#14 put the read API behind a session or an API key. External agents
reach the wiki through wiki-index.renner.uno with BasicAuth only, and
since #14 they got 401: a regression.
The service now checks `Authorization: Basic` itself against the same
htpasswd lines Traefik's wiki-api-basic-auth middleware uses
(Auth:Basic:UsersPath: a file, or a mounted Secret directory where every
visible key is read, as Traefik reads every key of an Opaque Secret).
This is a credential check in the service, not trust in a proxy: the
header proves nothing unless the password matches the hash.
- Hash formats Traefik accepts: bcrypt ($2a$/$2b$/$2y$, BCrypt.Net-Next
4.2.1), Apache MD5 ($apr1$, implemented here and checked against the
Traefik docs example and `openssl passwd -apr1`), SHA1 ({SHA}). Any
other format authenticates nobody.
- The identity is a machine named after the BasicAuth user.
- Policy "read" accepts it; "write" stays API-key only, and the UI stays
OIDC only.
- A verified (user, password, hash) triple is cached in memory, so bcrypt
costs once per process.
- With no UsersPath configured the scheme authenticates nobody, so this
image is safe before the deployment mounts the Secret.
Tests: all three hash formats read the API; a wrong password, an unknown
user, an empty password and a user found only in a hidden Secret entry
are 401; BasicAuth cannot write and does not open the UI.
Auth:Oidc:PublicOrigin was a hand copy of the Ingress host. Traefik routes on
the Host header and Authentik matches redirect URIs strictly, so the URI is
https://<request host>/signin-oidc; https because TLS ends at Traefik and
browsers only reach the wiki over https.
The ProxyHeader scheme took the username from X-Authentik-Username for any
peer inside Auth:ReverseProxy:TrustedProxies. The deployment set that to
the whole pod CIDR (a hand copy of the Talos podSubnets), so every pod the
network policies let reach :8080 (Hermes, Gitea, Traefik, all of ai and
monitoring) could claim any human's name on the Facts page and on saved
answers. Authentik's proxy provider sends no signed identity header, so the
header cannot be verified at all.
Humans now sign in with OpenID Connect against Authentik: a public client
with PKCE, the username is preferred_username from the ID token, whose
signature is checked against the provider's JWKS. Nothing about the
network path or the request headers names a user any more.
- Auth:Oidc:Authority, ClientId and PublicOrigin are required; the host
refuses to start without them. PublicOrigin builds the redirect URI,
because TLS ends at Traefik and the request scheme is http.
- UI pages, static assets and the Blazor hub need a session (fallback
policy); an anonymous browser is redirected to the provider.
- The read API (/api/v1 reads, /ask, /questions, /rank, /status) needs a
session or an API key (policy "read") and answers 401, never a redirect.
Before, it was anonymous and relied on the gate, BasicAuth or network
policy in front.
- Writes still need an API key (policy "write").
- /health, /ready and the HMAC-signed /webhook stay anonymous.
- ProxyHeaderAuthenticationHandler and Auth:ReverseProxy are removed.
Tests: an in-process OpenID provider issues RS256 ID tokens. They cover
the PKCE challenge and redirect URI, the full code flow ending in a
session named after the token's user, rejection of a token signed by a
foreign key, the Authentik header signing in nobody, and 401s for
credential-less API calls.
A statement may carry images (a type label, for instance). The drafter
sees them as image blocks and transcribes what is printed; the deposit
stores each picture beside the note as raw/<stem>-<n>.<ext>, lists it in
the frontmatter and embeds it in the body. INoteStore gains a binary
create and a browse URL, so the raw page can link to the pictures. The
page takes pictures from a file picker (camera on phones) or from paste,
resized in the browser to 1600px.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
IFactDrafter port: a statement in, title/content/tags out. Anthropic
implements it over the transport the answer model already uses (extracted
into MessagesTransport). RecordStatementUseCase drafts, sets the person as
source, keeps the statement verbatim in the body and deposits through the
existing use case. The Facts page is the text box and the receipt; the API
shape for machines is unchanged.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Same DepositFactUseCase as POST /api/v1/facts, attributed as a human
identity from the Authentik proxy header. Cascading authentication state
carries the principal into the Blazor circuit; without a signed-in user
the form is replaced by a notice. Lists the raw sources awaiting the
nightly distill.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The state dominates a System One request and each question is scored in
isolation against it, so sending context choices separately only re-sent
the state per request and added round trips. A live ask carried 22,957
input tokens over two requests for one question; one request carries it
once. Drops the now unused AskOptions.MaxParallel.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Both phases register their use cases and map their endpoints in the same host
files. The lab host test asserted exactly one labelled query, which held only
while no other test in the shared fixture wrote to output/; it now checks the
seeded query is among the rows.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
POST /api/v1/questions runs ask (rank + judge), selects the paragraphs that
carry probability mass, and streams one answer-model call over them. Below
the coverage gate no model is called and the response names the gap. With
save: true (signed-in user or API key) the answer lands in output/ as a
synth-shaped document with sources, confidence and the answering model.
