- _ATOM_LIST_SQL via LATERAL: zusaetzlich cpl.source_article (Gesetzes-Artikel)
im atom-grain Response. Spalte control_parent_links.source_article verifiziert
(macmini + prod).
- Registry-Mapper-Test (neue Domaenen) nach compliance/tests/ verschoben — CI
faehrt compliance/tests/, nicht tests/; schliesst die CI-Luecke der
6-neue-Use-Cases-Erweiterung.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Frontend (CRA/Cyber-Tab):
- Erklär-Zwischensätze je Ebene (Befund -> CRA-Anforderung -> Best-Practice-
Standard -> Maßnahmen) + "So liest du einen Befund"-Legende.
- Kuratierte M-Maßnahmen und atom-grain "Regulatorische Breite" in EINE Sektion
"Maßnahmen (wählbar)" zusammengeführt (statt zwei konkurrierender Listen).
- Standalone "Empfohlene Maßnahmen (Sollzustand)" entfernt (jetzt je Befund).
Backend:
- Atom-Controls-Query liefert jetzt cpl.source_article (Artikel/Anhang/Erwägungs-
grund-Anker) zusätzlich zu source_regulation; via LATERAL-Join.
- enrich_findings_with_breadth trägt source_article in regulatory_breadth.
- Daten waren schon ingestiert (682/691 CRA-Atome haben source_article) — wurden
nur nicht selektiert/angezeigt.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Schliesst die CI-Luecke (Tests lagen in tests/, CI faehrt compliance/tests/) und
flaggt backend in detect-changes, damit der zuvor uebersprungene Backend-Build
(43 Use Cases, /corpus, + Migration 153 der CRA-Session) auf Prod nachgezogen wird.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
apt-get install git scheiterte (exit 100) auf dem Runner — Debian-apt-Mirrors
nicht erreichbar — und brach damit den Checkout ALLER python:3.12-slim-Jobs
(validate-canonical-controls, test-python-backend, iace-gt-coverage, …) seit
#863. Dadurch wurde CI nie grün und Orca hat nie deployt. Das volle python:3.12
bringt git mit -> apt-get-Zeile entfällt. (dep-audits nodejs/golang-apt ist
PR-only und ausserhalb des Deploy-Pfads.)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Readiness check: legally tighter + sales-sharper copy per review — names both
regulations cleanly (CRA + Machinery Reg 2023/1230 in plain language), frames CRA
Art. 13 as "more than a yearly pentest: assess/document/handle cyber risk across
the lifecycle" (not over-claiming a "continuously documented risk assessment"),
adds the "we turn regulation into code" positioning, and reorders the 8 questions
in CRA order (machine -> connectivity -> software -> updates -> remote -> app ->
personal data -> critical env).
Track B: the Compliance Agent Pre-Scan wizard now detects the shared
CompanyProfile and offers "Aus Profil übernehmen" — tolerant mapping (legal_form,
industry, employee_count) across the differing module vocabularies, user-
triggered (never silent), so company context isn't re-asked.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Migration 153 adds compliance_cra_projects.linked_iace_project_id (additive,
idempotent). New thin router cra_link_routes.py: POST /projects/{id}/link-iace
sets the reference; GET /by-iace/{iace_project_id} returns the linked CRA project
+ its latest assessment snapshot. The IACE "CRA / Cyber" tab now resolves the
linked CRA assessment first (real, from the snapshot) and only falls back to the
demo scenario when nothing is linked. One assessment, two views.
[migration-approved] — user approved the new column for the CRA<->IACE reference.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CRA frontend pages hardcoded tenant 00000000-…-001 while IACE uses the dev
tenant 9282a473-… → a demo customer was split/invisible across modules. Align all
app/sdk/cra pages to 9282a473-… so the whole CRA<->IACE journey lives under ONE
tenant. Add scripts/seed_demo_customer.py: seeds CompanyProfile + IACE project
(components, hazards, mitigations) + CRA project (intake, scope-check, assessment
snapshot from faked repo findings + components + safety functions) — the source-
repo layer is faked so the full frontend is walkable once.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Follow-up to the machinery_reg_cyber.py removal: the readiness endpoint now pulls
Machinery Regulation 2023/1230 cyber-with-safety obligations from the shared
Controls-API (use_case=maschinen), tagged "Maschinen-VO", best-effort.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Machinery Regulation 2023/1230 cyber-with-safety obligations are already in the
shared Controls-API (use_case=maschinen, atom-grain, classified, license-clean) —
so remove the hand-authored machinery_reg_cyber.py spine. The readiness check now
fetches them from use_case=maschinen (sub_topics sicherheitsanforderungen ->
code, risikomanagement -> process, konformitaetsbewertung -> document), tagged
source "Maschinen-VO" alongside the CRA obligations. Same pattern as the security
cluster; no own formulation, no license question.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Machine/plant builders are hit by BOTH the CRA and the new Machinery Regulation.
New machinery_reg_cyber.py models its two well-corroborated Annex III cyber-with-
safety essential requirements (1.1.9 protection against corruption, 1.2.1 control-
system safety incl. foreseeable manipulation) in our own words; EU legal text is
freely reusable (Commission Decision 2011/833/EU, source acknowledged), harmonised
standards referenced by identifier only. The readiness check asks "is it
machinery?" and, if so, adds these obligations tagged "Maschinen-VO" alongside the
CRA ones — the combination is visible (regulations list + per-item source badge).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Intro now states machine/plant builders are hit by BOTH the CRA (manufacturer
duties from 2027) and the new Machinery Regulation 2023/1230 (cyber-affecting-
safety in CE), and frames CRA Art. 13 as a continuously documented risk
assessment over the lifecycle — not a yearly pentest — which we run as a living
system (versioned snapshots). Educates the lead to win them.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
For hardware CE projects (no repo) each networked component (controller/hmi/
gateway/drive/remote_access/sensor) yields typical ICS vulnerability CLASSES
(real CWE + "CISA-ICS — product-specific check" framing, NO fabricated CVEs);
they flow through the same CRA engine. /assess accepts components[]. MappedFinding
now echoes title/location/cwe so the response is self-contained for any finding
source. Live CISA-ICS/NVD per-product CVE lookup is the later enrichment.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
All 6 security use_cases are atom-grain now. Per finding we draw two lanes: the
CRA corpus (use_case=cra, the most on-point CRA obligations) + the technical
depth (code_security for secure-dev, else network_security). Controls merged,
deduped, each tagged with its use_case (shown in the best-practice depth).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Both network_security and code_security are now atom-grain. Per-sub_topic
use_case routing: secure_development -> code_security (best for secure-dev
findings), everything else -> network_security. Findings carry breadth_use_case
so the source context (which atom corpus) is visible under the best-practice depth.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Semantic breadth (2): each finding's CRA-AI is mapped to a network_security
sub_topic and enriched with atom-grain, framework-traceable obligations from the
shared Controls-API (compliance.atom_classification) — at the endpoint/view layer
(SessionLocal), NOT in the pure mapper. CRA-AI anchor + curated measure +
NIST/OWASP crosswalk stay the lead; this is breadth + source evidence. Only
network_security is queried (atom-grain), scoped by sub_topic + limit. Frontend
renders it under the collapsible best-practice depth (control_id · title · source).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Snapshot/history: "Snapshot speichern" + a version list (status, date, coverage)
you can click through — makes the CRA Art. 13 running system visible (backend
endpoints already live). Measure-class: each finding shows a remediation-class
badge from its CRA evidence_type ("Code-nah" = scan-locatable, code-fix in the
ticket possible; otherwise Prozess/Doku), and the measures section is relabelled
as the Sollzustand (process/build) — no auto-fix buttons on process measures.
Backend: MappedFinding now carries evidence_type.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
controls-augmentation now renders control_id + sub_topic + source_regulation
(handles both atom-grain and master-grain API shapes). The compliance-advisor
answers from atom_classification (precise, sub-topic-organized, framework-
traceable) via the shared get_controls_for_use_case API. 100% local at runtime.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Reads compliance.atom_classification (Haiku pass: relevant + sub_topic +
canonical_obligation) when present -> precise, sub-topic-organized controls per
topic; master-grain seed stays as fallback for unprocessed topics. New optional
sub_topic filter + subtopic_counts facet + granularity flag in the response.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Persist each CRA assessment as a versioned, auditable snapshot over the product
lifecycle. Reuses the existing compliance_cra_documents table (NO new schema,
frozen DB respected): doc_type='doc_risk_assessment', full assessment in
generation_context, requirements_coverage summary, auto-incrementing version,
prior version superseded. New endpoints: POST /projects/{id}/assess-snapshot,
GET /projects/{id}/assess-snapshots (history), GET /assess-snapshots/{id}.
Additive (no contract baseline change).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add-only table for the one-time Haiku pass result: per atomic control x use-case
-> relevant? + sub_topic + canonical_obligation. Atom-grain successor to the
master-grain mc_use_case_mappings (master clustering = gpre2 object-only ->
mega-clusters, unusable). Runtime reads only this table (no live LLM).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Deterministic bridge (cra_safety_bridge.py): a cyber finding's attack capability
(remote_actuation / code_tampering / integrity_loss / auth_bypass, derived from
its CRA category) is matched against what each CE safety function is vulnerable
to. A match re-opens the mitigated hazard, flags the finding safety_impact (which
floors it to P0), and produces the cross-link. Endpoint accepts safety_functions;
frontend passes the project's safety functions and renders the LIVE cross-links
(no more hardcode). Safety functions are demo input now; come from the CE risk
assessment in production.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CRA tab now shows the priority layer: a weights control (5 business objectives,
high/medium/low) that re-computes the assessment live; a Prio column with
P0..P3 tier badges (P0 = non-negotiable floor, reason on hover); the table in
backend priority order; and a Quick-Wins block (high impact, low effort). Demo
flags the safety-cross-linked findings as safety_impact so the P0 floor shows.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Deterministic prioritisation on top of the mapper (cra_prioritizer.py): a
non-negotiable P0 floor (safety-function compromise / actively exploited /
CRITICAL — customer weights cannot demote) plus a discretionary tier ranked by
severity x the customer's weight (high/medium/low) for the 5 business objectives
(access/data/network_api/supply_updates/monitoring). Quick-win flag (high impact,
low effort) for a second view; each finding carries a short priority reason.
Endpoint accepts weights + per-finding safety_impact/exploited. Rough pre-sort
only (devs re-sort in Jira). No DB.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CRA tab now computes the assessment live: useCRA POSTs the scenario findings
through a new /api/v1/cra/* proxy to the backend mapper and merges the live
mapping (CRA requirement, risk, measures, NIST/OWASP crosswalk) with the
frontend scenario constants (full measure texts + cyber->safety cross-links,
until those move server-side in step 2). Falls back to the static scenario if
the backend is unreachable.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Live HTTP entry for the deterministic CRA assessment — repo-scanner findings
in, CRA Annex I mapping + risk + curated measures + NIST/OWASP golden-set
crosswalk out. Project-less (works for any customer, no CE-RA/FMEA required);
reuses the tested mapper, same logic the MCP server exposes. Additive endpoint
(no contract baseline change); no DB.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Prompt-augments the RAG-only advisor with the shared use-case->controls API:
deterministic topic detection -> local controls API -> context block, so the
agent can answer from real Control-IDs. 100% local at runtime (no Anthropic).
NOT pushed/deployed: the shared API currently returns MASTER-grain controls,
whose composition is broken (gpre2 object-only clustering -> mega-clusters).
Pending the atom-grain rework of the API. tsc + vitest green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per the Co-Pilot-calm principle: the findings table stays compact (Befund /
CRA-Anforderung / Risiko / Maßnahmen) and the NIST/OWASP + ISO 27001
best-practice depth is revealed per finding via a "NIST/OWASP" toggle. Keeps
the 3-layer model (CRA obligation -> measure -> best-practice depth) tidy.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Read-only layer (service + thin route + tests) that returns the controls
mapped to a use-case/topic, ranked by a deterministic precision proxy
(is_primary + mapping confidence + registry keyword relevance) over the
existing mc_use_case_mappings seed. No schema change.
Shared handoff point: the document specialist agents AND the CRA
finding-mapper draw from this one controls index instead of separate
retrievals.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Crosswalk (cra_security_crosswalk.py): deterministic, hand-curated CRA Annex I ->
NIST 800-53 Rev5 + OWASP Top 10:2021 mapping, the authoritative Security Golden
Set (no RAG; semantic breadth comes later via the shared Controls-API). Mapper
attaches NIST/OWASP refs per finding; golden-set completeness pinned by test
(every requirement has >=1 NIST ref). CRA tab now shows the NIST/OWASP best-
practice refs per finding and the full curated measure texts + norm references
(from measures_library_cra.go).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
New "CRA / Cyber" tab in the IACE project (Zusatzmodule). Treats the
Kistenhubgeraet CE project as if it had an IoT module; invented cyber findings
are mapped to CRA Annex I requirements via the REAL backend mapper output
(faithful), and crucially cross-linked to the existing CE safety hazards they
re-open (cyber defeats a mechanically-mitigated guard -> CRA x Machinery Reg).
Frontend fixture for now; live wiring to the mapper endpoint follows.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Deterministic mapper (no DB/LLM): repo-scanner findings -> the CRA Annex I
essential requirement(s) they violate -> risk level -> remediation measures +
coverage. Reuses the existing Annex I spine (cra_annex_i_data). The MCP server
(compliance/mcp/server.py, stdio) is the thin transport the external scanner
queries; all logic lives in the fully-tested mapper. Works standalone (no
project/FMEA required). No DB migrations.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- compliance-advisor.soul: Quellenschutz/Anti-Leak ersetzt durch Transparenz-
Modus (nur Dev nutzt den Agent; offene Meta-Antworten erlaubt) + ehrlicher
Hinweis, dass der Agent nur RAG sieht, NICHT die MC-DB.
- Snapshot: Impressum/DSE/AGB-Tabs immer sichtbar (Hinweis statt Verstecken bei
fehlendem Text).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Storage-Keys fehlen in Phase A (vor Consent) oft → provider blieb leer. Jetzt
fällt detect_consent_history auf den bereits aus dem Banner-DOM erkannten
banner_provider zurück → korrekter Anbieter + history_capable auch ohne
Storage-Eintrag.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cumulative resume run added 64 new NASA NTRS docs (query/page pool then
exhausted): 164 total, 73 applicable failures, all public-use licensed.
