question_lint becomes `iris check`, needs_replay becomes `iris
replay`, and the needs module and local-checkout loaders come with
them. Portal is a library dependency pinned to the release iris
matches (v0.3.36), so every type iris reads is the site's own and the
two never disagree about what a page is. `iris --path questions`
still works, so a content repo's one-line CI needs only a new path.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
The 0.3.34 rule listed every followup "further down" from the current
page, so the front door offered the thank-you pages before anything
was submitted, and a followup one typed in offered the ones beyond
it. The chain half was weaker than it looked too: it matched the
question the chain answered, which is a front door anyway.
Now a followup is offered only when the visitor has earned it: their
verified chain answered the alternative that leads there (the chain
index records the alternative from here on; older nodes fall back to
every recording alternative of the answered question, and nothing
beyond it), or it hangs off an earned page by a link that records
nothing. A page behind an answer they have not given is never offered.
The walk ends at any front door, so "back to the start" never fans out.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Forty releases carried only a version bump in their commit, so what
changed lived in the commits between them and in the content repos'
pin bumps. The changelog is reconstructed from that history, newest
first, with an Unreleased section on top and a table of which version
each of the four sites runs. The publish workflow now posts the tag's
section as the Gitea release notes, and the README says to move the
Unreleased lines under a version heading before cutting.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Until now nothing outside a live site exercised the NATS side: the
aggregate engine was tested by in-memory replay only. This test
submits an answer, moves it from a desk through the engine, checks
that an undeclared move and an unknown record are refused, and reads
the record back from its own event log. It runs when
PORTAL_TEST_NATS_URL is set; CI starts a throwaway nats:2.14.6-alpine
with JetStream for it and tears it down. Without the variable the
test passes as skipped, so `cargo test` stays green offline.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
content.rs was 2,373 lines doing five jobs. It is now a directory:
mod.rs keeps the types and the resolution rules (router, refs,
reachability, build_questions) and the tests; gitea.rs loads from
Gitea and watches for reload, through crate::http; validate.rs holds
the checks a repo must pass; markdown.rs the inline markdown;
handlers.rs the Axum routes. Each submodule is re-exported from
content, so no caller changes. The loader module is server-only and
gated as such. No behaviour change; 65 unit tests unchanged and green.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Every outbound fetch - Gitea content at boot and on reload, the raw
asset proxy, the repo-info endpoint, the content-declared url resource
- built its own reqwest client with no timeout and read whole bodies
with .text(). `crate::http` is now the one client (5 s connect, 20 s
per request, 5 redirects) with capped body readers: 2 MB for a url
resource, 8 MB for a Gitea reply, 32 MB for a proxied asset; one byte
over is an error, never a truncated document.
The upload handler read a whole field into memory before comparing it
to the 20 MB cap, and axum's 2 MB default body limit had made that cap
unreachable anyway. It now reads chunk by chunk and refuses the moment
the cap is crossed; the /upload route raises the body limit to match.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
plan.yaml: `moves_by` says who presses the button into a stage when it
is not `handled_by` - other groups, several groups, or `submitter`,
the person who sent the form, through their own link (the form must
then collect a required email). `watched_by` puts a bucket on a
group's desk with no buttons. Each group's desk carries its own buttons
only; a submitter's moves become a followup page of self_transitions;
needs.yaml gains `also_moved_by` for portal's verifier.
simulate: a route is now a relay across every role that holds a legal
button, the submitter included, and each step is judged as the role
that presses it. `--dry` asks no model anything and walks every simple
route of every aggregate, for needs_replay to put through the real
engine.
plans/contractor-full.yaml: 13 needs, 6 roles, 10 aggregates. Verifier
green by construction; 40 routes, 78 transitions accepted by the real
engine, none refused. What it exposed is in its run README.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Roles overlap on one aggregate: a project manager prices a change, a
site manager completes it. A need may list the other groups whose
desks hold some of its moves; their buttons count toward reaching
`done_when` and toward the strand check exactly as `handled_by`'s do.
`handled_by` must still hold a desk of its own, so that someone is
answerable for the record.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Fallback words read like finished copy, and nemotron handed the front
door straight back. The prompt now shows every string as a slot saying
what belongs there; a slot echoed back counts as no answer and keeps
the fallback.
runs/contractor-build-2026-09-21: nemotron-3.5-lightning:30b-a3b, the
model that produced no loadable site free-form, words a complete
contractor site in 130 s. Verifier green by construction; held-out
personas route 8/8 with descriptions and 8/8 on headings alone
(qwen3:8b as the visitor). "What happens next" is still 0/2 on the
front door: no heading or description says who reads a submission.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
A plan.yaml is a needs spec that also says what each form collects
(and which stages a case may pass over). `portal-trainer build`
derives everything structural from it with no model involved:
needs.yaml, aggregates.yaml with the state machine read off `stages`
(walk the line, be turned down anywhere, succeed only from the end),
the front door, a followup page per need, a gated directory per group
holding its desk and the pages only that group may open, and every
desk button the machine allows.
