9 Commits

Author	SHA1	Message	Date
Claude Opus 5	3310ae3926	#305: a full pond archive before deletion and before purge ... Deleting a pond already had a strict prompt — typing the pond name, stricter than a confirm dialog. That was never the gap. The gap is that the person who deletes it loses access the moment they do: the pond leaves their view, only a Site Admin can bring it back, and the export is no longer reachable for them. So the archive is offered INSIDE the deletion flow, before the button. What it contains, and why it is not the existing export: - Every page the requester may read, as Markdown, as before. - **Every attachment of the pond**, not only the embedded ones. An attachment nobody put on a page would otherwise vanish unnoticed — which is the whole reason this issue exists. - `manifest.json`: pond settings (EFFECTIVE, defaults filled in — a preservation format must not require its reader to know Dorfteich's defaults), labels, the page hierarchy and sort keys, comments, and attachment metadata including the #199 hash so a reader can verify bytes. It extends the #210 manifest rather than adding a second descriptor, and carries an explicit `formatVersion`. - `README.txt`, because the manifest is for machines: whoever unpacks a folder of Markdown a year from now must not believe they hold a one-click restore. Decisions worth naming: - **"Complete" describes the RESULT, not the route.** A pond admin who may read every page gets `complete: true`; only an archive that actually leaves pages out is incomplete. The Site-Admin route skips the read filter (an archive taken before an irreversible purge must not depend on which ponds the operator happens to be a member of) — those are two different questions and the first version of this conflated them. - **The omission is named before the download**, with its number, in the UI and in the manifest. An archive silently missing content is worse than no archive, because it ends the search. - **Not downloading stays allowed.** A pond of test pages should not require one, and the server cannot tell whether a file arrived anyway — so the finality is stated in text instead of enforced. - **A plain link, not fetch-into-a-blob.** The api streams the ZIP; buffering a whole pond in the tab to draw a progress bar would trade memory for cosmetics. The browser reports progress and completion; what it cannot say — that the archive is being BUILT — is announced in a live region. - Read trail unchanged in kind (ADR 0023): one `export` event per classified page before any classified byte enters the stream. Attachments never travel without their page, so the same events cover them. - New audit action `pond.archived` (catalogue v1.7) with page and attachment counts, omitted pages, and completeness. Format documented in `docs/architecture/pond-archive-format.md`, including what is deliberately NOT in it (history, permissions, trash). Verified by hand, not only asserted: a real pond's archive downloaded and unpacked — README, manifest, three page files, the media file; the manifest's effective settings, per-page classification, the VS-NfD frontmatter and marking preserved in the classified page's Markdown, and the attachment's sha256 present. Plus six api tests (including that an unembedded attachment travels and that a Site Admin gets a complete archive without membership) and the a11y pack 11/11 in both schemes, which now also scans the pond settings screen. Not done, because there is nothing to attach it to: the Site Admin's purge dialog (#193) exists only as an api endpoint — there is no pond-trash UI in the web app. The api half is here and tested, so it becomes a link when that screen is built.	2026-08-01 20:24:35 +02:00
Claude Opus 5	6377faf332	#306: instance branding — logo and favicon, cropped in the browser ... An instance had no way to look like itself: the top bar said "Dorfteich" whatever the operator called their instance, `instance.name` was never rendered in the running app at all, and there was no favicon anywhere — `index.html` had no `<link rel="icon">` and `public/` held only fonts and theme-init.js. Where the line is drawn, and why: - **The api never decodes an image.