NASA stays a generic component failure-mode library; not scaled further.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
consent_history.detect_consent_history: erkennt CMP-Anbieter (Borlabs/
Usercentrics/OneTrust/Cookiebot/…) aus Storage+Cookies, versionierten Consent
(historie-fähig) + dauerhaftes Widerruf-/Einstellungs-Widget. consent_scanner
ruft es in Phase A; scan_matrix_summary surft summary.consent_history;
browser_cross_finding: positiver Befund wenn vorhanden, sonst Best-Practice-LOW
(„Nutzer sehen, wann sie welcher Version zugestimmt haben"); BrowserBehaviorView
zeigt es im Engine-Detail. Tests: 7 (classify/versioned) + 2 Cross-Finding.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
User-Bug (BMW): Banner überlagert, aber Footer-Links (Impressum/DSE) bleiben
klickbar → fehlender In-Banner-Link ist dann nur Best Practice, kein Verstoß.
banner_text_checker misst per Playwright-trial-Klick, ob der Footer-Impressum/
DSE-Link trotz Banner erreichbar ist: erreichbar → LOW/Best-Practice (+ Borlabs-
Consent-Historie-Hinweis), blockiert → HIGH/MEDIUM wie bisher. browser_cross_
finding: redundante (nicht footer-bewusste) "Link im Banner fehlt"-Befunde raus.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
User-Feedback umgesetzt: Cookie-Titel-Fix (rendern nicht mehr als nacktes
"Befund" — Titel aus cookie/type/vendor), Executive Summary oben, je Modul
Statistik (Counts + Severity-Balken + MCs) + nur Top-3 Befunde + Verweis auf
"N weitere" mit Frontend-Link (snapshot_id) + Zwischenfazit, Browser-Übersicht,
Gesamtanalyse, klarerer "Grenzen"-Satz, Report-Versionsnummer. 6 Tests grün.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Load the prior manifest, seed the seen-set from it + anthropic.jsonl, and add
MAX_DOCS NEW docs per run (100->200->...) instead of re-processing the first
batch and overwriting the register. Widen NTRS paging to page.from 0..400 so
enough fresh usable docs are found after the skip.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
100 NASA NTRS docs processed (Claude Haiku 4.5), 55 applicable failures
extracted with verbatim source quotes; all licences public-use-permitted
(NTRS GOV_PUBLIC_USE_PERMITTED / PUBLIC_USE_PERMITTED), each passes the
failure-knowledge allowlist. Register now serves the real corpus in the
FMEA "Quelldokumente" panel.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Quote-verifiable failure extraction via Claude (Haiku 4.5): PDF sent
directly, tool-schema forces verbatim source quotes + applicable flag +
confidence — replaces the unreliable local llama run. Only applicable=true
tuples ingest into bp_iace_failure_kb; every processed doc lands in the
source manifest.
Frontend: FMEA tab now shows a "Quelldokumente" register (every document we
use, with source + licence + link + what was extracted) served from the
embedded manifest via GET /iace/failure-knowledge/sources. Manifest is
placeholder until the 100-doc Haiku run is folded in.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Schnelltest-nur-Chrome wieder entfernt (User: Banner-Test soll IMMER alle
Browser abdecken). Ein primärer Button im Leerzustand + "Erneut testen" im
Ergebnis-Kopf; beide lösen die volle Matrix aus.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Stage 1 of the FailureKnowledge bulk loader: harvest NASA NTRS
lessons-learned with a strict public-reuse gate (NTRSUsable: public
release, not export-controlled/EAR/ITAR, not CUI, PUBLIC_USE_PERMITTED,
no third-party copyright). NTRSPDFURL prefers the PDF download for
downstream text/OCR extraction. GET /iace/failure-knowledge/ntrs runs
the live harvest and returns only the licence-clean records.
Pure parse/gate helpers are fixture-tested (usable vs ITAR / third-party
/ restricted / video-only); accepted licences also pass the FK allowlist.
Next: tuple extraction (abstract -> FailureKnowledge) + Playwright/OCR for
scanned PDFs -> bp_iace_failure_kb.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Korrektheit (§ 25 TDDDG): "Cookies vor Consent" ist KEIN Verstoss per se —
technisch notwendige Cookies inkl. des Consent-Cookies (speichert die
Ablehnung) sind nach Abs. 2 erlaubt. Verstoss ist nur nicht-essentielles
TRACKING vor Consent.
- browser_cross_finding: Befund haengt jetzt an violations.before_consent
(Tracking), nicht an der Cookie-Rohzahl; § 25 Abs. 2-Hinweis im Detail.
Regressionstest: Cookies-ohne-Tracking → KEIN Befund.
- multi_browser_scanner._extract_dimensions: Score nutzt Tracking-Violations
+ reject_respected-Verdikt statt Rohzahl (Fallback erhalten).
- BrowserBehaviorView: "Cookies vor Consent" nur rot/⚠ bei Tracking,
"nach Ablehnen" neutral (Verdikt = reject-Spalte); erklaerende Zeile.
Speed: run_consent_test ueberspringt im Matrix-Modus (browser_profile gesetzt)
die teuren Phasen C/D-F/G — nur A+B noetig. Verhindert das 504 beim
Multi-Engine-Scan (BMW 4 Engines lief sonst in den 338s-Gateway-Timeout).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
E2E auf BMW (macmini, arm64) zeigte: nur Chromium lief, Firefox/WebKit/Mobile-
Safari scheiterten mit "Host system is missing dependencies to run browsers".
Die manuell gepflegte apt-Lib-Liste war fuer Gecko/WebKit unvollstaendig.
`playwright install-deps chromium firefox webkit` (als root) installiert den
vollstaendigen OS-Dep-Satz → alle Engines starten. Betrifft beide Arches.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- browser_cross_finding: deterministische Sicht ueber die Matrix (keine 2.
Engine, kein LLM). Findet Inkonsistenzen ZWISCHEN Browsern (Cookies vor
Consent / Ablehnen nicht universell respektiert / Banner-Links fehlend) und
ordnet ein: Safari-ITP / Brave-Shields / Firefox-ETP maskieren Verstoesse
clientseitig → strenge Engine "sauber" ist KEIN Compliance-Beleg, massgeblich
sind die nachgiebigen (Chrome/Edge). Coverage-Hinweis fuer nicht verfuegbare
Browser. Je Befund Titel/Detail/Severity/affected/Massnahme.
- snapshot_check_routes: cross_findings frisch in run + GET (nicht persistiert).
- BrowserBehaviorView: "Cross-Browser-Befunde"-Block ueber der Tabelle.
- Tests: test_browser_cross_finding (6).
Offen (Folge-Task): Borlabs-Consent-Historie-Live-Erkennung (braucht
consent-tester-Storage-Scan).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- check_snapshot: update_browser_matrix/load_browser_matrix — migrationsfrei
in banner_result.browser_matrix (JSONB jsonb_set, eigener scanned_at)
- snapshot_check_routes: POST /snapshots/{id}/browser-behavior/run laeuft
/scan-matrix LIVE (Re-Crawl je Engine, nur live messbar), persistiert das
Ergebnis; GET /snapshots/{id}/browser-behavior liefert die gespeicherte
Matrix ohne Re-Crawl. Profil-Set = 4 Default-Engines + Brave/Chrome/Edge.
- consent-tester multi_browser_scanner: Semaphore(2) gegen OOM (7 Browser
parallel sprengten das 2g-mem_limit)
- Pydantic-Modell mit Optional[List[...]] (nicht `| None`) → Py3.9-sicher
- Tests: _snapshot_scan_url + Request-Defaults (5)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
P1 of the auto-FMEA build plan: establish the public-domain methodology
foundation (no AIAG-VDA/SAE/IEC tables reproduced).
- fmea_data_sources.go: MIL-STD-882E severity (Cat I-IV→1-10) + probability
(A-F→1-10 with per-hour λ bands), OccurrenceFromRate(λp·α), SeverityForCategory,
MIL-STD-1629A CriticalityCm = λp·α·β·t. Own 1-10 projection, government-anchored.
- 4 versioned source docs (MIL-STD-1629A, MIL-STD-882E, NASA RCM, FMD-91/NPRD-91)
ingested into the new RAG collection bp_iace_fmea_kb (whitelisted).
- Tests for all scales/mappings/criticality (green).
Next (P1 step 2): fetch FMD-91/NPRD-91 bulk λ/α tables from DTIC.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Dockerfile: Brave-apt-Repo + `playwright install --with-deps chrome msedge`,
beide hinter TARGETARCH=amd64-Gate und best-effort (|| echo) → arm64-Dev-
Builds (macmini) brechen NICHT, laufen mit den 4 Default-Engines; Brave/
Chrome/Edge sind amd64-only opt-in-Extras (EXTRA_PROFILES).
- docker-compose.hetzner.yml: consent-tester auf linux/amd64 (statt arm64-
Emulation auf Orca) — Voraussetzung dafuer, dass die echten Browser
ueberhaupt installiert werden.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
run_consent_test nimmt jetzt browser_profile (browser_profiles.py): Firefox/Gecko,
WebKit/Safari oder Blink (Chromium-Default / Chrome-/Edge-Channel / Brave via
executable_path). Rückwärtskompatibel: None → Chromium wie bisher. Fundament für
die echte /scan-matrix (Stage-1.b-Shim), die als nächstes Profile durchreicht.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
#1/#2 (kein-Banner-affirmativ) feuerte nicht, weil der no-banner-Pfad bei
Zeile 220 früh zurückkehrt — vor dem Edge-Case-Block am Funktionsende.
Logik in _apply_edge_case_findings extrahiert und an BEIDEN Return-Pfaden
aufgerufen (Früh-Return + Ende). Damit greift #1 jetzt auf statischen Seiten.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
#1/#2: Wenn KEIN Banner erkannt UND kein Tracking vor Consent (statische Seite
oder nur technisch notwendige Cookies, §25 Abs.2 TDDDG) → affirmativer LOW-Befund
"konform, kein Banner nötig" statt stillem "Banner fehlt". Inkl. Geo-Caveat
(Scan außerhalb EU sieht geo-getargetete Banner evtl. nicht).
#3: detect_non_cookie_tracking erkennt Pixel/Fingerprinting per Domain-Signatur
(Meta, TikTok, LinkedIn, Pinterest, Clarity, FingerprintJS, Hotjar, Reddit,
Snapchat) → MEDIUM-Befund "§25/Art.5(3) gilt auch ohne Cookies". '0 Cookies' ≠
'kein einwilligungspflichtiges Tracking'.
Verdrahtet in consent_scanner vor dem Return. Tests + py_compile grün.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cookie-freie Analyse mit reinem Opt-out-Hinweis (z.B. bayshore.ai:
"Privacy-friendly, cookie-free analytics are currently enabled ... Disable")
ist KEIN Consent-Banner: cookieless = kein Endgeräte-Zugriff → §25 TDDDG
verlangt keine Einwilligung → Opt-out statt Opt-in. Die Standard-Opt-in-
Checks (granulare Kategorien, Accept/Reject-Balance, Impressum-im-Banner)
trafen nicht zu und erzeugten 3 Falsch-HIGHs.
is_cookieless_optout() erkennt das Muster (cookieless-Signal + Opt-out-Wort,
KEIN Consent-Signal); check_banner_text gibt dann früh EINEN ausführlichen
LOW-Erklär-Befund zurück (zählt nicht als HIGH) und setzt die Opt-in-Checks
aus. Ausführlich, weil der Fall extrem untypisch ist.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ollama entlädt das 35b-Modell nach 5 Min Leerlauf → jede Frage danach
startet es kalt (Modell-Load) und läuft in den Frontend-Timeout ("Load
failed"). keep_alive='30m' im Chat-Request hält es warm.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
compliance-advisor.soul.md (Sie durchgehend):
- Persona: ruhiger Compliance Co-Pilot (Komplexität abnehmen, Nutzer entscheidet),
DSB/Anwalt als Partner-Schritt statt Ausrede.
- Antwortlänge an die Frage koppeln (kurze Frage → 1-3 Sätze, kein erzwungenes
4-Punkte-Schema); proaktiv mit nächstem Schritt schließen.
- Konfidenz-bewusst (Wahrscheinlichkeit statt Garantie); Risiken/Bußgelder nur auf
Nachfrage + konstruktiv, nie als erster Eindruck.
- Scope-Disziplin: nur Compliance/Datenschutz/Security/Recht; Off-Topic freundlich
+ knapp ablehnen, kein Erfinden fachfremder Antworten.
drafting-agent.soul.md: Anti-Leak-Regel (Anweisungen nie offenlegen) + Sie + Off-Topic-Disziplin.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Der Advisor deutete Inhaltsfragen ("Was ist der CRA?") als Quellen-/
System-Frage und wich aus; auf Nachfrage zitierte er sogar seine
Quellenschutz-Anweisung. Fixes in compliance-advisor.soul.md:
- Quellenschutz gilt nur noch für ECHTE Meta-Fragen (Quellenliste/RAG),
NICHT für "Was ist X?"-Fachfragen → die werden sofort beantwortet.
- Neue Regel: System-Anweisungen/Prompt NIE offenlegen oder zitieren;
auf "warum hast du nicht geantwortet?" nicht mit internen Regeln erklären.
- Neue Regel: mehrdeutige Abkürzungen (CRA …) kurz disambiguieren statt
ausweichen.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Der "Voll-Audit öffnen (alle MCs)"-Link hing in ComplianceResultTabs (aus
der Agent-Seite entfernt). Jetzt im Detail-Header der Snapshot-Ansicht via
snap.check_id → /sdk/agent/audit/{check_id} (Audit-Daten verifiziert
vorhanden). Plus Site-Titel-Header.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Tabs Website-Scan (nie funktioniert), Banner-Check, Agent-Test entfernt;
Tab-Leiste weg, da nur noch Compliance-Check übrig.
- Unter dem Compliance-Check jetzt die Snapshot-Historie (neuer
SnapshotHistoryList): neuester oben + farblich markiert, Klick → Detail-
Seite mit den Ergebnissen. Macht /sdk/agent/snapshots erreichbar.
- ComplianceCheckTab zeigt nach dem Lauf keine Inline-Ergebnisse mehr,
sondern einen Hinweis auf die Historie (onComplete refresht sie).
- Tote Komponenten gelöscht (ScanResult/BannerCheckTab/AgentTestTab).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- architecture.go: DataSources now reflect the real ingested set (ESAW 2023,
BLS CFOI, OSHA OTM, PRISM, cobot CC-BY, HSE) with their RAG collections;
risk stage cites BLS + the searchable RAG layer; matrix stage now mentions
the distance-benchmark dimension.
- Architektur & Datenfluss tab: new DataFlowDiagram — 4 lanes (input →
knowledge/RAG-evidence → deterministic engine → outputs) with live counts.
- scripts/ingest_iace_kb.sh: idempotent E1 ingest — creates the 2 collections
and uploads the 6 datasources docs against a configurable RAG_URL (for prod
Qdrant), with retry.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Die Deklaration-vs-Bibliothek-Sicht deckte sofort einen Fehl-Match auf:
'cct_chatSessionToken' (Genesys-Webchat) traf die Library-Basis 'cct'
(actual_category Marketing, purpose 'shopping cart') → falsches
'necessary→Marketing'-Finding. Ursache: gekürzte 3-Zeichen-Basis ohne
führenden _.
_is_distinctive_base: gekürzte Präfix-Basis nur akzeptieren bei ≥4 Zeichen
ODER führendem '_' (kanonische Cookies wie '_ga'). GTM-/AdobeOrg-/Hash-
Suffix-Stripping bleibt erhalten (Tests grün), generische 'cct'/'sid'/'gtm'
über-matchen nicht mehr.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Für die Library-getroffene Teilmesse (~32%) pro Cookie die Feld-
Abweichungen deklariert→Library (Kategorie/Laufzeit/Zweck) als Diff-Karte,
plus ehrlicher Funnel (gesamt → geprüft → abweichend) — nicht-getroffene
Cookies sind nicht prüfbar (kein Pass/Fail), passend zur Tonalität.
- analyze_cookies: 'expected'-Soll-Wert an tracker_as_necessary/
excessive_lifetime/missing_purpose (+ _CAT_LABEL_DE).
- neues cookie_declaration_diff.build_declaration_diff: reine Regroup-
Aggregation der Findings pro Cookie (single source = analyze_cookies),
Hinweis-Typen (third_country/eu_alternative) bewusst ausgeschlossen.
