /tags/nothing/ and /tags/ 404.
A site's templates/ is parsed after the embedded set, so the last definition of a
name wins and a theme redefines one block while inheriting the document
(ADR-0019). Per kind exactly two files are overlaid — base.html and that kind's
block — because overlaying every site template into every set lets a listing's
"main" leak into bundle pages, which is the collision per-kind sets exist to
prevent. Both directions are tested.
templates/theme.css in the site root replaces the reference stylesheet outright;
there is no merging to reason about. static/ is served verbatim under /static/,
through the same os.Root-backed fs.FS, and directory paths answer 404 so it never
indexes its own contents.
Queue entries 6 and 7 are deferred again with triggers: the cascade's consumers are
stage toggles, view selection and cache flags — none of which exist — and declared
types are read by ordering, feeds and check, none of which have landed. Building
either now is the speculation that justified withdrawing them.
Evidence: a real site override renders <section class="mine"> inside the embedded
document with canonical intact, the listing still uses the embedded block, and
/static/site.css and /static/img/logo.svg are 200 while /static/ and /static/img/
are 404.
The first collection page earns the Query primitive: content.Query{Section, Lang}
with Site.Run, newest first, undated after dated, ties broken by key so the same
query always answers in the same order. No cache signature — nothing caches, and a
signature with no consumer is speculation.
Pagination lives in the path (ADR-0028): page one is the bare listing URL,
/page/1/ redirects to it, and a page past the end is 404 rather than an empty page,
because an empty page is a URL that means nothing. `page` is therefore a reserved
segment inside a section, now recorded in content-model.md.
Two kinds of page means two parsed template sets already — base plus the block that
kind defines — which is ADR-0019's per-type shape arriving by need rather than by
anticipation. A head struct is embedded in both Page and List so base.html has one
contract, and theme-contract.md gains the listing fields.
Bundle gains Date, accepting an unquoted YAML date or an RFC 3339 string, since
yaml.v3 hands back time.Time for one and a string for the other.
Evidence: 12 posts → /posts/ shows 10 with rel=next to /posts/page/2/,
/posts/page/2/ shows 3 with rel=prev to /posts/, ordering is post-12 11 10,
/posts/page/1/ 301s to /posts/, /posts/page/9/ is 404, /bn/posts/ is 200.
/invariants at the arc boundary: 1, 3 and 5 held and were checked rather than
assumed; 2, 4, 6, 7, 8 and 9 are not yet applicable, with the pre-freeze core
figure recorded as 619.
Invariant 2 turned up a load-bearing default nobody had written down: goldmark
omits raw HTML unless WithUnsafe is set, verified, and that is the only reason
authored Markdown cannot inject script today. The obvious move when a shortcode
needs to emit HTML removes it, so it is now a latent item triggered by queue
entry 12 rather than a surprise.
The other two latent items are triaged as the arc-close rule requires: the
counter-correctness gap is accepted with a reason, the untrusted-boundary gate is
scheduled to Arc 3.
The %w advisory counted every fmt.Errorf without %w, including calls that create
an error rather than wrap one — so it fired permanently on correct code. It now
looks for a call that passes an err and drops the %w, which is the actual rule.
Recorded the general form in the loop's Verify step and in HARNESS.md: a warning
that fires on correct code and keeps firing is a defect in the check or the code,
resolved in that change. Two advisories have now been narrowed after firing on
code the harness itself mandates, and the count creeping from one to two was the
only signal either time.
verify.sh is at zero warnings, which is what makes the next one legible.
The advisory fired on stringList(v any), which reads the open page object that
ADR-0002 mandates — so its false-positive rate was total, and a warning that is
always wrong teaches you to skim warnings. conventions.md bans interface{} as an
API escape hatch, which is what the check now looks for: any or interface{} in an
exported func or method signature. Verified by adding an exported func Do(x any)
and watching it fire, then reverting.
An alias is a promise that an old URL keeps working, so it answers 301 to the
canonical one rather than serving the same content twice (ADR-0008). Frontmatter
takes a scalar or a list and tolerates surrounding slashes, because authors write
both.
Ambiguity is dropped, not resolved: an alias naming a real bundle, or claimed by
two bundles, is logged and ignored so the real bundle keeps its URL. Aliases
compose with language prefixes for free, since the resolver splits the language
before the key is looked up.
The redirect still fires only for an alias that exists, so a nonexistent path
cannot be probed by 301 — the property prompt 3 established.
Evidence: /pages/bio/ and /about/ both 301 to /pages/about/, /bn/pages/bio/ 301s
to /bn/pages/about/, and /pages/nothing/ is 404.
The default locale stays at the root; every other language is the same key under
/{lang}/ (ADR-0009). /en/… is never live and redirects to the root form so the URL
space cannot fork. Lookup now takes a language and reports which one it served,
following requested → default → any rather than 404ing when a translation is
missing.
That is the second routing case, so the resolver is extracted to resolve.go and
the mux keeps one entry: URL shape is the resolver's business. A leading segment
counts as a language only when some bundle is written in it, so an unknown prefix
is a 404 rather than a stripped path — and a section may not be named after a
language in use, now recorded in content-model.md.
Because the served variant can differ from the URL requested, Page gained
.Canonical (the variant actually served) and .Alternates for hreflang. A theme
must never build a path, so both come from the engine.
Evidence: /bn/pages/about/ serves the Bengali body with lang="bn" and canonical
/bn/pages/about/; /bn/posts/hello-world/ falls back to English with canonical
/posts/hello-world/; /en/pages/about/ 301s to /pages/about/; /fr/… is 404.
Claude Opus 5andbdeshi