** Cropping, scaling and the conversion to PNG happen on a canvas in the browser; the api checks the PNG signature, reads the IHDR dimensions at their fixed offsets and enforces the caps. An image library would put a decoder in front of attacker-supplied bytes AND would have to be carried through the `--network none` offline build. Reading two big-endian integers is not decoding. - **SVG is refused**, with its own error message rather than a generic "not a PNG": it can carry script, and serving it from our own origin would be a cross-site-scripting vector. An operator who tried one should learn that it is deliberate. - **The crop is driven by number inputs, not by dragging.** A drag-only cropper excludes keyboard and switch users outright; a number input is arrow-key operable and screen-reader readable without any custom aria. The resulting pixel size is stated in text, not only drawn as a frame. - **The variant is chosen by CSS, not JavaScript.** `theme-init.js` has already resolved `data-theme` before first paint, so the correct logo is the one painted rather than the one that appears after a flash. Without a dark variant the LIGHT logo carries both themes — the operator's own asset shown unchanged beats one they did not choose (the rule #307 extends to ponds). The settings screen warns; it never blocks. - **The favicon link is static, its resource dynamic.** index.html stays a static file and the api answers with the uploaded icon or a shipped default — that route must never 404, or the browser keeps its generic icon for good. The default is generated by a script from Node's own zlib (`gen-default-favicon.mjs`), for the same offline-build reason. - Both favicon sizes are uploaded together: one source, one crop, so the tab icon and the home-screen icon can never disagree. - Branding is served WITHOUT a session, because the login screen carries it and the browser fetches the favicon before anyone signs in. The admin screen says so — an operator may not expect their logo to be public. - The metadata is not writable through the settings endpoint: it describes bytes on disk, and hand-writing it would claim an asset that is not there. `./data/branding` follows the three-step rule #303 paid for: env default + `data-dirs.ts` entry, compose volume (repo AND the stages on ONE), and the `mkdir`/`chown` line in the api Dockerfile. `data-dirs.test.ts` is new and closes the hole that made #303's variant invisible: the nightly archive skips a missing directory WORDLESSLY, so the fence now demands that every `*_DIR` the backup env declares actually travels in the archive. Verified against the real defect — removing the line fails it by name. Audit catalogue v1.7 (`branding.changed`), carrying `scope` from the start so #307 is the same event with a different scope, not a second id. Verified: api suite 103 files green (a lone `public-api` ECONNRESET under local parallel load, green in isolation — the documented local flake); branding suite 12 tests against a real directory; crop arithmetic unit tests; a11y pack 11/11 in both schemes; /admin measured at 320px with the new section (overflow 0); and the whole flow walked in the browser: upload → crop 780×180 → stored as 512×118 → logo in the sidebar linking home with the instance name as its accessible name → topbar wordmark following `instance.name` → light logo still shown under `data-theme="dark"`.	2026-08-01 19:30:52 +02:00
Claude Opus 5	b96997501a	#303: operator-uploaded fonts — storage, API, PDF embedding, backup ... An operator holding a font licence could only use it by baking the file into a custom image, which tied every change to a rebuild and left the file out of the backup. ADR 0016 said there is no runtime font management. It also listed this exact case under "Alternatives considered" — *may become a Site-Admin- level feature later*. The amendment takes that option and answers the two objections it raised: licensing risk (Site Admins only, licence recorded with the family) and file-format attack surface (magic-byte check and a size cap, never a parse). - `CUSTOM_FONTS_DIR` (default `./data/fonts`) — a sibling of uploads and plugins, NOT inside the image-baked `FONTS_DIR`, where a deploy would overwrite it and no backup would ever see it. - One list of data directories (`apps/backup/src/data-dirs.ts`) now feeds both the nightly archive and the restore, so they cannot drift. #306 and #307 add one line each instead of a second mechanism. - Both Dockerfiles bake the path. The backup image sets its volume paths itself ("self-sufficient without compose env" — #71's lesson) and reads no *_DIR from compose; without the ENV entry the archive would have skipped the directory silently. - The PDF path already read WOFF2 