- cookie-check exponiert out['declaration_diff'].
- CookieDeclarationDiff.tsx oben im Cookie-Tab (vor Panel/ResultView).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cross-checked cobot_biomech_limits.md against both source papers:
- Behrens et al. 2022 (Frontiers): 10 body regions spot-checked, force
values match the paper EXACTLY in both columns (pinching + impact).
- Park et al. 2019 (PLOS ONE): lowest/highest/range pressure values exact.
Fix: 28 -> 29 body locations; add a verification stamp. Threshold VALUES
were already correct (no data change), so no RAG re-ingest needed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
#3 Storage-Filter: cookie-check exponiert per-Cookie-Speichertyp
(storage_inventory.per_cookie); CookieResultView bekommt Filter-Chips
(Cookie/Local Storage/Framework …) + eine Speicher-Spalte, Anbieter ohne
passenden Treffer werden ausgeblendet, KPI zeigt gefilterte Zahl.
A-Routing: legal_notice ist jetzt ein kanonischer Doc-Type. Eigene
Discovery-Regel (legal-disclaimer/rechtlicher-hinweis) VOR impressum →
die Disclaimer-Seite wird nicht mehr als Impressum substituiert (Ursache,
dass die Cross-Doc-Reconciliation nie zündete). capture-only: als
doc_entry für B persistiert, aber nicht einzeln gescort (keine 0%-Noise,
da ohne eigene Checkliste). Im Scan-Form als Option auswählbar.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Cookies werden je Vendor nach Name dedupliziert (Consent-Phasen-Dubletten;
BMW 2196 → ~772) — in cookie-check + get_snapshot, behebt aufgeblähte
Kachel-/Finding-Zahlen.
- Impressum-Snapshot-Check überspringt den ~40s-LLM-Schritt (context skip_llm)
→ Tab lädt sofort statt leer zu bleiben.
- Vendor-Tabelle zeigt nur die Cookie-Zahl (kein 'Cookies'-Wort je Zeile).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Versioned, license-tagged source docs for the multi-layer GT knowledge base,
ingested into the new core RAG collection bp_iace_safety_kb (whitelisted in
the RAG search handler):
- prism_risk_methodology.md — OPSS PRISM v2 (OGL v3): full severity(4)×
probability(8) → risk-level matrix (Serious/High/Medium/Low), RAPEX-aligned.
- cobot_biomech_limits.md — CC BY 4.0 papers (Behrens 2022 / Park 2019):
force (N) & pressure (N/cm²) pain thresholds by body region (the data behind
ISO/TS 15066, cited from the open papers — standard tables NOT reproduced).
- hse_example_risk_assessments.md — HSE (OGL v3): qualitative hazard→control.
- osha_robot_safety.md — OSHA OTM (public domain): 250 mm/s teach anchor,
robot hazard taxonomy, safeguarding hierarchy.
No DIN/EN/ISO/IEC/DGUV content reproduced; each doc states its license + attribution.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Surface result.distances in the benchmark module: a DistanceComparison
panel showing agreement %, covered values (green), GT-only gaps (amber)
and engine-only extras — mirroring the RiskComparison panel.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
US severity anchor complementing ESAW: BLS Census of Fatal Occupational
Injuries (public domain), event/exposure distribution 2023-24 + the
machine-relevant "Contact incidents" breakdown (struck/caught/compressed
by running powered equipment: 226/213). Key finding: in MANUFACTURING,
contact is the leading fatal event (104/353 = 29.5%) — independent support
for the model's mechanical-contact emphasis. Ingested into the core RAG
collection bp_iace_accident_stats.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ein 'X fehlt'/'zu prüfen'-Finding wird unterdrückt, wenn die Pflicht in einem
ANDEREN Snapshot-Dokument erfüllt ist (z.B. § 36 VSBG / OS-Link stehen bei BMW
in AGB/'Rechtlicher Hinweis', nicht im Impressum → war False Positive).
Konservative Allowlist (impressum: verbraucher_streitbeilegung, odr_link) gegen
False-Reconciliation. Verdrahtet in _run_doc_agent (alle Doc-Checks). Frontend:
'In anderem Dokument abgedeckt'-Sektion. Greift voll nach Scan + Legal-Capture.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
load_big_library matchte nur EXAKT → nur ~27% der BMW-Cookies trafen die
Open-Cookie-DB, weil Per-Instanz-Suffixe abweichen (_ga_GTM-XYZ, AMCVS_###@
AdobeOrg, _pk_id.5.7d8). Jetzt: Library einmal laden, Namen entwildcarden,
über _candidate_keys (exact + Präfix an Trennzeichen, Mindestlänge 3 gegen
Über-Match) matchen. Reuse der bewährten _strip_wildcards-Logik.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
C1: drop the misleading OSHA §1910.212(a)(5) fan-guard citation from M602
(overhead lift clearance) — EN 349 + EN ISO 13854 already cover it.
C2: frame M237's 25/500 mm as Richtwerte to be determined per EN ISO 13854
(single factual values in prose are facts, not table reproduction — but
keep the conservative caveat).
C3: keep ergonomic W=2 deliberately and document why — ESAW ranks it the most
frequent non-fatal mode (24.7%) but that population doesn't transfer to an
acute machine point-hazard; the machine GT governs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adding M605 (drive-limited general speed) and M606 (limited descent on
energy loss) to the library wasn't enough — measures only get suggested
if a pattern's SuggestedMeasureIDs references them. Add M605 to the three
lift crush patterns and M606 to the floor-stop descent pattern (HP2100),
so a re-seed actually attaches them and the distance benchmark closes the
≤150 mm/s gap.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CompareBenchmark now also compares the engine's numeric dimensions
(mm gaps, mm/s speeds) against the professional's GT measures: parses
distance tokens from both sides (German thousands/decimal aware),
reports matched / gt_only (gaps) / engine_only + an agreement %.
Surfaces as result.distances on the existing benchmark endpoint.
Deterministic, no LLM. On the GT-derived seed sessions it mainly guards
DRIFT; its real value is new sessions. Real-GT test pins that the engine
covers the Bremse (250 mm/s, 250/850 mm) and Kistenhub (25/120 mm,
150/75 mm/s) headline dimensions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The GT distance benchmark surfaced three Fachmann lift values the engine
carried no measure for: general lift/lower speed (≤150 mm/s), the low-zone
inching regime (<200 mm floor clearance, ≤75 mm/s), and limited descent on
power loss (≤100 mm). Extend M603 (inching) and add M605 (drive-limited
general speed) + M606 (load-holding on energy loss). Values framed as
generic hoist recommendations with EN 1570-1 reference, not GT-memorised.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The Risikobewertung page only mentioned the data sources as static prose.
Add a collapsible "Datenquellen & Evidenz" panel sourced from
/iace/risk-data-sources: the real Eurostat ESAW 2023 contact-mode shares
per mode, with license + ready-to-print attribution, and the note that
tiers anchor the ordering while values stay GT-calibrated.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Linter: FORBIDDEN_OUTPUT_TERMS per Wortgrenze → 'Schutzgarantien'/'geeignete
Garantien' (Art. 46) passieren, 'garantiert'-Claims bleiben geblockt.
- DSE: L2-Detail wird übersprungen statt 'na', wenn die L1-Pflichtangabe fehlt
(kein irreführendes 'nicht anwendbar' für z.B. Transfermechanismus).
- DSE: Drittland → HIGH bei dokumentiertem Drittlandtransfer (scan_context via
AgentInput.context) — BMW (Konzern, US-Provider) ist kein weiches MEDIUM.
- DSE: Titel/Maßnahme kurz (treibt den Recommendation-Titel); ausführliche
Begründung als evidence — behebt 120-Zeichen-abgeschnittene Überschriften.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Makes the OSHA minimum-distance anchor visible per measure in a project
without a DB schema change or re-seed: persisted mitigations store the
measure NAME verbatim (not the catalog ID), and measure names are unique
across the 578-entry library (pinned by test), so a name→ID resolver
bridges the gap.
Backend: MeasureIDByName + MinimumDistancesForMeasureName/LinksForMeasureName;
/iace/minimum-distances now accepts ?measure_name=; link table enriched with
measure_name for one-request UI matching.
Frontend: useMinimumDistances loads the link table once and keys it by name;
OshaDistanceNote renders the anchor (value/CFR/license/EU-hint/relation) on the
matching measure group in the Maßnahmen tab.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
DSEAgent wrappt die existierende ART13_CHECKLIST (33 kuratierte Pflichtangaben
L1 + Detailchecks L2) → strukturierter AgentOutput, NICHT der 90k-Library-
Firehose (eCall/Gesundheit/Telekom-Lärm). GET /snapshots/{id}/dse-check spiegelt
impressum-check; doc_input_from_snapshot generalisiert. Frontend: generischer
AgentModuleTab (lazy → AgentResultTab) für Impressum + DSE; DSE-Tab in der
Snapshot-Seite. Plus HRB-Pattern \d→\d+ (volle Registernummer als Beleg).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
_match_value gibt genau den gematchten Bereich zurück (nur die E-Mail unter
Email, nur die USt-IdNr, nur die Telefonnummer) — nicht mehr ein Fenster/den
umgebenden Satz. Behebt die Wiederholung desselben Anfangssatzes bei Texten
ohne Zeilenumbrüche (BMW = ein Block).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Executes the accident-statistics pipeline for the risk anchors:
- Refresh contactModeEvidence with real Eurostat ESAW figures
(dataset hsw_ph3_08, reference year 2023): impact 24.0%/21.4%,
struck-by 13.0%/23.8%, sharp 14.5%, trapped/crushed 13.8% (fatal),
+ new physical/mental-stress mode 24.7% → ergonomic. GT-calibrated
tier VALUES unchanged; the real data confirms the ordering.
- Add the versioned source document (datasources/esaw_accident_stats_2023.md,
ESAW CC BY 4.0 + OSHA public-domain context) that is ingested into the
core RAG collection bp_iace_accident_stats for searchable evidence.
- Whitelist bp_iace_accident_stats in the RAG search handler so seeding
can full-text search the statistics with citation at seed time.
Two-layer design: the small license-tagged code table stays the deterministic
tier/citation lookup; the RAG holds the searchable source evidence.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Drittland: unbekannte Herkunft ('N/A') + Self-Hosting feuern nicht mehr —
First-Party-Session-Cookies (PHPSESSID/JSESSIONID) waren False Positives.
- Impressum _line_of: enges Fenster um den Treffer bei Texten ohne Umbrüche
(BMW = ein Block) → jede Pflichtangabe zeigt IHREN Beleg statt denselben Satz.
- Neuer Finding-Typ missing_opt_out: einwilligungspflichtiger Anbieter mit
Cookies ohne Opt-Out-/Widerspruchs-Link (Art. 7 Abs. 3 + Art. 21).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Snapshot-Detailseite wird zu Modul-Tabs (Cookies & Tracking | Impressum).
Backend GET /snapshots/{id}/impressum-check laeuft den v3 ImpressumAgent auf
dem gespeicherten Impressum-Text (kein Re-Crawl); Input-Erzeugung in
impressum_input_from_snapshot() ausgelagert (pure + getestet: Text/Scope/
company_name-Fallback/None-Pfad). Frontend laedt lazy beim Tab-Wechsel und
rendert mit dem bestehenden AgentResultTab (keine zweite Engine).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The May-built OSHA distance library (minimum_distances.go, 29 CFR 1910,
US public domain) was dead code — zero callers, no route, no test, while
the mm values that actually appear in measures are independent hand-prose
(some carrying ISO 13854/13857 values, not OSHA).
This surfaces it without touching the measures response contract:
- GET /iace/minimum-distances (+ ?measure_id=) returns the distances, the
curated measure→distance link table and the licensing note.
- AllMeasureDistanceLinks/MinimumDistancesForMeasure resolve only the
defensible links (M600 value_source; M254/M065 public-domain crossref to
ISO), with the relation made explicit so the join stays honest.
- architecture.go lists the OSHA library so it shows in the audit explainer.
- Tests: inch→mm conversion + license completeness, link integrity, and a
consistency test pinning that a value_source measure's prose still
matches the OSHA source (codifies the audit finding as a regression gate).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CookieResultView bekommt einen Umschalter [Rechtliche Rolle] ↔
[Banner-Kategorie] (Notwendig/Funktional/Statistik/Marketing). In beiden
Sichten zeigt jede Cookie-Zeile '→ sollte: Marketing', wenn die tatsächliche
Kategorie laut Library von der deklarierten abweicht (rot bei Tracker als
notwendig, § 25 TDDDG). Neue KPI 'Falsch einsortiert'. Backend liefert dazu
cookie_categories (name→actual_category) aus big_lib im cookie-check-Output;
Seite lädt cookie-check einmal und reicht es an beide Komponenten.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
app/api/webhooks/woodpecker/route.ts (committed in 529c37d) imports
WoodpeckerWebhookPayload, ExtractedError + BacklogSource from
@/types/infrastructure-modules, but that file was never committed. Clean
checkouts (Docker build, CI) fail with 'Cannot find module'. Restore the
file so the admin build is green again. Pure type declarations, no logic.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Vendor-Ebenen-Finding: greift, wenn ein Vendor eine Verarbeitung deklariert
(Kategorie/Zweck), aber KEINE Cookies gelistet sind UND keine persistence
angegeben ist (z.B. Nayoki GmbH — 'necessary' Auftragsverarbeiter ohne
Löschfrist). Die Pro-Cookie-Schleife sah solche Vendors nie (0 Cookies →
0 Findings). Remediation = Ticket-Text 'bitte Löschfrist festlegen'.
Art. 5 Abs. 1 lit. e + Art. 13 Abs. 2 lit. a → Control AUTH-2051-A03.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds an auditor-facing view of the IACE engine: a clickable 10-stage
pipeline flow (Grenzen-Formular → ParseNarrative → Pattern-Gates →
Relevanz → Caps → Gefährdungen → Maßnahmen → Risiko → Normen → Matrix),
plus live library counts, the data-source/license register (incl. the
DIN/Beuth + DGUV exclusions), and the norm-matching logic that reconciles
DIN/ISO/OSHA machine-type vocabulary via canonicalMachineType folding.
Backend: BuildArchitecture() with LIVE counts so the diagram can never
drift; GET /iace/architecture; collectAllNorms() extracted from
SuggestNorms as the single source of truth for the norm-library count.
Frontend: useArchitecture hook + page + new IACE nav tab.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cookie-Check-Endpoint liefert jetzt out["drift"] (audit_cookie_compliance):
deklariert (Cookie-Richtlinie-Text) vs. tatsaechlich geladen (Browser).
Frontend zeigt den Reality-Check-Strip oben im Panel: X dokumentiert ·
Y geladen · Z undokumentiert. Pinnt den Vertrag mit test_cookie_drift.py
(undokumentiert-geladen + beide Drift-Richtungen) + Vitest Drift-Strip.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Surfaces the public-statistics provenance for the contact-mode probability
tiers so generated risk numbers are auditable and attributed (not RAG —
~a dozen stable aggregate facts are better as a license-tagged code table).
- risk_data_sources.go: RiskEvidence register (Eurostat ESAW figures + CC BY
4.0 attribution) for the documented contact modes; RiskDataSourcesNote.
- risk_suggestion.go: the W justification now cites the actual ESAW share +
license where documented; RiskSuggestion gains a data_source field.