The model is asked for words only, one page at a time, as a JSON object
whose shape is fixed by built-in fallbacks. Its strings are merged over
the fallbacks key by key, so a missing, malformed or partial reply
costs wording, never a working site, and it can neither add nor drop
structure. `via` in needs.yaml is filled mechanically from the heading
the words gave.
plans/contractor.yaml builds with no model into a site portal's
question_lint accepts: 6 needs met, 12 pages, 5 aggregates.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Every prompt and reply is now kept under <out>/llm/. An empty reply is
a retry, not a fatal error; a spec that never parses stops the run
before the structure phase; --visitor-llm lets a different model play
the visitor. The file-block parser finds markers anywhere in a line,
accepts a short END and lets a new FILE close an open one, after a
reply glued two files together with `<<<END>></think>`. The reference
gains a language section: copy in the intake's language, identifiers
in ASCII.
runs/rehab-entreprenor-no-2026-09-21: nemotron-3.5-lightning:30b-a3b
did not produce a loadable site. Written up with the reason and the
way forward: structure from needs.yaml deterministically, words from
the model.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Readability is counted (LIX), not judged. Understanding is three
yes-or-no questions per persona, with a gutted twin of every page as
the control: an instrument that still says yes to "Option 1, Option 2"
is not measuring the page. Friendliness is a preference between two
versions shown in both orders; a pick that flips with the order counts
as no preference. Page load (options offered, required fields in the
step) flags a page that should be split into a further question or
have fields moved to a later stage.
First run on the generated contractor site, qwen3:8b as the visitor:
control passed on every need; "what happens next" was 0/2 everywhere
and 2/2 everywhere once the page said who reads it and when (LIX 28 to
30); an official-voice control reads LIX 63 against 27. Friendliness
mostly came back "no preference" at two personas per need: directionally
right (never preferred the cold version), too thin to rely on.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
A simulation that lives in JSON never meets the engine. needs_replay
takes a simulation's cases and runs them through the code the site
runs: answers::store_answer for the submission, aggregates::transition
for each desk move, against a real JetStream with its event log and
its refusal of undeclared moves. Buckets are then read back by
replaying each record's events and reported per bucket against
`done_when` and `success`. `--report-only` takes the same scorecard
from any JetStream, a live site's included.
First run, 30 cases from a generated contractor site on a throwaway
in-memory NATS: 30 aggregates, 56 transitions accepted, none refused.
The same routes against the earlier machine that skipped a stage: 8
refused, 8 of 12 enquiries left in `open`.
The local-checkout loaders move from the lint binary to `portal::local`
so both tools share them.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
`portal-trainer simulate` reads the model `question_lint --needs-sim`
exports. A visitor model picks an alternative and fills the form in a
manner that varies by run; the case is then walked along the handling
group's own desk buttons to a finished state, a different one each
run, so every state and button is exercised. At each button the
handler model is asked one thing: with only what the form collected,
could this role really do what the button claims? Its question goes
back to the visitor, who may not know.
Who wins is deliberately not simulated. The first version let the
handler decide: qwen3:8b sent every case to "won", and after one
sentence of prompt change, every case to "lost". Outcomes belong to
the live bucket; `success` in needs.yaml is what a report over real
records counts.
`compare` sets two score files side by side. engines/ollama.sh honours
SEED, FORMAT=json and TEMPERATURE so the same case replays the same
way against two versions of a site. generator/reference.md documents
`success` and `stages`.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Found by simulating a generated site: the owner wrote "they come in,
we go and look at the building, we send a quote", and the generated
machine went open -> quoted. Nothing compared the machine to what the
business said, so it passed. A need may now list its stages as the
business describes them; each must be a state of the bucket and
reachable through the handling group's desk.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
A need may name which of its `done_when` states are a good outcome
for the business ("lost" is finished, "won" is success), checked to be
a subset of `done_when`.
`question_lint --needs-sim <model.json>` writes everything a
simulation of the business needs, resolved the way the runtime
resolves content: each need's path with the choice on every page and
the form that gets filled, each bucket's state machine with every
desk's labelled buttons, and the personas. Portal exports it because
only portal loads content correctly; the simulator (uhhm/portal-trainer)
reads plain JSON.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
runs/contractor-2026-09-21: the generator's first end-to-end run, on
intakes/contractor.yaml. Needs parsed first round, structure green
first round, 10/10 on its own personas - and 8/10 headings-only on
ten held-out personas with qwen3:8b as the visitor, because the pages
had been written with the personas in view. Phase 2 now shows the
model the needs without the personas.
`portal-trainer answer` exposes the blind answerer on its own. README
rewritten around what this is: an intake in, a verified content repo
out.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
`portal-trainer generate` turns a plain-words business intake into a
content repo in three phases, each held to portal's question_lint:
needs.yaml (then frozen, so the generator cannot weaken the spec to
pass), structure (revised against the lint's findings until green),
and wording (personas answered blind, misroutes fed back; a rewrite
that breaks structure is reverted). Model output is written to disk
only under aggregates.yaml, site.yaml and questions/**.yaml.