from disk at request time, so it only had to pick the other base directory for a custom family. - `fontStack`/`fontEntry` take the instance's uploaded families as an argument — they are runtime data. The catalog is searched first, and a colliding family name is rejected at upload, so a custom font can never shadow a catalog one. - Deletion is never blocked by usage: an unknown family already falls back to the system stack, so affected ponds degrade instead of breaking. The count of affected ponds travels into the audit entry. - Audit catalogue v1.6 (`font.uploaded`, `font.deleted`). Verified: api full suite against a fresh database, 102 files / 571 tests. The upload suite writes into a real temp directory and reads the bytes back off disk, so the storage layer is exercised rather than mocked.	2026-08-01 14:49:13 +02:00
Claude Fable 5	9b7acab294	#232: plugin allowlist with SHA-256 hash pinning ... The install path records the SHA-256 of the delivered bundle ZIP (plugins.bundle_hash; pre-#232 installs show it as unknown until reinstalled). plugins.allowlist in instance_settings names permitted ids with their pinned hashes: empty (default) = not enforced, existing instances unchanged; non-empty = installs of unlisted or deviating bundles are rejected (plugin_not_pinned / plugin_hash_mismatch, 403), and an installed plugin outside the list or with a deviating hash does not load — absent from pond mount lists, frame/assets 404. Every rejection is audited (plugin.rejected, catalogue v1.5). A version bump changes the hash and therefore requires an explicit re-pin — the intended friction (ADR 0025). Admin UI shows observed vs pinned hash per plugin with pin/re-pin/unpin. Scope stated honestly in plugin-architecture.md: the pin answers "is this the reviewed bundle"; post-install disk tampering is platform integrity (ADR 0019), sandbox containment stays the sandbox's job. Hardening guide row + catalog advisory triage; residual risk R-03 resolved. e2e: empty-allowlist compatibility, pinned load, unpinned and tampered installs rejected and audited, pin drift blocks loading while the admin still sees the mismatch, version bump needs re-pin. Full api suite 101 files / 561 green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AUtYMxwTCMHG9mVHnwbFg8	2026-07-31 21:26:36 +02:00
Claude Fable 5	4c7f001cab	#215: trusted reverse-proxy header / mTLS client-certificate path ... For perimeters that authenticate before the application (ADR 0021 §4). Off unless BOTH AUTH_PROXY_HEADER and AUTH_PROXY_TRUSTED_PEERS are set — nothing about the header is guessed. The peer check runs against the TCP peer address only (a forwarded header is attacker-influenced): a request carrying the header from any other peer is rejected outright and audited as auth.proxy_rejected (catalogue v1.4) — that is a spoof attempt, not a misconfiguration — even when a valid session cookie rides along. From a trusted peer the header IS the identity; a session cookie never escalates beyond it; with the feature off the header is inert. Mapping is explicit (AUTH_PROXY_MAP: username or e-mail); deliberately no just-in-time creation — the header carries no verified address. The mTLS variant (AUTH_PROXY_MODE=mtls-dn) maps the configured attribute (default CN) out of the certificate subject DN the TLS terminator forwards, under the same peer rules. Session-less proxy requests key the read trail per user (user:<id>). The trust boundary is stated in security.md (the section an assessor reads closest), the VS-NfD security documentation and the hardening guide's deploy table. Tests cover all four decisions: off = inert, trusted peer authenticates (username and DN mapping), untrusted peer rejected + audited, no escalation past a session cookie. Refs #215. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AUtYMxwTCMHG9mVHnwbFg8	2026-07-31 12:50:09 +02:00