- GET /iace/risk-data-sources returns the evidence register + attribution.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds GET /projects/:id/risk-matrix — a confidence-aware risk view computed
on read from each hazard's category/scenario/lifecycle using the SAME model
as the GT benchmark (no persistence, so it never goes stale against the
model; the hand-defaulted iace_hazards risk columns stay untouched).
- risk_matrix.go: EstimateHazardRisk (single source of truth for S/F/W/P +
range + level + confidence) and BuildRiskMatrix (per-hazard list + a 5×5
Severity×Probability aggregation grid with dominant level per cell).
- Frontend: RiskMatrix grid in the Risikobewertung tab (muted colours per
the confidence-aware tonality), level counts + tool-confidence summary,
fed by useRiskMatrix. Shows risk for EVERY project, not only GT ones.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Jeder Cookie-Befund traegt jetzt ein strukturiertes control-Feld
(control_id aus doc_check_controls + regulation + article) statt nur
hardcodeter Strings: vague_duration->AUTH-2051-A03 (Art.5(1)e+13),
tracker_as_necessary->DATA-2851-A05 (§25 TDDDG), third_country->
DATA-1624-A04 (Art.44). Kette Regulation->Article->Control->Finding.
Frontend zeigt die Rechtsgrundlage je Befund. (Controls tragen
regulation/article noch NULL -> hier mitgeliefert bis gepflegt.)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Flaggt Laufzeit-Angaben ohne konkrete Dauer/Kriterium ('dauerhaft', 'bis zur
Loeschung', 'bis Nutzer deaktiviert', 'unbegrenzt' …) — Art. 5(1)(e) + Art. 13
DSGVO. Library-unabhaengig, gilt fuer ALLE Cookies (Coverage auf BMWs 780).
'13 Monate'/'Session'/'bis Widerruf, max. X' bleiben ok.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
These were pre-existing failures (stale tests, not source bugs):
- getNextStep walks steps ordered by `seq`, not array order (ai-act seq 350
sits before import 400). The tests assumed array order; derive the
expectations from the seq-sorted sequence instead.
- buildDocumentScope: a document required only by the level matrix is
`mandatory` but may be `medium` priority — only trigger-mandated docs (and
the high-priority doc types) are forced to high. The test wrongly asserted
ALL mandatory docs are high; now it checks the trigger-mandated ones.
Full vitest suite: 414/414 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
recipient_type=CONTROLLER (Meta/LinkedIn/Criteo) gehoert zu Art. 26
(eigenverantwortliche Dritte / Joint Controller), nicht zu den eigenen
Verarbeitungen. BMW: 58 eigene / 16 AVV / 7 joint / 2 sonstige (= Mail-VVT).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Eliminate the pre-existing TS errors that were masked by
next.config.js `typescript.ignoreBuildErrors: true`, then turn the flag
OFF so the compiler is a real safety net for future changes. `next build`
and `tsc --noEmit` now pass with 0 errors.
The errors were not cosmetic — several exposed real latent bugs hidden by
the flag, e.g. the drafting-engine ConstraintEnforcer read non-existent
fields (`t.rule.dsfaRequired`, `d.required`, `r.title`), so its DSFA hard
gate and risk-flag checks were silently no-ops; scopeDefaults read
snake_case CompanyProfile fields that never matched the camelCase type
(generator defaults never populated). Both fixed by aligning code to the
current types.
Highlights:
- Vitest globals: add vitest-globals.d.ts (config already had globals:true)
so the test files type-check; exclude Playwright specs from vitest.
- Add a minimal ambient `pg` module declaration (no @types/pg installed).
- Fix Next 15 route handlers to await Promise params.
- Reconcile drifted types across loeschfristen, compliance-scope, document-
generator, drafting-engine, vendor-compliance, agent and more.
Pre-existing (NOT caused here, proven by stashing the diff): 3 vitest
logic tests still fail — getNextStep (2) and buildDocumentScope priority (1).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
RemediationPlan: aus den offenen Punkten (result.results, Haupt-Engine) je
Finding eine Massnahme + fertigen Ticket-Text ableiten, nach Prioritaet
sortiert, mit Kopieren + JSON-Export als Uebergabe. SCOPE: BreakPilot
formuliert nur — Ticketsystem/Jira/Feedback-Loop baut ein anderes Team.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
ResultSummary: Titel (Firma aus extracted_profile) + check_id + 4 Kacheln
(Dokumente, Konform, Offene Pflichtangaben, Zu pruefen), gerechnet aus
result.results. Co-Pilot-Ton: gruen/gelb/rot nur bei echten Werten.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Report the tool's risk number as a plausible range with a confidence
label instead of a false-precision point value (confidence-aware
tonality — the assessment is confirmed by the DSB / safety expert).
- risk_estimation.go: EstimateConfidence (hoch/mittel/niedrig from how the
contact mode resolved), EstimateRiskRange (S±1 and aggregate L=F+W+P ±1,
the empirically validated per-parameter accuracy), RiskLevelRange; share
the riskBandLabel thresholds with EstimateRiskLevel.
- risk_benchmark.go: RiskComparisonPair gains eng_risk_point/low/high +
level + level_range + confidence; RiskAgreement gains high_confidence_pct.
- RiskComparison.tsx: per-hazard range "low–high (level range)" + point,
confidence chip, and an aggregate confidence line; types in useBenchmark.ts.
- Unit tests for the range/confidence helpers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Domain-gate ~15 foreign machine classes (pool, amusement, paint booth,
tank farm, reactor, lathe/chips, saw, film/carton, robot, mobile cab,
asbestos, playground swing) in pattern_domain_gates.go so ungated hazard
patterns stop leaking into unrelated machines; matching emit keywords
added in keyword_dictionary.go (gate+emit share one vocabulary).
- Extend the cross-domain precision guard to 6 machine classes (press,
cobot, motor, welding + the 2 GTs) with per-case homeDomains, so a
machine's own domain terms are never flagged. GT coverage stays 100%.
- Reconcile the fine-grained norm machine-type vocabulary (455 keys) with
the 68 canonical dropdown keys via canonicalMachineType() family folding
in matchNorm — welding 0->17, robotics_cobot 0->6, press 8->13,
circular_saw 1->35 machine-specific C-norms. Pattern gating left strict.
- Fix initialize?force=true summary index-shift that mislabeled counts
(reported matched-patterns as "hazards"); now uses named step vars.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Resolved .claude/rules/loc-exceptions.txt: removed the temporary
iace_handler_init_helpers.go exception — the file is now split to 455 lines
(< 500) in commit afb3f83, so the exception is no longer needed (per the note
the other session left on that entry).
[guardrail-change]
- Components view: three presence sections (Vorhanden / Nicht vorhanden /
Geloescht) with bidirectional move + soft-delete (audit-visible, restorable),
so the expert corrects the engine's best-effort negation in both directions.
- CE marking per component (bought robot/actuator/SPS) with a clear
"validate the integrated safety function (PL/SIL)" note when also safety-relevant.
Safe semantics: hazards are not suppressed, only provenance is surfaced.
- Project-create form: machine type is now a grouped dropdown from the engine's
controlled vocabulary (GET /machine-types) instead of free text.
- Knowledge graph: component→hazard edges use the real component_id.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Engine precision (stop foreign-machine patterns leaking into a project):
- Wire project.MachineType into the engine machine-type gate (empty input no
longer fires every machine class — press/cnc/excavator/crane/medical...).
- Capability-domain gating extended by 7 domains (outdoor, ventilation,
machining, bulk, palletizer, playground, fitness) so domain-specific hazards
only fire when the narrative names that domain; emitted via keyword_dictionary.
- Relevance backstop moved into iace (single gating contract, testable), and its
dominant false-anchor class removed (a long pattern word no longer matches a
short common token; prepositions/leitung added to the generic stoplist).
- New guard tests: TestCrossDomainPrecision (full pipeline, 0 foreign per GT) and
TestPatternReachability now asserts 0 dead patterns. Both GTs keep coverage 1.0.
Reachability fix: the 51 dead patterns required electrical/pneumatic/hydraulic
tags nothing produced — renamed to the canonical electrical_energy/
pneumatic_pressure/hydraulic_pressure/hydraulic_part.
Component review (negation is best-effort + expert-correctable):
- Parser surfaces negated components (ComponentMatch.Negated) instead of dropping
them; negated contribute no tags/energy → no phantom hazards.
- presence_status (vorhanden|nicht_vorhanden|geloescht) + ce_marked on components;
only `vorhanden` feed matching. CE+safety-relevant flags the PL/SIL obligation.
- Force re-seed preserves the expert's component decisions instead of wiping them.
- Tag-based component→hazard assignment (was: all on the first component).
- Negation-aware narrative parsing ("keine Pneumatik" no longer extracts it).
Local-dev DB: ai-sdk sets search_path=compliance,core,public; reconcile migrations
152-156 bring the consolidated local iace tables to the current schema + add the
presence_status/ce_marked columns. Machine-type vocabulary endpoint for the form.
[migration-approved]
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The guard probes mc_use_case_mappings as the existence sentinel, but the route
also queries mc_verification, mc_regulations and mc_use_case_sync_state. Document
that they are seeded together and that a half-seeded DB (sentinel present, a
sibling missing) still 500s on the sibling's queries.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
test-go (go vet runs as part of go test) failed on two pre-existing iace spots:
- cmd/iace-audit/main.go: 6x fmt.Println with redundant trailing \n
- internal/iace/document_export_sources.go: duplicate `r == ';'` clause
build-sha-integrity failed because the alpine job installs python3 but not
pyyaml, so `import yaml` raised ModuleNotFoundError. Add py3-yaml to apk.
loc-budget flagged iace_handler_init_helpers.go (530 lines, committed state).
The other session already split it to 455 in the working tree (uncommitted);
grandfather it until that split lands, then remove the exception.
Verified locally: go test ./... all ok, go vet clean, check-loc.sh exit 0.
[guardrail-change]
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pre-existing tech-debt file (~535 LOC in the CI tree) that grew past the
500-line hard cap and has blocked the repo-wide loc-budget check since #657.
Not related to the IACE work in flight. Documented with a Phase-2 split
rationale; the exceptions list stays the escape hatch the check itself points to.
[guardrail-change]
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 2: Live-Filter an /sdk/master-controls (Use Case, Quell-Regulierung,
Verifikations-Methode, Coverage, Primärzweck-Toggle, category via Member-EXISTS).
API mit EXISTS-Filtern + gecachten Meta-Counts in master-controls/route.ts.
Phase A: neue UseCase telekommunikation + Fix der Impressum-Fehlrouten im
Register (TKG/AT-TKG->telekommunikation, telemedien->dse, GewO->handelsrecht);
echte Impressum-Quellen (TMG/Mediengesetz) bleiben impressum. Deterministischer
Seed aus source_regulation; Tests grün.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosis: engine F mean 3.56 vs professional 2.56; the dominant disagreement was
normal-operation hazards getting F=4 where the professional assigned 2. Lowered
the lifecycle→F mapping (normal operation 4→3, occasional phases 3→2). New
TestGT_RiskComparison_CrossGT runs the exact production comparison on BOTH GTs:
F within±1 rose to 95% (robot cell) and 94% (lift) — generic, not lift-tuned.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The "Zugeordnet" tab already expanded to a GT-vs-Engine detail comparison; the
"Fehlend" and "Engine Findings" tabs were flat and could not be inspected.
Extracted GTDetailBlock / EngineDetailBlock from DetailComparison and made both
tables expandable (chevron) — missing rows show the full GT entry, extra rows
show the full engine hazard (incl. measures, norms, clarification status).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
#1 Risk-number comparison in the benchmark: ComputeRiskComparison derives the
tool's S/F/W/P + Fine-Kinney per matched hazard and compares to the GT values;
exposed on the benchmark response and rendered in a new RiskComparison table
with GREEN/YELLOW/RED traffic lights on the risk number R (like the Excel),
plus per-axis within-1 agreement cards.
#2 Generic misuse pattern HP2103 "Personenbefoerderung auf Hebezeug" — gated to
lift-family machine types, fires for ANY lifting device (not machine-specific).
#3 Benchmark matcher is now 1:n — one broad engine hazard may cover several
fine-grained GT sub-scenarios (foot/hand/leg crush), so coverage reflects real
risk coverage rather than 1:1 wording matches.
Validated on BOTH ground truths (robot cell + lift): leakage 0, ghosts 0,
coverage held.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Member-canonical_controls tragen meist kein evidence_type/verification_method
(wie schon source_citation). primary_verification_method() leitet die Methode
deterministisch aus dem Primaer-Use-Case ab (impressum->document,
code_security->source_code, ...). Populiert mc_verification beim naechsten Seed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
(1) extractNarrativeFromMetadata now reads every limits-form field generically
(no whitelist) — intended use, foreseeable misuse, all machine limits and all
four interface groups (electrical/mechanical/pneumatic/software). Field-schema
drift no longer silently drops hazard sources.
(2) withUniversalLifecycles always adds normal_operation/setup/maintenance/
cleaning to the matched lifecycle phases — these occur on virtually every
machine and the professional assesses them, so their hazards must be derived
even when the form omits them.
Kistenhubgeraet recall jumped 42.9% -> 74.3% (electrical 9% -> 82%) from the
field-name fix alone; this broadens it further.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
extractNarrativeFromMetadata looked for field names that don't exist in the real
limits-form schema (interfaces_description, control_system_description,
energy_sources, space_limits, foreseeable_misuse), so it effectively read only
general_description + intended_purpose. The electrical/mechanical/pneumatic/
software interface fields — each a hazard source — were silently dropped, which
is why electrical hazard coverage was 9% for the Kistenhubgeraet.
Now reads the actual schema fields incl. electrical_interfaces /
mechanical_interfaces / pneumatic_hydraulic_interfaces / software_interfaces /
energy_supply / spatial_limits / foreseeable_misuses, plus array fields
(operating_modes, person_groups, industry_sectors). Legacy names kept.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
A "Splitterflug bei Werkzeugbruch" pattern leaked into a lift re-seed because
its press hint ("Pressraum") lives in ZoneDE, which applyDomainGates did not
scan. Add ZoneDE to the gated text. Leakage stays 0, ghosts 0, coverage held.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Observed on a real Kistenhubgeraet (lift) project: generic mechanical patterns
(e.g. HP1000 "Quetschen Arm zwischen Pressenteilen") carry NO machine type and
only generic tags (crush_point, rotating_part), so they fired for a lift; the
narrow domain-gate terms missed their press/welding/glass wording.
Broadens domainGateTerms (pressenteil, pressraum, blechbearbeitung,
punktschweiss, schweisselektrod, elektrodenspalt) and adds a dom_glass domain
(glasschneid/glasbearbeitung/...) with its emit keywords. New test pins that the
four observed leakers now require a dom_* tag. Ghost=0, Leakage=0, coverage held
on both GTs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
New tab /sdk/iace/[projectId]/risikobewertung. Per hazard it shows BOTH models
side by side — EN-62061-style (S/F/W/P) and Fine-Kinney (P/E/C) — with
BreakPilot's justified suggested values from public data, the visible formula,
and editable fields that recompute the score + risk band live. The professional
adjusts the values (e.g. from his own licensed DIN/Beuth data); we only supply
the formula + inputs, reproduce no norm table.
