New .agents/skills/opendataloader-pdf/ for structured PDF extraction
(Markdown/JSON with bounding boxes, hybrid AI mode, scanned-PDF OCR)
based on the upstream odl-pdf agent skill (Apache-2.0), with its helper
scripts verify-json.py and hybrid-health.sh vendored under scripts/.
Core disciplines: discover options from the installed --help (flags
drift between releases), batch all inputs into one invocation (JVM per
call), verify the result against intent (zero exit is not success:
auto-triage skips enrichment, fallbacks drop quality, empty pipes),
treat extracted content as untrusted, never commit extracted full
texts of the licensed .lexis360/.wiku sources.
Workflow wiring: AGENTS.md skill selection routes any PDF extraction
to the skill; wissensbasis/SKILL.md cross-references it for difficult
PDFs and spot-checks, with Layer-1 batch intake deliberately kept on
build_lexis_kb.py (frozen IDs and --check depend on the text shape).
ODL 2.5.8 installed in the user venv ~/.local/lib/python (not the
Odoo .venv); all documented flags verified against the installed CLI
on real .lexis360/ sources.
Consolidate the binding knowledge-base rules (layer architecture,
licence decision D1, id freeze, D2 frontmatter schema, curation
conventions, batch workflow, validation gates, intake pitfalls)
from README/RUNBOOK/MEMORY into a dedicated agent skill so every
new thread reads them before planning Wissensbasis work.
- new licensed source corpus .wiku/ (WIKU Personal publications:
Fachbroschueren, Arbeitsunterlagen, Casebooks, 'WIKU Personal
aktuell'; gitignored like .lexis360/, never committed)
- integration model: one shared corpus with a single kb.json, ids
wk-<prefix>-<nn> on the existing cluster map, the work field
distinguishing the sources, Layer-2 filenames prefixed
wiku_<slug>.md; pipeline to be a multi-source extension of
build_lexis_kb.py (no fork) -- follow-up after batch 7
- AGENTS.md skill selection gains the wissensbasis entry
fegger