The Answers page streams the same use case and lists saved answers.
Anthropic is the first IAnswerModel adapter; the answer model is chosen per
request or from Wiki:AnswerModels:Active.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adds EvaluateUseCase, the C# port of bin/wiki_rank_eval.py: every output/
document with a question and a sources list labels a query, gold is the
cited slugs that still exist as wiki notes, and RankUseCase ranks each
one to report recall@k, hit-rate@k, MRR and a one-hop recall variant that
also credits a gold note reachable from a top-k note by wikilink or
related:. Exposes it as POST /api/v1/lab/eval (write-policy) and a new
/lab page for running it against any registered embedding model and
comparing runs side by side.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Implements IAnswerModel over Anthropic's streaming Messages API: SSE
parsing (text_delta/usage/error), 429/529 retry with Retry-After or
exponential backoff, and a pure AnswerPrompt.Build for the evidence
and judgements prompt shape. Health check hits GET /v1/models.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Rank every query, unite the top notes into one context under a per-document
and a total budget, expand context choices over paragraph or document ids
plus NONE, add the covered gate, judge context choices alone and the rest
batched, attach the answering paragraphs' text. A missing TypeSafe key is a
call-time error shown in status, not a startup failure.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adds / (search, with recent notes as default view), /notes/{slug}, /browse,
/raw/{slug}, /reports/{slug} and /maintenance, all InteractiveServer, built on
NoteQueries and IEmbeddingModel injected directly (no HTTP). Moves the former
status page from / to /status and factors its markup into a shared
StatusPanel used by both /status and /maintenance. Adds a MudDrawer/NavMenu
nav in MainLayout.
Wikilinks ([[slug]], [[slug|alias]], [[slug#heading]]) are rewritten to note
links (or a muted "no such note" span when unresolved) by a new
Wiki.Host.Rendering.WikiMarkdown, before the body goes through Markdig with
DisableHtml() so note content can't inject script. Fenced code blocks and
inline code spans are left untouched. Covered by WikiMarkdownTests.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
IJudge over TypeSafe's System One API: serialises noul/choice/score
questions and the caller's state, retries 429/529 with Retry-After or
backoff to 30s, rejects Choice past the 255-option cap before sending,
and skips the request entirely for an empty question set.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The fact sink was a second adapter over the same backend. INoteStore
gains CreateAsync (DocumentExistsException on a taken path); the raw
note's shape, slug and collision numbering live in DepositFactUseCase;
GiteaRepository only moves text.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Two classes over the same backend showed as two adapters on the status
page. One GiteaRepository, one HttpClient, one health entry; the host
lists an adapter once even when it serves several ports.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A long sync over one keep-alive connection to Gitea behind Traefik saw a
TLS connection reset after ~1000 documents and aborted the whole run.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
INoteStore reads the vault straight from Gitea (branch HEAD, a paged
recursive tree listing, raw file reads); IFactSink appends a new raw/
file via the contents API, numbering around any path collision. Both
verified against the live gitea.renner.uno API for response shape
rather than the swagger doc.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
FileVectorIndex persists one binary file per model under index/ with
atomic tmp+rename writes; FileDocumentCatalog persists a single
catalog.json, rebuilding the link graph via WikiLinks.Extract on every
upsert. Both start empty and report health Ok with a reset detail if
their file on disk is corrupt, rather than crashing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Batches requests (default size 2, matching llama-server's limits for long
notes), sorts responses by index, and normalises vectors to unit length.
CheckAsync prefers GET /health, falling back to a one-word embed call when
the server doesn't expose it. Registers as a keyed IEmbeddingModel so
multiple models can be wired up under different names.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
k.renner
hermes-bot
claude-botandClaude Fable 5.1