Consumes GET .../hazards/:hid/risk-suggestion. Registered in IACE_NAV_ITEMS.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
GET /projects/:id/hazards/:hid/risk-suggestion returns BreakPilot's justified
starting values for BOTH risk models per hazard:
- EN-62061-style F/W/P/S (the Excel format the professional knows)
- Fine-Kinney P/E/C (US-recognized)
each with a plain-language justification + the visible formula. Read-only and
computed from public-data anchors (ESAW/NIOSH/OSHA via the engine estimators) —
the professional adjusts the values; no norm table is stored or reproduced.
Adds EstimateFrequency (lifecycle -> 1-5) and BuildRiskSuggestion. Go SDK has no
OpenAPI baseline, so the only contract surface is the frontend consumer (the new
Risikobewertung tab, next).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Fine-Kinney (Fine 1971 / Kinney-Wiruth 1976): Risk = Probability x Exposure x
Consequence — a PUBLISHED, freely-usable method (not a DIN/Beuth/ISO standard),
widely used incl. CE-marking. Gives the professional a second, US-recognized
model alongside the EN-62061-style one; German exporters get both for free and
adjust with their own licensed norm data.
risk_fine_kinney.go: SuggestFineKinney derives justified P/E/C from public
anchors (ESAW frequency -> P, lifecycle -> E, de-biased severity -> C on the
Fine-Kinney consequence scale) + ComputeFineKinney(p,e,c) so the professional
can override with his own values. No norm table stored.
GT benchmark (rank concordance vs the professional): Fine-Kinney 75.4% — beats
the EN-62061-style model (69.3%) and the raw engine (57%).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The engine's hand-set DefaultSeverity systematically over-estimates severity
(GT shows crushing 3.3 vs 2.2, struck_by 3.1 vs 2.5; electrical was already
close). EstimateSeverity blends the pattern default 50/50 with the contact
mode's GT-calibrated typical severity (baseS) — keeps pattern-specific signal,
removes the bias. Our own model, no norm table.
Effect across both GTs: severity within +-1 78%->88%; risk RANK concordance
57%->69% (Kistenhub 45%->70%). Wired into iace_handler_init.go so the
BreakPilot risk line uses the de-biased severity.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Der Agent lieferte "alles gruen": _load_controls gab auf macmini nur 3 von 75
doc_type='impressum'-MCs zurueck (Sidecar mc_classification.db hat nur 4/75 als
text-matchbar klassifiziert). Tiefere Ursache: die 75 doc_type='impressum'-MCs
sind fehl-klassifiziert (60/75 canonical_scope='other'; Prefixes TRD/SEC/GOV =
Geschaeftsbriefe/Marktplatz/Bestellung, NICHT §5 TMG Website-Impressum).
Fix: Der Impressum-Agent erzeugt Findings jetzt aus seinen 12 autoritativen
§5-TMG/DDG-Pattern-MCs (mcs.py) statt aus dem verunreinigten DB-Set —
deterministisch, scope-aware, field_id = semantisches Feld. Semantic-Validator-
Demote + Massnahmen + Rollup bleiben. Die 5-Impressum-GT-Tests laufen jetzt
echt durch: 0 Falsch-Positive.
DB-Master-Controls fuer Impressum deaktiviert bis zum MC-Re-Filtering (separate
Aufgabe: die doc_type-Klassifizierung der Vorgaenger-Session muss bereinigt
werden).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
IP/copyright fix: ComputePLr reproduced the EN ISO 13849-1 Anhang A risk-graph
decision table (S/F/P -> PLr a..e) and SeverityToS/ExposureToF its parameter
binning, emitted into every hazard description. Removed — we may not reproduce
DIN/Beuth norm logic.
Replaced with BreakPilot's OWN risk model:
- risk_estimation.go: probability (W) + avoidance (P) estimated from public,
permissively-licensed accident statistics (Eurostat ESAW, CC BY 4.0) by
contact mode, calibrated to our ground-truth corpus; own risk index + bands.
- iace_handler_init.go now emits "Risikoeinschaetzung (BreakPilot-Modell):
S F W P -> Risiko: <level>" instead of the norm PLr string.
- DATA_SOURCES.md: data provenance + license register (ESAW CC BY 4.0; BLS/OSHA
public domain; HSE OGL; DGUV + DIN/Beuth explicitly excluded).
- gt_risk_benchmark_test.go: first GT validation of risk numbers — W within +-1
99%, P 93% vs the professional across both ground truths.
Removed risk_graph_test.go (pinned the reproduced norm table).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Damit die Specialist-Agents auf vollstaendigem Website-Content arbeiten:
A — _find_dsi_links pierct jetzt Shadow-DOM (Web-Components wie Usercentrics/
Mercedes) rekursiv; versteckte (display:none) Links werden erfasst + als
Coverage-Metadatum geflaggt.
B — _expand_to_fixpoint klappt Akkordeons/Tabs/Hover-Menues in einer Schleife
auf, bis das DOM stabil ist (statt 1 Pass); erweiterte Selektoren;
Coverage-Telemetrie (Runden, expandierte Elemente, DOM-Wachstum, Shadow-/
versteckte Links) → Response + Backend-Log.
C — legacy_url_cdx.cdx_enumerate listet via Wayback-CDX-API ALLE je
archivierten URLs der Domain → findet Orphan-/Legacy-Seiten, die nie im
Slug-Raster standen (z.B. nicht mehr verlinktes /datenschutz, per Direkt-
URL noch erreichbar). Fliesst durch das bestehende Legacy-URL-Inventar.
Tests: test_legacy_url_cdx.py (6) + consent-tester/tests/test_dsi_discovery.py
(Pure-Helper + Real-Browser-Integration). Alle gruen, LOC-Gate gruen.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Regression: Der v3-Agent-Pfad baute eine parallele MC-Pipeline
(_load_impressum_mcs / _load_cookie_mcs, Roh-SELECT) und lief damit an
allen Schutzmechanismen der Engine vorbei → GOV/Branchen-MCs als HIGH bei
OEM/Zulieferer, fremde MCs (Bestellbestätigung), und action=check_question
(Fragen statt Maßnahmen im Frontend).
- Agent delegiert MC-Laden an rag_document_checker._load_controls
(P72-Scope, check_type='text', fits_doc_type/scope_requires).
- Subtraktives Sektor-Gate (SECTOR_PREFIXES) + Themen-Gate am Agent-Rand.
- action = konkrete Maßnahme (Imperativ) statt check_question.
- rag_document_checker: from __future__ import annotations (3.9-Import).
- mcs: Name-Pattern erkennt "Aktiengesellschaft" (OEM-Impressums).
- Tote GT-/Semantic-/Routes-Tests wiederbelebt (v3-Mismatch +
agent.cascade-Patch-Target). Alle 72 Specialist-Tests grün.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
User-Korrektur 2026-06-09:
(1) Begriff 'MC' steht im Projekt fuer Master-Control aus
canonical_controls (314k Eintraege, ~1.800 fuer dieses Tool). Mein
neuer Agent-Code hatte 'MC' als Abkuerzung fuer 'Machine-Check'
verwendet — Naming-Konflikt. Frontend-Methodik-Box jetzt:
- 'Pattern-Check' statt 'Machine-Check'
- Explizit: 'Diese Pattern-IDs (IMP-MC-001) sind interne Test-IDs,
NICHT die Master-Control-IDs aus der canonical_controls-DB'
- Roadmap-Hinweis: formale Verknuepfung Pattern→Master-Control folgt
Backend-Variablen mc_id bleiben technisch unveraendert (Refactor
waere gross), aber UI darf sie nicht als 'Master-Control' bezeichnen.
(2) LLM-Modell-Default war 'qwen2.5:7b' — Projekt nutzt aber das
groessere 'qwen3.5:35b-a3b' auf macmini (ENV SELF_HOSTED_LLM_MODEL).
_escalation.py default jetzt: SELF_HOSTED_LLM_MODEL als Fallback,
und Methodik-Erklaerung nennt das richtige Modell.
(3) Methodik-Erklaerung erweitert um Sprint-1.10 Semantic-Validator
und Sprint-1.11 Auto-Learning-Pattern-Library + Cross-Placement.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bug bei BMW: dsi-discovery liefert HTML-Entities ( ) als
Literal-Strings ohne Decode. Beispiel im BMW-Impressum:
'wird gesetzlich durch den Vorstand (Milan Nedeljkovic, …)'
Mein Pattern erwartet ':' / '.' / Whitespace nach Vorstand →
matched nicht das '&' → false-positive HIGH-Finding.
Fix 1 (Hauptfix): Test-Harness ruft html.unescape() vor agent.evaluate()
auf, so dass jeder Agent sauberen Text bekommt — entkoppelt von
dsi-discovery-Eigenarten.
Fix 2 (Belt-and-suspenders): Pattern duldet jetzt auch '(' direkt
nach Vorstand/Geschaeftsfuehrer (falls Decode mal fehlschlaegt).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Statt der simplen dsi-discovery-Wrapper-Funktion ruft der Test-Harness
jetzt _fetch_text() aus agent_check/_fetch.py — die VOLLE Pipeline
die auch der produktive Compliance-Check verwendet:
- consent-tester dsi-discovery mit 240s Timeout (statt 120s)
- doc_type-aware max_documents (1 für cookie/dse, 3 für impressum)
- CMP-Payload-Capture (ePaaS, OneTrust …)
- HTTP-Fallback mit Browser-User-Agent + DomainRateLimiter
- HTML-Tag-Strip wenn Playwright fail
Damit funktionieren Cloudflare-/Anti-Bot-geschützte Sites wie BMW
und Elli auch im Test-Harness — vorher Timeout nach 90s.
Plus: bei leerem Fetch klare Fehlermeldung im Slot
('Cloudflare-/Anti-Bot-geschützt — Tipp: Text manuell einfügen')
statt silent-fail. cmp_payloads landen jetzt auch im Vault.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Safetykon-Bug: 'Geschäftsführung:' (Sammelbegriff für GF einer GmbH)
matched das alte Pattern 'Geschäftsführer' nicht — False-Positive
IMPRESSUM-AGENT-VERTRETUNGSBERECHTIGTE_LABEL_KORREKT.
Pattern erweitert: Geschäftsführer|Geschäftsführung|Geschäftsführerin
+ Vorstand|Vorstandsvorsitzender + Inhaber|persönlich haftend.
Test test_safetykon_geschaeftsfuehrung_passes ergänzt (11/11 grün).
frontend: SlotCard zeigt jetzt Badge bei 0/0/0-Slots
('Dokument konnte nicht geladen werden') statt silent-fail, +
bei 0 Findings ein 'alle MCs OK'-Badge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
/app/artifacts gehört root und appuser darf nicht mkdir machen — Endpoint
crashte mit PermissionError. Default jetzt /tmp/breakpilot/agent_runs.
EVIDENCE_VAULT_ROOT-Env-Var bleibt für persistente Volumes nutzbar.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Backend registriert specialist-agent-Routes über den compliance-Router,
prefix wird /api/compliance/specialist-agent/* (statt /api/v1/...).
Frontend-Proxy hat auf /api/v1/specialist-agent/* gezeigt — 404.
Verifiziert auf macmini:
curl http://localhost:8002/api/compliance/specialist-agent/agents
→ 200 {"agents": [{"agent_id": "impressum", ...},
{"agent_id": "cookie_policy", ...}]}
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
(1) B22 Cross-Domain (fix#59):
Elli-Test fand AGB auf logpay.de NICHT obwohl URL in doc_entries
korrekt. Vermutete Ursache: Discovery-Phase A drops/überschreibt
Original-URL bei PDF-Fetch-Fail (word_count=0).
Fix: _collect_audit_urls() iteriert über state.doc_entries +
rejected_url + req.documents — Cross-Domain-Hosting ist
unabhängig vom Text-Inhalt. Plus Trace-Logging für künftige
Diagnose. Dedup per (doc_type, host_sld).
(2) B17 Audit-Walk-Fail-Fallback (fix#60):
BMW v5 hatte audit_walk=None ohne Mail-Hinweis. Vermutlich
180s-Timeout bei OneTrust-CMP-Banner-Tour.
Fix: Timeout 180s → 300s. Plus: Bei Fail wird ein Hinweis-
Stub mit error-Grund in state["audit_walk"] + HTML-Block
geschrieben — Reviewer sieht den Fail statt silent-skip.
(3) company_name + origin_domain im Backend (fix#61):
Frontend sendet seit ec03317 die zwei Felder — Backend ignorierte
sie.
Fix: ComplianceCheckRequest-Schema um company_name +
origin_domain erweitert. phase_e_email priorisiert User-Input
vor URL-Heuristik für site_name. Bei origin_domain ohne
ableitbare doc_entries-domain wird der User-Input als domain
übernommen.
(4) Plausibility-LLM Fallback-Modell (fix#62):
qwen3:30b-a3b liefert auf großen DSEs (BMW 122 FAIL) gehäuft
leere format='json'-Responses — Circuit-Breaker griff aber
Phase blieb nutzlos.
Fix: Default-Modell auf qwen2.5:7b umgestellt (4× kleiner,
zuverlässiger bei format=json, ausreichendes Reasoning für
PASS/MODIFY/DROP-Klassifikation). Plus Strategy-C eingeführt
— Fallback-Modell (llama3.2:3b) wenn primary leer bleibt.
BATCH_SIZE 4 → 3. ENV-Switches PLAUSIBILITY_LLM_MODEL +
PLAUSIBILITY_FALLBACK_MODEL für Tuning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ComplianceCheckTab.tsx bekommt zwei neue UI-Felder oberhalb des
PreScanWizard:
- Firma → z.B. 'Tesla Germany GmbH'
- Domain (Site-Origin) → z.B. 'https://www.tesla.com/de_de'
Beide werden:
- in localStorage persistiert (Hook _useCompanyOrigin.ts)
- im POST-Body als company_name + origin_domain mitgeschickt
- haben Vorrang vor LLM-extracted_profile (Backend nutzt
eingegebene Werte falls vorhanden, fallback auf Inferenz)
Datei jetzt 489 LOC (war vorher 461 + 28 für die Inputs).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ComplianceCheckTab.tsx war 519 LOC und blockte jeden weiteren Edit
(500-LOC-Hard-Cap). Drei Concerns ausgelagert:
- _document_types.ts: DOCUMENT_TYPES + DocTypeId (inkl. news doc_type)
- _compliance_storage.ts: STORAGE_KEY_*, DocState/HistoryEntry types,
emptyDocState/initState helpers, countWords
- _useCompliancePolling.ts: Resume-Polling-Hook (importierbar,
Inline-Polling bleibt für Stabilität)
ComplianceCheckTab.tsx ist jetzt 461 LOC (-58).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User-Feedback BMW v5: "740 Cookies verschwunden auf 31, Übersicht
verloren". Drei Anpassungen:
Mail-Restrukturierung (_executive_summary.py + _compose.py):
- render_executive_summary(): Top-of-mail TL;DR mit
Compliance-Score (gross + farbig), Top-3-Findings nach
Severity, Cookie-Statistik (deklariert/Browser/Drittland),
Severity-Verteilungs-Chips.
- collapsible(): wrapt jeden Block in <details>/<summary>.
Mailpit + alle modernen Mail-Clients rendern das nativ.
- _compose.py: alle 18+ B-Blöcke + per_doc + per_theme +
legacy_html in Akkordeons. NUR Critical-Findings + Sofort-
massnahmen sind immer offen — Reviewer sieht ~15 Zeilen
Übersicht und klappt selektiv auf.
- Cookie-Inventar (742) hat jetzt eigene Sektion ganz oben
(Akkordeon "🍪 Cookie-Inventar"), Vendor-Karten parallel.