Claude Fable 5	5796b7a5dd	#214: OIDC Authorization Code with PKCE, Keycloak as reference IdP ... External authentication (ADR 0021) built on jose (#188's vetted library) plus fetch — no new dependency enters the supply chain for a security base function. Discovery-configured; ID tokens validate against the IdP's JWKS under an explicit RS256/ES256 allowlist with issuer, audience, expiry and nonce binding. State, nonce and the PKCE verifier travel in a signed HttpOnly Lax cookie keyed by a dedicated HKDF purpose (oidc-state, ADR 0020). Deploy-level configuration (OIDC_ISSUER/CLIENT_ID/CLIENT_SECRET/SCOPES/ PROVIDER_LABEL): who authenticates users is a platform decision. The login page discovers the provider via GET /auth/methods and renders the SSO button (i18n de+en). Identities use the existing slot (provider oidc:<issuer>, subject from the token). First login creates the account just-in-time — ACTIVE and mail-verified only when the IdP asserts a verified address. An existing local account is NEVER adopted silently by e-mail (account-takeover path): login refuses with oidc_link_required and the owner links explicitly via GET /auth/oidc/link (audited auth.identity_linked, catalogue v1.3). Sessions come from the one existing session service. Tests run the full flow against a protocol-faithful fake IdP: PKCE verifier at the token endpoint, JIT creation incl. personal pond, invalid state/nonce/signature/issuer/audience/expiry each rejected, the linking refusal and the explicit link flow. Verified end-to-end against a real Keycloak 26.0 (repeatable procedure documented in security.md §External authentication). Refs #214. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AUtYMxwTCMHG9mVHnwbFg8	2026-07-31 12:44:52 +02:00
Claude Fable 5	2bdb0ec2cf	#224: read-trail storage — partitioning, retention, admin query path ... Convert read_events to monthly RANGE partitions on occurred_at, with a DEFAULT partition as safety net: a lagging maintenance job must never turn the trail's hard-failure semantics into an outage for classified reads. The dedup unique pair (#223) moves to per-partition indexes (PostgreSQL cannot carry it on the parent); a bucket spanning a month boundary may record one duplicate — over-recording is acceptable, gaps are not. New daily job read-trail-maintenance (job-count fence 9 -> 10) creates months ahead — each with its dedup index — and applies the trail's own retention readTrail.retentionDays (default 365, deliberately independent of audit.retentionDays): whole expired months are DROPped without scanning, remainders deleted by range, every run audited as read_trail.pruned (catalogue v1.2; the fence regex now admits an underscore namespace). Site-Admin query path GET /admin/system/read-events answers "who read page X" and "what did user Y read" within a period — API-only by design, documented. Growth measured and documented in data-model.md: ~1 MB per 1000 events including indexes. Tests: retention pruning + audited deletion + admin queries on the shared database; the partitioned shape, per-partition P2002 dedup, months-ahead creation and DROP-based pruning against a fresh database built by the real migration chain. Refs #224. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AUtYMxwTCMHG9mVHnwbFg8	2026-07-31 12:21:45 +02:00
Claude Fable 5	488d0d06f1	#205: classification inherits down the tree; lowering is a guarded, audited act ... New pages take max(instance default, parent level); moving a subtree under a higher-classified parent raises every member below that level. No move-like path (reposition, trash-promote, purge-promote) lowers a level as a side effect — pinned by test. Raising is ordinary editorial work; lowering requires the dedicated capability canLowerClassification (pond-wide Pond Admin) in the central permission model. Both directions are audited (page.classification_raised/_lowered, catalogue v1.1) with old value, new value, actor and page. Co-Authored-By: Claude Fable 5 (1M context) <noreply@anthropic.com>	2026-07-31 06:12:26 +02:00
Claude Fable 5	000d110727	#201: stable audit event catalogue for syslog/SIEM export ... The 36 audit action ids become a typed union (AUDIT_EVENTS in audit-actions.ts) — an uncatalogued id is now a compile error; every existing id keeps its name. The published, versioned catalogue (docs/architecture/audit-events.md, v1.0) documents per event: trigger, severity, actor and target semantics, and every field, plus the compatibility promise (ids are never repurposed; retiring keeps the row forever) and the stable stdout field set. audit-catalogue.test.ts is the fence: it parses the document's event tables and fails when ids or severities drift from the code (negative case verified). Audit stdout lines now carry the catalogue severity as a routing hint — pino level stays 30 so transport is unaffected; no DB migration. Forwarding path documented: container stdout -> operator's collector; deliberately no application-side syslog client. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0168Ph5uBmHm8X28CSVpbpnJ	2026-07-31 04:51:09 +02:00