`portal-trainer answer` is the blind answerer alone.
The LLM is an external command, prompt on stdin, text on stdout:
engines/ollama.sh (local Ollama), engines/anthropic.sh (API key),
engines/mailbox.sh (hand each prompt to an agent or a person).
portal-portal/ is the onboarding site as a portal content repo: the
front door is the intake interview, records land in `onboardings`
(open -> drafted <-> revising -> live | declined), and it passes its
own needs spec.
runs/uhhm-2026-09-21/ollama: qwen3:8b and granite4.1:8b as blind
answerers on the same eight tasks.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
The trainer procedure as an agent skill, the engine contract, two
sector specs (software-studio: UHHM's own, green; construction: a
target with no pages yet), and the first wording run against uhhm.no.
The verifier itself is portal's question_lint (uhhm/portal 188f9bf).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
A content repo may now state what the business needs the site to do,
apart from how the pages word it. question_lint checks each need
against the pages and aggregates.yaml:
- a path of submissions gets the answer into `lands_in`, within
`max_steps` and a required-field budget (`max_fields`);
- `handled_by` holds a desk over that bucket;
- every `done_when` state is reachable through the moves that desk
(or a self_transition) offers, and no record strands on the way;
- a disabled alternative is named as the reason when it is one;
- `planned: true` reports the gap as a warning instead of failing.
Personas - visitors in their own words - become typed `choice` tasks
(the page's question, its alternatives as the options) written as
JSONL by `--needs-tasks` (`--skim`: headings only). Any external
answering engine returns `{id, choice, confidence}` lines, graded by
`--needs-score [--min-accuracy]`: a misroute is a wording defect. The
engine stays outside this crate on purpose so classifiers and LLMs
can be compared on the same tasks. The runtime never reads the file.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
`list_qualifying_questions` reduced the visitor's `?chain=` to a
bool, so one followup path unlocked every followup on the site -
redoal's `/watching` (an email-link page whose header says "never in
the nav") showed up for anyone who had just subscribed.
The nav now resolves the chain hash to the one question it verifiably
answered (the same `lookup_node` a `requires_chain` gate uses) and
offers a followup only when it is on that path or further down from
the current page over `action` edges. `content::outcomes_reachable_from`
walks those edges through followup pages only, so a "back to the
start" action ends the walk at the front door instead of fanning out.
Templated actions (`/shape/{curve.key}`) resolve to their dynamic
page's pattern. No schema change: the `followup` flag keeps meaning
"not a front door", and the edges were already in the content.
Also: `main.loading`/`main.not-found` gets the same `svh` ladder as
`.hero-piece`, the last bare `vh` in the stylesheet, so wrapper
browsers that resize the webview as their toolbar animates don't
reflow the placeholder.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Fonts hide behind the stylesheet - the browser parses CSS before it
finds them - so the latin Quicksand subset (every page's critical
path) is announced in the shell head; latin-ext/vietnamese stay
unicode-range-gated. The wordmark preloads from the resolved site
config, with a preconnect to its origin when it lives on the content
host (redoal.com's does), opening that TLS session early.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
A content repo can ship a favicon (svg/png/ico) named in site.yaml;
portal serves it at /site/<path> and injects a <link rel=icon> after
the static defaults so it wins. Also serves .ico assets.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
Was cropped to a 14rem band via object-fit: cover; now shows the whole
image at natural aspect, bounded at 32rem so a tall one can't tower.
The image deck keeps its fixed-height cards.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
Content-only image control: a markdown image title carries layout —
left/right to float text around it, a number for max-width in rem
(validated, capped) — translated to md-float-* classes and a numeric
inline max-width. No arbitrary CSS reaches the page. Portal also now
serves .avif site assets.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
A content-shipped colour logo (Klingenberg's red tile) was being
inverted to teal in light theme by the .svg heuristic. Invert is now
an explicit site.yaml flag defaulting to house-mark-only; content
logos keep their colours unless they opt in.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
A select requirement can now declare a plain options: [A, B, C] list
instead of a resource - each string is both value and label. Restores
the ancestor question format's inline enums (used by the klingenberg
port) without needing a resource endpoint for a fixed list.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
A content repo can now ship a CSS file named in site.yaml's
stylesheet field; portal serves it same-origin at /site/<path> and
leptos_meta appends it after the default stylesheet so content rules
win. Plain repo-relative .css path only, validated on load.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
The strings the portal itself says - Asked by, Also worth asking,
Sign in, the not-found page, defaults - translate via a small en/no
table selected by site.yaml's lang, which also sets the html lang
attribute. Content keeps speaking for itself; unknown languages fall
back to English per key.
Also: a scheme-gated markdown image now swallows its alt text instead
of leaking it as stray text - this is the fix for the test that has
been red since 0.3.19 (publish is tag-triggered and does not gate on
the test workflow; caught late).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y42TyF8Zu7NGRR2893vNcZ