B22 Cross-Domain-Legal-Doc-Detector (cross_domain_doc_check.py):
Real-Beispiel User-Feedback: Elli's AGB liegt auf docs.logpay.de
statt elli.eco. Detektor erkennt SLD-Mismatch:
- HIGH bei agb / widerruf (vertragsrelevant)
- MEDIUM bei dse / nutzungsbedingungen
- INFO bei cookie / impressum (Best-Practice)
Norm: DSGVO Art. 28 (AVV-Pflicht für Hosting) + Art. 13 Abs. 1
lit. e (Empfänger) + § 312i BGB (Cool-URLs).
9/9 Tests grün inkl. Elli/LogPay Pattern.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Erweitert Phase 1 (Backend 5-Stage Lifecycle, Migration 148) jetzt auch
im Frontend: Status-Pills, Buttons und Modal-Texte differenzieren nun
zwischen DSB- und Mandanten-Pruefung.
- WorkflowStatusBar zeigt 5 Schritte: draft -> review_internal ->
review_client -> approved -> published, mit status-spezifischen
Action-Buttons (Save/Submit, DSB-Freigabe, Mandant-Freigabe, Publish).
- ApprovalModal differenziert Mode 'approve-internal' / 'approve-client' /
'reject' mit eigenen Titles und Button-Labels.
- useWorkflowActions ruft neue Endpoints /approve-internal und
/approve-client (Backend Phase 1); approveVersion bleibt als
Backward-Compat-Alias.
- page.tsx leitet Modal-Confirm an passende Action weiter und akzeptiert
review_internal/review_client im draftVersion-Filter.
- _types.ts: Status-Union + STATUS_LABELS um beide Review-Stufen
erweitert; alter 'review'-Wert bleibt fuer Bestandsdaten erhalten.
- CompareView, SplitViewEditor, HistoryPanel: Status-Rendering und neue
Action-Labels (submitted_internal, approved_internal, approved_client).
LOC-Exception fuer admin-compliance/lib/sdk/types/sdk-steps.ts (525):
zentrale SDK-Step-Registry mit kanonischer Reihenfolge — splits wuerden
die globale seq-Garantie zerreissen.
[guardrail-change]
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drei verwandte Mechanismen für DSE-Beweisbarkeit + URL-Hygiene.
Plan B + PDF — Versions-Beweisbarkeit-MCs (dse_checks.py):
- mc-dse_version_date (HIGH) — sichtbares Stand/Versionsdatum
Pflicht. 12 Regex-Pattern: "Stand: April 2024", ISO-Datum,
"Letzte Aktualisierung", "Version 3.2", englische
Varianten ("Last updated", "Effective date as of …").
Norm: Art. 7 Abs. 1 DSGVO (Nachweisbarkeit Einwilligung).
- mc-dse_version_proof (MED) — PDF-Download oder
versionierte Archiv-URL. Reine HTML-DSE ohne Snapshot ist
juristisch fragil. 8 Pattern: .pdf, Download-Hinweis,
web.archive.org, /dse-vNNN.html.
Norm: DSK-Orientierungshilfe 2024.
Plan A — Legacy-URL-Discovery (legacy_url_discovery.py + B20):
Vier komplementäre Quellen:
A.1 /sitemap.xml + Sub-Sitemaps parsen, auf compliance-
relevante Slugs filtern
A.2 archive.org/wayback/available pro Slug — wenn Wayback
zeigt ≥18 Monate alten Snapshot UND Seite heute noch
200 liefert UND nicht im Footer → Legacy-Verdacht
A.3 Slug-Permutations: 6 doc_types × 6 Slug-Varianten ×
5 Lang-Prefixe × 4 Brand-Parameter
A.4 Banner-Modal-Links (über consent-tester Stufe 4 Tour)
Mail-Block "🗂️ Legacy-URL-Inventar" mit Tabelle: URL · HTTP ·
Wayback-Alter · Footer · Empfehlung (301/Offline/Behalten).
Engine entscheidet NICHT was Legacy ist — präsentiert das
Inventar, Kunde wählt.
Real-World-Smoke Elli:
/en/cookies → HTTP 200, Wayback 69 Mo alt, nicht im Footer
→ "Legacy-Verdacht, 301 setzen"
/en/impressum → HTTP 302, redirected → "behalten"
Plan C — Multi-Version-DSE-Analyse (multi_version_dse.py):
Wenn ≥2 DSE-URLs reachable: pro Variante DSB-Name + Datum +
Wortzahl + SHA-256 extrahieren, Inkonsistenzen flaggen
(date_divergent, dsb_divergent, no_date_count).
Mail-Block "📑 Mehrere DSE-Versionen erkannt" mit
Vergleichstabelle + rotem Hinweis "Nur eine Version kann
gültig sein". Beispiel Elli: /de/datenschutz (Mollstr-DSB,
2022) vs /de/datenschutzerklaerung?brand=elli (Proliance,
ohne Datum).
API-Response erweitert um legacy_url_inventory +
html_blocks.legacy_urls + multi_version_dse_html im V2-Layout.
ENV-Override: LEGACY_URL_DISABLED=1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Neue Compliance-Admin-Seite /sdk/document-library: zeigt alle compliance_
legal_documents mit aktueller Version, gruppiert nach Empfehlungs-Klassi-
fikation, filterbar nach Status + Volltextsuche.
Backend (Service + Routes):
- LegalDocumentService.list_documents_with_versions() — JOIN über docs +
latest/published version in einem Roundtrip statt N+1
- GET /api/v1/compliance/legal-documents/documents-with-versions
liefert {documents:[{...doc, latest_version, published_version}]}
Admin-Frontend:
- app/sdk/document-library/page.tsx (350 LOC)
- Lädt Docs + Recommend parallel
- Mapped jedes Doc per .type → Recommend-Item (klassifiziert in
required/recommended/optional/uncategorized)
- 4 Sektionen mit Klassifikations-Chip + Anzahl-Badge
- Tabelle pro Sektion: Titel · Type · Status · Version · Geändert · Override
- Status-Filter (alle / draft / review_internal / review_client /
approved / published / archived / rejected)
- Klick auf Zeile → /sdk/workflow?doc=<uuid>
- Empty state mit Link zum Generator (Bulk-Modus)
- workflow/page.tsx: auto-select bei ?doc=<uuid> URL-Param
- lib/sdk/types/sdk-steps.ts: 'document-library' bei seq=2500 im Paket
'dokumentation' registriert (sichtbar in der SDK-Sidebar)
Workflow-Hookup vervollständigt: Library → click → Workflow öffnet
direkt das gewünschte Dokument im SplitViewEditor, keine manuelle
Selektion über DocumentSelectorBar mehr nötig.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 of the workspace-cutover initiative: the Document Generator
gets a Bulk-Generate mode that produces every recommended document
in one click instead of forcing the user through 25+ per-template
clicks.
New: BulkGenerateModal.tsx (430 LOC)
- On open: POSTs current CompanyProfile + ComplianceScope answers
to /api/sdk/v1/compliance/recommend (Phase 1 endpoint)
- Matches each recommendation's document_type against allTemplates
- Shows tabular list: classification chip, title, document_type,
source citation; checkboxes pre-selected for required+recommended
(only where a template exists)
- On submit: sequentially renders each selected template using the
same pipeline as GeneratorSection (runRuleset → applyBlockRemoval
→ applyConditionalBlocks → placeholder replace), then POSTs
documents + version v1.0 draft
- Per-row progress: ⏳ generiere → ✓ erstellt / ✗ Fehler / —
übersprungen; final summary counts
page.tsx:
- Imports BulkGenerateModal
- Adds prominent "Empfohlene generieren →" CTA above the
RecommendedDocuments block
- Wires SDK state (companyProfile, complianceScope) into the modal
Profile mapper:
- CompanyProfile (camelCase): employeeCount, businessModel,
isDataProcessor → org_employee_count, org_business_model,
comp_has_processors
- ComplianceScope answers (questionId/value): pass through 1:1
since the rule system uses the same field names as the wizard
- compliance_depth_level pulled from decision.determinedLevel
End-to-end flow:
1. User completes CompanyProfile + ComplianceScope
2. Clicks "Empfohlene generieren →"
3. Reviews 25-30 prefilled checkboxes
4. Clicks "Generieren" — modal iterates, all docs land as drafts
in compliance_legal_documents + version v1.0
5. Phase 3 (next): document-library tab makes them findable
6. Phase 4 (next-next): workspace consumes these directly
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 1 of the workspace-cutover initiative: compliance becomes the
single source of truth for documents. Step one is making the existing
compliance_legal_documents workflow rich enough to express the DSB→
Mandant approval pattern that the workspace's 5-stage UI needed.
Migration 148:
- Adds CHECK constraint on status (was free-form VARCHAR20)
- Allows: draft, review, review_internal, review_client, approved,
published, archived, rejected (legacy "review" kept for backward
compat — 0 existing rows so no backfill needed)
- Adds CHECK on approvals.action with extended values:
submitted_internal, submitted_client, approved_internal,
approved_client, rejected_internal, rejected_client
- Adds 6 new columns for the richer audit trail: submitted_by/at,
approved_internal_by/at, approved_client_by/at
Service:
- New methods submit_internal_review, approve_internal, approve_client
- submit_review / approve kept as backwards-compat aliases that map to
the new methods
- reject() now reads current status to log specific rejected_internal
or rejected_client action
- _version_to_response includes all new audit fields
Routes:
- POST /versions/{id}/submit-internal-review
- POST /versions/{id}/approve-internal (DSB sagt OK → Mandant ist dran)
- POST /versions/{id}/approve-client (Mandant sagt OK → approved)
- Existing submit-review / approve endpoints stay but map through aliases
Schema:
- VersionResponse extended with optional submitted_by/at,
approved_internal_by/at, approved_client_by/at fields
This unlocks Phase 2 (Generate-All in compliance generator), Phase 3
(Document-Library tab in admin), Phase 4 (workspace cutover — drop its
own document storage and route everything through this lifecycle).
[migration-approved]
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
BMW4 zeigte 1037 UNK-Findings — die Mail wurde damit unleserlich.
Drei pragmatische Anpassungen:
1. UNK severity: LOW → INFO. Mail-Renderer zeigt jetzt nur
HIGH/MEDIUM/LOW; INFO bleibt im API-Payload + CSV.
2. UNK wird NICHT emittiert wenn Vendor=First-Party-Owner
(z.B. "BMW AG" auf bmw.de). Heuristik _is_first_party_owner
vergleicht Vendor-Name gegen Domain-SLD.
3. auto_learning threshold ≥3 Sites → ≥1 Site. Second-time-Audit
einer Site hat ihre eigenen Cookies bereits gelernt → kein
UNK mehr. Single-site Auto-Learning ist absichtlich
konservativ (Annotation, kein Truth).
Effekt: erwartete Reduktion bei BMW von 1037 UNK → ~50-100
(nur unbekannte 3rd-party-Vendoren). Mail wird lesbar, MAE-
Findings (Salesforce-as-essential) bleiben prominent sichtbar.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
KRITISCH: Mein vorheriger B19-Edit hatte send_email() versehentlich
in den _build_cookie_csv_extra-Helper geschoben (NACH dem return {}).
Mail wurde nie versendet (email_status=skipped war Folge — state[
"email_result"] nie gesetzt).
Fix:
- send_email + state["email_result"]/site_name/domain/doc_count
zurück in run_phase_e (BMW4 hat 1520 findings produziert aber
keine Mail verschickt).
- _build_cookie_csv_extra ist jetzt eine echte Modul-Funktion
NACH run_phase_e.
Plus: phase_f_persist.response.html_blocks um "cookie_coherence"
ergänzt (B19-HTML-Block fehlte im API-Schema).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stufe 4 — Cookie-Banner-Tour vor dem Accept-Klick:
- audit_walk_banner_tour.tour_cookie_banner(): öffnet Settings
(16 Phrase-Varianten), scrollt vertikal, aktiviert jedes
[role=tab], expandet jedes [aria-expanded=false] / details /
summary + 14 CMP-spezifische Selektoren. Max 35 Klicks,
Best-Effort.
- audit_walk_recorder ruft tour_cookie_banner() VOR
_try_accept_banner auf — Reviewer sieht den vollen Consent-
Katalog im Video (Vendor-Liste, Kategorien, Zwecke).
- Recorder unter 500 LOC (412+155 split).
Stufe 5 — Annotierte Screenshots pro Finding:
- finding_annotator.annotate_url(): WebKit headless, JS-Inject
eines rot-banner-Labels oben + roter Outline um das Element
(Selector oder Text-Match).
- finding_annotator.annotate_findings(): dispatched 3 Cases —
B1 Tap-Target (Anchor markiert mit "Tap-Target X×Y px"),
B16 URL-Slug-Drift (404-Seite mit "/<slug> 404"),
B13 Widerruf (Footer markiert "Widerruf-Link fehlt").
- routes_audit_walk.POST /annotate-findings (consent-tester).
- _b17_wiring ruft annotate-findings nach record_audit_walk und
speichert annotations in walk.annotations.
- audit_walk_zip_builder packt PNGs nach findings/<name>.png ins
ZIP — Reviewer hat Beweis-Bilder im Postfach.
Plausibility Circuit-Breaker:
- Nach 6 consecutive empty batches (PLAUSIBILITY_EMPTY_BUDGET=6)
bricht die ganze Phase ab statt 200 Calls zu warten. Fix für
qwen3-down + große DSE-Sites (BMW: ohne Breaker 21min, mit
Breaker ~3min).
audit_walk_zip_builder fängt walk.annotations ab und legt sie unter
findings/<fname>.png im ZIP-Anhang ab.
V2-Default:
- docker-compose.yml backend-compliance.environment.MAIL_RENDER_V2:
default 'true'. Ohne diesen Override liefert die Engine
weiterhin das alte Legacy-Mail-Layout, in dem die B-Wiring-
Blöcke nicht sichtbar sind.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 Backlog-Items aus dem Multi-Site-Briefing in einem Sprint:
1. B13 B2C-Soft-Hints — Versicherungs/Tarif/Buchungs-Marker
_B2C_WEAK erweitert um "Reiseversicherung", "Tarifrechner",
"Online-Antrag", "Flug buchen", "Stromtarif" etc.
Fängt Allianz-Reise-Chatbot (vorher False-Negative).
2. Chatbot-Policy-Discovery (chatbot_policy_discovery.py)
Probt 14 Standard-Slugs (privacypolicychatbot, chatbot-datenschutz,
ai-policy, ki-datenschutz, ...) × 5 Lang-Prefixe auf jeder
submitted Origin. Successful >300-Wort-Findings werden in
doc_texts['dse'] gemerged. Audit-Trail über
doc_entries[dse].chatbot_policy_sources.
Hebt Westfield-iAdvize-Lücke.
3. API-Response-Payload erweitert
phase_f_persist.response um extra_findings, audit_walk und
html_blocks erweitert. B-Wiring-Output (B1, B3-B18) ist nicht
mehr nur im Mail-HTML versteckt — externe Aufrufer sehen jeden
Finding. Schema additiv, legacy clients ignorieren neue Felder.
4. Plausibility-LLM Empty-Response-Fix
Resilienz-Strategie A→B→C→D:
A) format='json' (strict, default)
B) format='' (loose, _try_extract_json mit ```json-fence + prose-
wrap-Unterstützung)
C) Split-Batch-Recursion (vorhanden)
D) Give up, leeres dict (callers behandeln als skipped)
Plus _post_llm() als isolierter LLM-Call-Helper, catched
Network-Errors.
