Release history

v0.1.0 put the eight-package table above the fold with “two real, two empty, four not published” as its headline, and everything after it inherited that frame. The feedback was fair and the numbers say so: 5,401 lines of source with 604 tests behind it, a CLI of 14 commands, a type system where edges carry domain and range constraints, four storage backends, four live graphs, 59 documents and eleven agentic roles is a substantial thing to have built, and a visitor was meeting the release hygiene before any of it.

The front page now leads with what exists. A proof strip in the hero — source, tests, commands, surfaces, live nodes, documents. Then the install and quickstart. Then What is built: six cards for the type system, the four backends, the CLI, the live graphs, the foundational corpus and the eleven roles. Then the three agent surfaces, all three marked as shipping or usable rather than one of them as an absence. The PyPI section is now “what you can install right now” — the two real packages with what is in each wheel — with the other four named as built and further along on dev, which is what they are.

The gaps are still published, in full, as a backlog. A closing section — the backlog, for when this project has energy again — carries the overdue dev→main merges, the two service placeholders, the README refresh, the three built-but-unwritten features, the unnamed copyright holder and the seven open questions. Nothing was removed and no number changed. What changed is the framing they are true in: the code has been ahead of its documentation since February, which is the good direction for a project to drift in, and every item is a chore rather than a defect. Most are one command.

Carried through to the pages a reader lands on next, so the tone does not snap back on the first click: /shipped/ leads with the two installable packages and describes the four unpublished repositories by what is in them; the release mechanism is titled for the mechanism rather than for the absence; corrections is a refresh list that opens with the direction of travel — more tests than claimed, more submodules, more roles; nothing promised and quietly dropped. llms.txt got the same treatment, since it is the front page for the reader this site is most built for: what is built and how to install it now come before anything that is missing. Nav labels follow.

The shared stylesheet is carried over from the sibling sites, and it scopes both its table treatment and its code-listing treatment to main.doc — reasonably, because on those sites a table always lives on a document page. This site puts a real table above the fold: the package table is the front page's most valuable single element, and it was rendering with no panel, no borders and the two version columns running into each other (0.7.0 v0.7.1). The band-level treatment now exists, and the two adjacent right-aligned columns are separated.

Second: every listing on the site was written as <pre class="term"> and the stylesheet defines pre.shell — a dark terminal — so forty-odd code, schema and file listings were rendering as unstyled full-bleed text. pre.term is now defined as this site's light listing panel, matching pre.tree, which suits the material better than a terminal: most of what is shown here is a schema, a config or a task file rather than a transcript. pre.shell stays available where a real terminal is the point.

Neither was caught by the gate, because both are visual rather than structural — which is worth recording: the gate checks that a page is correct, not that it looks right, and the two failure modes need different instruments.

The pipeline before the content, so that every release from here goes through a gate that already works: validate → auto-tag → deploy, carried over from the sibling sites with the merge-commit anchoring and the SIGPIPE fix those sites had to learn the hard way. Two checks were added for this site specifically. Every page must carry a “for an agent” block — the house rule is that each page serves three readers and here the third one is the product, so it is checked rather than remembered. And no page may repeat a corrected claim as fact: the four false README claims are pattern-matched across the tree, and a page may show one as false only by marking the element data-corrected-claim. The corrections page exists because of that escape hatch, and cannot silently lose it.

/lite/ — the commissioning priority, and it was homeless. The Issues-FS-lite specification lived in §7 of an 8,229-word manual about email, in a different repository, where an agent looking for how to track issues would never find it. It is published standalone: the three folders, the task format with required and recommended front-matter, the four mv operations, the check-in cycle, and the writer rule. Two lines are featured because they carry the design position — “Be honest — if you don't know how to approach it yet, write that” and “each agent owns their own work plan.” It carries a cross-boundary licence provenance note, because the specification comes from a CC BY 4.0 repository while the Issues-FS repositories are Apache 2.0. Plus three pages the corpus did not have: a five-minute worked example (three tasks, one blocked and unblocked, one closed, with the actual commands — nothing in the corpus had this), lite vs full including the three things that do not travel between them and the honest statement that there is no converter, and the team view — two find commands, free, out of the layout.

/shipped/ — the navigation job. Eight packages against live PyPI with every wheel downloaded and unzipped: two real, two published-but-empty, four not published at all. The two empty wheels are stated rather than hidden — 8 files and no functional code, while their published summaries describe a FastAPI server and a Python client. The dev→main release mechanism answers every “why isn't it on PyPI” question in one table, including the sharpest one: Issues-FS__Docs's main is still at Initial commit, 24 commits behind, which is exactly why pip install issues-fs-docs 404s under a PyPI badge. And the corrections page publishes all four false claims with what was measured instead — following the convention pki.sgit.ai established, where the page listing what earlier briefs got wrong turned out to be one of the strongest on the site.

/dsl/ — the first prose documentation of the .issues flat-file DSL anywhere. Eleven source files, ~55 tests, three live example files, wired into Graph__Repository.issues_files_* — and no README, brief or architecture document had ever mentioned it. The page describes the shape, the five-stage pipeline, the repository API and the four reasons it went unnoticed, and deliberately stops short of specifying a grammar that has never been written down: the parser tests and three live files are the only specification that exists, and inventing one would be the failure this site was built to correct.

The rest of the site. /start/ and /cli/ — every one of the 14 commands verified against source with its real flags, the three output formats, and the two warts documented rather than hidden (-p means --path on init and --priority elsewhere; --depth is -D). /model/ from the schema files rather than from prose: the thirteen-field node, the doubly-stored edge that explains why 71 nodes report 141 link entries, the two identities per node, and the five regex primitives — with the layout page stating that three incompatible layouts are in circulation and that _index.json is a cache, not a source (one index says 22 issues where the disk has 24; the healthiest graph has no index at all, and that is an argument for the design). /examples/ with the four live graphs measured. /concepts/ giving the foundational five one screen each and linking out for the argument, plus the lexicon published as argued design. /roles/, /origins/, /curiosities/, /network/ and /documents/.

Two boundaries settled in public. graphs.sgit.ai owns the philosophy; issues-fs.sgit.ai owns the implementation — and the origin, with a practical test for which side a page belongs on. And the licence position: this site's content is CC BY 4.0, the upstream repositories are Apache 2.0 and name no copyright holder, the single README sentence claiming CC BY for documentation is not treated as backing, and upstream documents are therefore cited by path rather than mirrored. Ten requests (N1–N10) and twelve tasks are on comms, and seven open questions are published unresolved.