5. Specialist-Agents Phase 2 LLM (MVP) — Impressum-Agent
impressum_agent_llm.py: qwen3:30b-a3b mit § 5 TMG System-Prompt,
business_scope-hints aus profile_dict. Output identisches Schema
wie pattern-agent für ein Merge ohne API-Bruch.
_b18_wiring.py orchestriert beide Agents + deduplet nach
field_id, rendert lila V2-Block mit KB/LLM-Tags pro Finding.
Pattern-first im Dedup (deterministisch + stable).
Tests: 107/107 grün (7 Test-Suites + chatbot-discovery + b18).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Video + walk.json werden nach Aufnahme zu DSMS-IPFS hochgeladen.
Die zurückgegebenen CIDs sind manipulationssichere Audit-Anker —
Reviewer können das Walk-Video Monate später noch verifizieren und
auf Unverändertheit prüfen.
consent-tester:
- _upload_to_dsms(): Best-Effort-Upload zu /api/v1/documents
(Bearer-Token, document_type=audit_walk_video|meta). DSMS-Down
bricht den Walk nicht ab — CID fehlt einfach im result.
- record_audit_walk(): nach video.webm + walk.json erzeugt, beide
hochladen. walk.json wird re-written sodass es BEIDE CIDs
selbstreferenziell enthält.
- ENV: DSMS_GATEWAY_URL + DSMS_BEARER konfigurierbar.
backend:
- _b17_wiring._publicize_gateway_url(): DSMS gibt intern
http://dsms-node:8080/ipfs/{cid} zurück. Für die Audit-Mail
wird das via env DSMS_PUBLIC_GATEWAY (default
https://dsms-dev.breakpilot.ai) durch eine extern erreichbare
URL ersetzt.
- Render-Block: gelber DSMS-Anchor-Hinweis mit Video-CID +
walk.json-CID, beide als klickbare Links zur public Gateway.
Real-World-Smoke gegen Elli:
- Video-CID: QmbdFwtSymPuWGYYdC6eNZ1eEvVLsTYmoRRxEo5L6BXgwt
- walk.json-CID: QmWaTqwZq4KVd5wYFVAKB12uZtAosPqoG1X4m1azysXYJi
- DSMS-Upload erfolgreich, gateway_url im response
Tests: 12/12 grün (+2 für DSMS-Anchor-Render-Pfade inkl.
Internal-Host → Public-Gateway-Rewrite).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Nach jedem Compliance-Doc-Aufruf werden alle Akkordeons /
<details> / [aria-expanded=false] / Trigger-Patterns geklickt
und im Video aufgenommen.
- _expand_accordions(): 7 Selektor-Patterns, max 25 Expansionen
pro Seite, Dedup nach inner_text (verhindert Endlos-Loops bei
nesteten Strukturen). Scroll-into-view + click + 400ms warten
sicher dass das Klick-Result im Video erfasst wird.
- _visit_link(): Returns (nav_event, expand_event) Tuple. Expand
läuft nur bei HTTP 2xx + ohne nav-error.
- 1500ms post-expand wait gibt der Kamera Zeit, den finalen
Zustand mitzuschneiden.
Backend B17 render: "expand_accordions" Action wird als "5
Akkordeon/Details-Sektion(en) entfaltet" gerendert. Bei 0:
"Keine Akkordeons gefunden" (neutraler Hinweis, kein Fehler).
Real-World-Smoke gegen Elli:
Impressum: 0 Akkordeons (keine)
Datenschutzerkl: 5 Akkordeons aufgeklappt
Nutzungsbeding: 0 Akkordeons
Video-Größe verdoppelt sich (581 KB → 1.14 MB) — Reviewer sieht
jetzt den vollen DSE-Vendor-Tabellen-Inhalt im Video.
Tests: 10/10 grün (+2 für Akkordeon-Render-Pfade).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Smoke gegen www.elli.eco hat 3 Bugs offengelegt, die in den
synthetischen Tests nicht greifbar waren — Real-Texte haben
Abkürzungen, HTML-Stripping-Artefakte, andere Formulierungen.
B9 Multi-Entity-Impressum — vorher: 13 "Entities" statt 2.
- Block-Boundary jetzt HRB-Anker-basiert (jeder HRB-Eintrag
markiert eine Entity). Robuster als Legal-Form-Anker, der bei
"Programmierung der Webseite Acme GmbH" über-matchte.
- _NAME_BLOCKLIST gegen 11 typische False-Positives
(programmierung, webseite, umsatzsteueridentifik, ...).
- _LEADING_NOISE_RE strippt Email-TLD-Artefakte ("eco "),
deutsche Artikel ("Die "), URL-Fragmente.
- _USTID_PAT fängt jetzt auch die Vollform
("Umsatzsteueridentifikationsnummer der … ist DE…") über eine
zweite Pattern-Alternative mit [\s\S]{0,80}? Bridge.
- Dedup gleicher Entity-Namen — Mehrfacherwähnung in einem Doc
zählt als EINE Entity.
- Fallback auf alten Legal-Form-Anker wenn keine HRBs vorhanden
(z.B. e.V. ohne HR-Pflicht).
B14 Retention-Conflict — Anchor-Liste erweitert:
- "protokolldat" / "protokollierung der zugriffe" /
"zugriffsdat" / "zugriffsprotokoll" als zusätzliche
Logfile-Anchors (Elli's reale DSE-Wortwahl statt "Logfile").
B15 AI-Legal-Basis — kein Code-Fix. Elli's aktuelle DSE enthält
keine LLM-Provider-Erwähnung mehr; der GT-Anker (2026-06-06) ist
seither veraltet. 0 Findings ist korrekt für den aktuellen Stand.
Tests: 3 neue Real-World-Regression-Tests in
test_impressum_multi_entity_check.py::TestRealWorldElliPattern.
Combined: 75/75 grün.
Real-World-Smoke gegen Elli (HTTP→Text via crude strip):
B9: Entities 13→2 ✓, IMPRESSUM-MULTI-UST_ID → VW ✓
B13: 1 Finding (b2c_strong) ✓
B14: 0 (Elli hat aktuell nur EINEN Retention-Wert für Logs)
B15: 0 (LLM nicht erwähnt, korrekt)
B16: 3 Findings (impressum/dse/cookie Standard-Slug-Brüche) ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Raw text() queries return JSONB columns as JSON-encoded Python strings,
not as Python list/dict objects. The existing isinstance check then fails
and silently falls back to defaults — so list-valued fields like
target_markets, offerings, processing_systems, ai_systems were always
returned as their defaults regardless of stored content.
Add a JSON-decode pass over _JSONB_FIELDS before the type check.
Verified: PATCH of target_markets=["DE","EU"] now round-trips through
GET correctly. Previously the DB had the right data but GET returned
["DE"] (the default).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SQLAlchemy's text() parser treats `:name::jsonb` ambiguously when the
trailing `::jsonb` follows immediately — psycopg2 receives the literal
`:name::jsonb` string and raises a SyntaxError because `:` isn't a
psycopg2 placeholder syntax.
The fix uses ANSI CAST(:name AS JSONB) which is semantically identical
in PostgreSQL but lets SQLAlchemy unambiguously substitute the
parameter.
Effects: PATCH and POST/upsert on /api/v1/company-profile now actually
update the row. Before this fix both endpoints returned 500 (or 200
with stale data) and never persisted edits.
Files touched:
- _company_profile_sql.py (build_upsert_params / execute_update /
execute_insert): 12 JSONB columns
- company_profile_service.py: PATCH dynamic JSONB column,
audit log insert
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- tests/test_elli_gt_coverage.py: 7 Charakterisierungstests die
einen synthetischen Elli-State konstruieren und sicherstellen,
dass die 5 neuen Detektoren (B13-B16 + B9-Cleanup) genau die
erwarteten GT-IDs fangen. Regressionsschutz.
- zeroclaw/docs/audits/2026-06-06-elli-gt-coverage-sprint.md:
Sprint-Zusammenfassung mit GT-Bilanz (12/13 voll, 1/13 wartet
auf #7), Commit-Liste und Morgen-Agenda-Kandidaten.
Combined Sprint-Test-Run: 72/72 grün.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Erkennt: LLM/GPAI-System (Vertex AI, OpenAI/GPT, Claude) wird in
DSE oder Cookie-Doc auf Art. 6 Abs. 1 lit. f (berechtigtes Interesse)
gestützt — statt auf lit. a (Einwilligung).
GT-Anker (Elli AI-ACT-RISK-001): Vertex-AI-Chatbot mit lit. f
deklariert. Bei LLM-Prompt/Output-Logging + US-Transfer +
Profiling-Ähnlichkeit ist Interessenabwägung fragwürdig.
Heuristik:
- KB-basiert (chat_providers.json filter: ai_capable + LLM-Type-Hint)
- LLM-Vendor-Aliases inkl. Marken-Familien (PaLM, Gemini, GPT-4,
ChatGPT, Claude 3, Azure OpenAI)
- Absatz-Boundary-Scope: Provider + lit. f im selben Absatz
- Negativ-Filter: wenn lit. a / Einwilligung ebenfalls im Absatz →
kein Finding (Side-Purpose-Erwähnung)
- Dedup pro (doc_type, provider_id)
Severity: MEDIUM.
Norm: DSGVO Art. 6 Abs. 1 lit. a vs lit. f + AI Act Art. 50 + 51.
Tests: 17/17 grün.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Der Multi-Entity-Check fängt Elli's USt-IdNr-Lücke (VW Group Charging
GmbH hat keine, Elli Mobility GmbH hat eine), aber Entity-Namen waren
mit Header-Noise verunreinigt:
'Impressum\n\nVolkswagen Group Charging GmbH'
'eco\n\nElli Mobility GmbH'
Behoben:
- _ENTITY_PAT lässt nur Space im Namen zu (kein \s/\n mehr)
- _clean_entity_name() trimmt Header-Worte (Impressum, Anbieter, ...)
und nimmt nur die letzte Zeile vor Legal-Form-Suffix
- 11 neue Tests, davon einer mit Elli-like Impressum als
Charakterisierungs-Test
Damit ist die finale Finding-Ausgabe für Audit-Reports lesbar
('Fehlt bei: Volkswagen Group Charging GmbH') statt verunreinigt.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Beide Funktionen wurden im run_compliance_check() aufgerufen aber nicht
oben importiert — NameError landete im except-Catch-all, jeder
Compliance-Check schlug auf "failed" um.
Bug stammt aus den letzten 2 Sprints (B12 + browser-matrix Stage 1.c)
wo die Aufruf-Stelle ergänzt, der Import vergessen wurde.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the "Meine Overrides" tab in /sdk/template-rule-editor — the
mechanism by which a Kanzlei tells the system "yes, the global
recommendation says required, but for MY mandanten this is only
optional / or disabled entirely (because we have an equivalent
control elsewhere)".
Components:
- TenantOverrideList.tsx (398 LOC): tabular view with search filter,
add/edit/delete operations; one row per override showing Rule Title,
Original Classification, My Override Classification (or "Deaktiviert"
badge for disabled), Reason, Created-by/at; sticky table header.
- OverrideDialog (inline): rule picker (locked in edit mode),
classification radio group (required/recommended/optional/disabled),
mandatory reason textarea, shows the original source_citation as
context above the radio group.
- ConfirmDialog (inline): delete confirmation.
Page integration:
- New Tab system at top of /sdk/template-rule-editor:
[Globale Regeln (n)] | [Meine Overrides (n)]
- TabButton helper component (border-bottom indicator).
- loadOverrides on mount.
- handleUpsertOverride / handleDeleteOverride reload overrides after
success.
Backend integration (already in place since Phase 1):
- GET /api/sdk/v1/compliance/tenant-rule-overrides
- POST /api/sdk/v1/compliance/tenant-rule-overrides (upsert)
- DELETE /api/sdk/v1/compliance/tenant-rule-overrides/{id}
Verified end-to-end against live Mac Mini backend:
Baseline: whistleblower_policy in required (for 250_999 MA)
Add override (optional + reason): moves to optional bucket with
override_applied=true and reason concatenation
"Trifft zu: ... · Quelle: ... · Tenant-Override: required → optional (Bei meinen Tier-1-Mandanten ...)"
Delete: 204
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduces the sustainable backend replacement for the hardcoded inline rules in
admin-compliance/app/sdk/document-generator/templateRecommendations.ts.
What's in this commit (Phase 1.1 - 1.5 of the rustling-yawning-boot plan):
- Migration 147: 4 new tables
- compliance_template_rules (rule shell, document_type, current_version_id)
- compliance_template_rule_versions (lifecycle, JSONB conditions,
source_citation, change_summary, approval timestamps)
- compliance_template_rule_approvals (audit trail)
- compliance_tenant_rule_overrides (per-tenant classification overrides)
Plus partial unique index for "only one is_live=1 version per rule".
- SQLAlchemy models: TemplateRuleDB, TemplateRuleVersionDB,
TemplateRuleApprovalDB, TenantRuleOverrideDB (compliance/db/).
- Pydantic schemas (compliance/schemas/template_rule.py): full request/response
set including RecommendationRequest/Result with reasons and override tracking.
- TemplateRuleService (compliance/services/): CRUD + Lifecycle transitions
(submit_for_review/approve/publish/reject) following legal_document_service.py
pattern with _transition() helper and approval audit trail. Plus tenant
override upsert.
- RecommendationService: condition evaluator (eq, neq, in, not_in, gte/lte/gt/lt,
exists, truthy) over JSONB conditions, override application, reason generation
for human-readable explanations in workspace UI.
- 18 FastAPI routes in compliance/api/template_rule_routes.py covering rule CRUD,
version lifecycle, override management and POST /recommend evaluation endpoint.
- Seed data: 33 initial rules ported from templateRecommendations.ts in
compliance/data/template_rule_seed_data.py, written as published versions
on first seed run. Idempotent via rule_key.
Phase 1.6 (pytest suite) and Phase 2 (editorial UI in admin-compliance) follow
in separate commits.
[migration-approved]
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wizard war bisher nur im DocCheckTab eingebaut, der aber nirgends im UI
gemountet ist. Daher: alle Compliance-Checks schickten scan_context=null,
P72 Branchen-Filter wirkte nie.
Fix: PreScanWizard ins ComplianceCheckTab über die Document-Rows
gestellt. Submit-Button disabled bis alle 8 Felder (Branche, B2B/B2C,
Direkt-Vertrieb, Rechtsform, Konzern, MA, Besondere Daten, Drittland)
gesetzt sind. scan_context wird im POST body mitgesendet.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Statt EIN full-page screenshot: full-page wird per PIL in viewport-grosse
Slices geschnitten, jede ueberlappt die vorherige um overlap_px Pixel.
Jeder Cookie erscheint in mind. einer Slice, an Slice-Grenzen sogar in
zwei → Dedup nach Name eliminiert die Doppel.
Warum nicht direkt scroll-based slicing in Playwright? VW's
Cookie-Page nutzt scroll-snap / fixed-position — alle viewport-shots
kamen identisch zurueck (Header-Overlay). PIL-cut auf dem full-page
PNG bypasst das Problem voellig.
VW smoke-test (32 slices):
per-slice: [0, 0, 2, 5, 5, 3, 4, 7, 4, 3, 4, 5, ...]
103 raw cookies → 79 unique nach dedup
14 vendor records (Google 9, Adobe-Familie 17, etc.)
Jeder Slice hat eigenen Timestamp + SHA256 → ZIP-Anhang fuer
juristische Beweiskette.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VW-Loop-Iteration 1: LLM cascade lieferte 14 vendors (Lucky-Hit via
Direct-Fallback). VW-Loop-Iteration 2: 0 vendors — qwen2.5:14b
ReadTimeout auch im 420s-Direct-Fallback (50k input + 16k output
output dauert > 7min auf M4 Pro).
Fix: max_text_chars 50000 → 20000. Erfasst die ersten ~3000 Worte der
Cookie-Tabelle (Tabellen-Kopf komplett). Vollstaendige Tabelle wird
ohnehin deterministisch von parse_flat_cookie_text geparsed. LLM ist
nur fuer Vendor-Namen die NICHT in der Tabelle stehen (z.B. aus
Prosa) und Inferenz-faehiger.
Erwartung: 60-120s LLM-call statt Timeout, reproduzierbar 10-15 LLM-
Vendors → Vendor-Normalizer-Total bleibt stabil bei 20+ statt 17.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bisheriges _FLAT_ROW_RE erwartete textContent-Output (Cookie-Tabelle
konkateniert ohne Whitespace zwischen Zellen). Bei VW lieferte das
deterministische 10 Vendors / 35 Cookies, aber nur weil der DSE-Text-
Fallback unvollstaendige Tabellen-Fragmente enthielt.
Beim echten cookie-richtlinie.html Fetch (8086 Worte HTML→text) sind
die Spalten durch Whitespace getrennt — und der Regex hat 0 gematcht.
Fix: \s* zwischen jedem Anker und dem Cookie-Namen erlaubt. Direct-Test
auf VW: 0 → 60 Cookies / 16 Vendors (Google 13, Adobe-Familie 16, Meta,
Salesforce, Cloudflare, Akamai etc.).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VW-Scan-Befunde aus 0a8aa16e:
1. TCF lookup failed 5x mit: column 'source' does not exist. Korrekt:
'source_name' (siehe DELETE-Query in derselben Datei). Mit dem Fix
funktioniert das TCF-Cross-Reference fuer alle Vendors statt 0.
2. Cascade tier-1 fail loggte leere message — jetzt mit type+model+base.
3. Cascade collapse (tier 2+3 unconfigured) wird beim ersten Aufruf
geloggt damit der Operator den ENV-Mangel sofort sieht.
4. vendor_llm_extractor loggt jetzt START + 0-vendor-Return (vorher
silent skip — sah aus als waere er nie aufgerufen worden).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Diese 5 Files verletzten den Hard-Cap und blockierten jeden PR der sie
touched. Pre-existing — keine neue Verletzung. Jedes Eintrag enthaelt
Refactor-Plan fuer Phase 2 (Charakterisierungs-Test + Sub-Module).
- consent-tester/services/vendor_detail_extractor.py (675)
- consent-tester/services/consent_scanner.py (567)
- backend-compliance/.../rag_document_checker.py (559)
- consent-tester/services/banner_text_checker.py (531)
- admin-compliance/app/sdk/ai-act/page.tsx (503)
Effekt: CI exit 0 ohne Verhaltensaenderung. Die exceptions-Liste muss
laut .claude/rules/architecture.md ueber Zeit schrumpfen, nicht wachsen
— d.h. diese 5 Eintraege sind explizite Tech-Debt-Marker.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
check-rebuild-needed.sh war seit Mai funktionsfähig nur fuer 3 von 10
Containern. Die anderen 7 Dockerfiles hatten kein ARG/ENV BUILD_SHA und
docker-compose.yml hat fuer KEINEN Service den Wert durchgereicht — daher
defaultete BUILD_SHA ueberall auf "unknown" und die Drift-Check war
zahnlos.
- ARG BUILD_SHA + ENV BUILD_SHA in 8 zusaetzlichen Dockerfiles
(ai-compliance-sdk, developer-portal, document-crawler, dsms-gateway,
compliance-tts-service, docs-src, docs-site, dsms-node)
- docker-compose.yml: BUILD_SHA: \${BUILD_SHA:-unknown} in jedem build:
Block (10 Services)
- .gitea/workflows/ci.yaml: neuer Job build-sha-integrity validiert dass
jedes Dockerfile ARG+ENV hat und jeder compose-build den Arg durchreicht.
Faellt bei jedem PR/Push gegen master, der einen neuen Service oder
Dockerfile ohne BUILD_SHA einfuehrt.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- _PROCESS_INTERNAL_PATTERNS: Patterns wurden gegen lowercased Blob
geprueft, aber Case-sensitive geschrieben (TOM/AVV/SCC). Matchen
nie. Auf lowercase normalisiert.
- "Ausnahmen ... dokumentieren": Pattern war zu eng, verlangte direkte
Adjazenz. Jetzt bis zu 60 Zeichen Wortabstand.
- Test-Suite mit 22 kuratierten DSGVO/AI-Act/eCall-MC-Labels. Alle
gruen (vorher 2/22 FAIL — beide vom User explizit als Beispiele
genannt: TOM, Ausnahmen).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DSMS Stufe 3 — making the parent_cid chain useful end-to-end.
Gateway (dsms-gateway):
- /api/v1/documents/{cid}/history alias added next to the legacy
/documents/{cid}/history (history endpoint itself was already there,
just under an inconsistent prefix).
- NEW /api/v1/documents/{cid_a}/diff/{cid_b}: fetches both packages from
IPFS, computes a metadata diff (per-field old/new), and renders a
unified text diff for utf-8 payloads. Binary payloads return only
metadata diff with a "binary — compare via rendered export" note.
- 4 new pytest cases (mocking ipfs_cat): text diff, binary fallback,
fetch error, history chain depth — all green.
Frontend (admin-compliance):
- CIDHistoryModal: lazy-loads /dsms/documents/:cid/history, renders the
version chain as a vertical timeline, marks the AKTUELL entry, and
per-step exposes a "Diff zu V<n>" button that loads + renders the diff
inline (metadata table + unified text diff in a monospace panel).
- AuditTimelinePage: existing CID badge now sits next to a "Verlauf
anzeigen" link that opens the modal. Handles both Python's plain-CID
audit values and the Go techfile flow's JSON envelope {cid, filename,
size} via extractCID() helper.
This makes "show me how this CE-Akte changed between V2 and V3"
self-service in the UI instead of a curl-against-IPFS workflow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Before: archiveTechFile called dsms.Archive() and discarded the result. The
file was archived to IPFS but no audit-trail entry was written, so there
was no way to later prove "this CE-Akte export went to DSMS with CID X".
After:
- archiveTechFile is now a method on IACEHandler with access to store + gin
context, and captures the CID from dsms.Archive().
- Writes an AuditAction "tech_file_export" audit entry whose new_values
JSON carries {cid, filename, size}, mirroring the Python evidence-upload
pattern.
- Applies to PDF, XLSX, DOCX, and Markdown exports.
Plus dsms package gets 3 unit tests pinning the contract: success-CID
extraction, gateway-unreachable returns nil, 500-response returns nil.
This closes DSMS Stufe 2 (evidence side was already wired; tech-file side
was missing the audit hook). Stufe 3 next: version chains + delta view.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the loose end from IACE Phase 5 handover: the LLM FM-suggest button
existed and the backend endpoint was wired, but accepted suggestions had
no path into the FMEA worksheet.
Hook (useFMEA.ts):
- acceptSuggestion(fm, componentId): builds an FMEARow from FM defaults,
prepends to rows (sorted by RPZ), removes the FM from suggestions.
No-ops + drops the suggestion when (component, fm.id) is already in rows.
- rejectSuggestion(fmId): drops the FM from suggestions list.
Page (fmea/page.tsx):
- Suggestion cards now have explicit Uebernehmen / Ablehnen buttons.
- Counter "X Vorschlaege uebernommen" tracks accept count for the run.
- RPZ in each suggestion is colour-coded (red >200, orange >100).
- Hinweis line explains S/O/D adjustability after acceptance.
- acceptedCount auto-resets when suggesting starts or panel closes.
Tests (useFMEA.test.ts):
- 8 calculateAP cases covering AIAG-VDA 2019 boundary points for severity
10 / 9 / 7 / 5 / 3, validating the H/M/L action priority matrix.
LOC: fmea/page.tsx hits 320 (soft target 300, well under 500 hard cap).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three follow-ups to the 671-norm cross-reference matrix:
1. Tech-file renderer (Go): standards_applied section now gets a deterministic
Markdown appendix with the DIN/ANSI/GB/JIS mappings for the project's
suggested norms. Built from registry, never hallucinated by LLM. Applied
both to LLM and fallback content paths.
2. Frontend NormCrossRefPanel (Next.js): expandable row in the IACE library
norms tab now has a "Internationale Aequivalenzen anzeigen" button that
lazy-loads /iace/norms-library/:id/crossref and renders a colour-coded
table (relation + confidence). Region labels humanised (US — ANSI,
China (GB), Japan (JIS), etc.).
3. Contract tests (Go): 4 new handler tests pinning the response shape of
GetNormCrossRef and ListNormCrossRefs. Equivalent to an OpenAPI snapshot
for these specific endpoints — ai-compliance-sdk has no full OpenAPI
baseline yet (separate ticket).
Tests: 6 renderer tests + 4 handler contract tests, all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Beide rufen jetzt llm_cascade.call_with_cascade() statt direkter Qwen/OVH-
Aufrufe. Damit:
* Cache-Hit auf identische Eingaben (Valkey, 7d TTL) → ~50ms statt
4-6min beim Re-Run derselben Cookie-Doc.
* Tiered Cascade automatisch: Qwen → OVH 120B → Anthropic Claude Haiku
wenn lower-tier under confidence-threshold.
* Confidence-Scoring (JSON-parse + items_per_input_size) entscheidet ob
weiter delegiert wird.
Fallback auf alte _call_ollama/_call_ovh bleibt bestehen wenn der
Cascade-Aufruf scheitert.
Erwartete Wirkung beim 2. VW-Lauf: ~10min statt ~25min (Cache-Hit auf
identische Cookie-Doc + MC-Solutions).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Batch 6 (100): EN 1870 saws, EN 81 lift sub-parts, hearing/glove PPE,
EN 50126 railway, EN 60974 welding, EN 60335-2-x cleaning appliances
- Batch 7 (71): IEC 60601 medical family, EN ISO 19085 woodworking, safety
footwear (ASTM F2413), fitness (ASTM F2276), chainsaws (OPEI B175.1),
ISO 4254 agri remainder, acoustics ISO 3743/3745/3747
671 of 671 norms now have at least DIN mapping; ~80% have a US (ANSI/NFPA/
UL/OSHA/ASME/ASTM/SAE/NIOSH) mapping; ~40% have CN-GB and/or JP-JIS.
Added TestCrossRef_SpotChecks with 15 manually vetted region mappings
(IEC 60601 → ANSI/AAMI ES60601, EN 13445 → ASME BPVC, EN 60204 → NFPA 79,
ISO 10218 → RIA R15.06, etc.).
Next steps for follow-up work:
- Add OpenAPI snapshot for new /norms-library/crossref endpoints
- Front-end: render crossref panel on /sdk/iace norm detail page
- Tech file: auto-emit "this requirement also satisfies X in market Y" hints
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a jurisdiction-cross-reference layer to the norms library. Each entry
maps an ISO/IEC/EN norm to its identifier in DIN (DE), ANSI/NFPA/UL/OSHA (US),
GB (CN), and JIS (JP), with explicit Relation (identical/equivalent/partial/
superseded_by/supersedes) and Confidence (verified/high/medium/low) fields.
Batch 1 covers IDs 1-100 in load order:
- 1a (50): A-norms + B1-norms + early B2-norms (ergonomics, vibration, noise)
- 1b (50): remaining B2 (ATEX, EMC, cybersec) + first C-norms (presses,
robots, conveyors, plastics, woodworking)
These are the foundational, internationally harmonized standards with the
strongest verified mappings (ISO 12100 ~> GB 15706 ~> JIS B 9700, EN 60204-1
~> NFPA 79 ~> GB 5226.1 ~> JIS B 9960-1, etc.).
API:
- GET /iace/norms-library?include_crossref=true → inline crossref
- GET /iace/norms-library/:id/crossref → single norm lookup
- GET /iace/norms-library/crossref → bulk dump
Strategic context: enables dual-use CE/US/CN/JP tech files without
re-authoring, and addresses the "Norm Translation Matrix" gap that the
US-export strategy memory entry calls out. 6 batches remaining (~571 norms)
to reach full library coverage.
Tests: 6 new tests; all pass via `go test -vet=off ./internal/iace/`.
(vet=off needed only to bypass an unrelated pre-existing typo in
document_export_sources.go.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-22 09:02:05 +02:00
643 changed files with 65686 additions and 4481 deletions
Unterhaltung): freundlich + KNAPP darauf hinweisen, dass das nicht Ihr Fachgebiet ist, und
zurueck zum Thema lenken — ohne belehrend oder abweisend zu wirken. Beispiel:
"Dafuer bin ich nicht der richtige Ansprechpartner — ich bin Ihr Co-Pilot fuer Compliance,
Datenschutz und Security. Womit kann ich Sie dort unterstuetzen?"
- Erfinde KEINE Antworten ausserhalb deines Fachs, auch nicht "nett gemeint".
## Eskalation
- Bei Fragen ausserhalb des Kompetenzbereichs: Wenn die Frage harmlos ist (z.B. "Hast Du Informationen zu X?"), kurz mit Ja/Nein antworten und anbieten konkreter zu helfen. NUR bei sensiblen oder rechtsberatenden Fragen hoeflich ablehnen und auf Fachanwalt verweisen.
- Bei widerspruechlichen Rechtslagen: Beide Positionen darstellen und DSB-Konsultation empfehlen
- Bei dringenden Datenpannen: Auf 72-Stunden-Frist (Art. 33 DSGVO) hinweisen und Notfallplan-Modul empfehlen
- Bei rechtsberatenden Einzelfaellen: hoeflich auf DSB/Fachanwalt verweisen — als sinnvollen
naechsten Schritt, nicht als Abwimmeln.
- Bei widerspruechlichen Rechtslagen: beide Positionen knapp darstellen + DSB-Konsultation empfehlen.
- Bei dringenden Datenpannen: auf die 72-Stunden-Frist (Art. 33 DSGVO) hinweisen und das
@@ -158,6 +163,10 @@ Der Nutzer hat "${countryLabel} (${validCountry})" gewaehlt.
systemContent+=`\n\n## Relevanter Kontext aus dem RAG-System\n\nNutze die folgenden Quellen fuer deine Antwort. Verweise in deiner Antwort auf die jeweilige Quelle:\n\n${ragContext}`
}
if(controlsContext){
systemContent+=`\n\n${controlsContext}`
}
systemContent+=`\n\n## Aktueller SDK-Schritt\nDer Nutzer befindet sich im SDK-Schritt: ${currentStep}`
// 3. Build messages array (limit history to last 6 messages)
@@ -179,6 +188,9 @@ Der Nutzer hat "${countryLabel} (${validCountry})" gewaehlt.
messages,
stream: true,
think: false,
// Modell im VRAM halten → kein Kaltstart bei der naechsten Frage
// (Kaltstart eines 35b-Modells war die Ursache fuer "Load failed").
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.