LOG.md

LOG

2026-09-21

• difensore-civico-ti-scrivo: new Category I, geo-blocking (PA sites unreachable from abroad via IP-geolocation restrictions), with the supporting norms in references/normativa.md (CAD art. 9 and art. 53, D.Lgs. 36/2006 art. 8); trigger description updated
• typst-cards: nuovo principio "Gender in the copy" - ruoli e titoli prendono il genere di chi li ricopre ("una Presidente", non "un Presidente"); per i gruppi va valutata la doppia forma ("bambine e bambini") e la scelta va lasciata all'utente. Aggiunto il controllo alla Phase 4

2026-09-19

• openalex: nuove sezioni su ricerca per campo, sintassi di search= e semantica, tutte misurate contro un articolo bersaglio. La ricerca per titolo perde i lavori che parlano di un paese senza nominarlo nel titolo: title_and_abstract.search unito a authorships.institutions.country_code dà insiemi piccoli abbastanza da leggersi per intero, che è quello che serve quando il ranking non aiuta. Aggiunto il test di richiamo su un lavoro noto (cercalo per DOI nei risultati e guarda a che posto sta) come passo della validazione: è quello che ha fatto emergere un articolo rimasto al diciannovesimo posto su 468 e mai letto. La sintassi filter=campo.search: è deprecata da OpenAlex ma non ha sostituto per lo scoping su un campo, e resta l'unica strada: annotato. Misurato anche il resto: corpus=all porta il 2024 da 10,8 a 28,0 milioni di lavori, un filtro .search si paga come ricerca (0,001) e non come filtro (0,0001), meta.cost_usd dice il costo reale di ogni chiamata, e la semantica rifiuta from_publication_date elencando nell'errore i filtri che accetta. Due punti in cui la documentazione non regge alla prova: dà per_page massimo 100 mentre l'API accetta 200 e lo dichiara nel proprio messaggio d'errore, e dice che la semantica accetta quasi tutti i filtri mentre la whitelist vera sta nell'errore. Corretta la ricetta 1, che passava title.search come parametro a sé (non valido) e senza api_key

• cup-cig: new reference references/openbdap-cli.md for openbdap-pp-cli, a CLI that covers source #2 only. It collapses the MOP branch — dossier <CUP> works out which regional partitions to query and returns project, payments, tenders with CIG, bidders, cost plan and owners in one call, where the reference documents five hand-built OData requests. mop and cerca return both ids per dataset (package for the CSV dump, XML resource for OData), which retires the two-ids trap, and colonne maps readable names to the mangled filter ids

• cup-cig: allinea is a prerequisite for cup, cig, dossier and opere too — they fetch live rows but read the local catalogue to know which dataset to query. Against an empty archive all four answer trovati: 0 and exit 0, told apart from a real miss only by a nota field, which is the one misreading this skill keeps warning about. Note for the record: --home does not relocate the archive, only --db does, so a fresh-home test silently keeps reading the populated one

• cup-cig: verified the one claim that justifies the tool — the API answers "__count" : "0" on a filter that matches rows, while conta and opere --solo-conteggio both answer 13489 for ANAS, and raw pagination counts 13489

• cup-cig: three traps survive the CLI and are documented as such. righe defaults to --limite 50 and returns exactly 50 rows with no warning where the true count is 347, so it must be paired with conta or --tutte --limite 0; scarica passes the dump through as latin-1 (à as the single byte 0xE0) while the OData path is clean UTF-8; cig without --regione costs 24s against 2s, measured with --no-cache — not serial requests, which already run six at a time, but their number: gar and pga for 21 regions, two calls each

• cup-cig: the decimals trap no longer holds. The reference claimed a comma in prg against a dot in gar and sal; today it is a dot in every family, on the same /datastore/dump/ endpoint the reference documents — 866.216 numeric fields with a dot and none with a comma over 66.633 rows of the national prg, and the six Sicilian regional files agree. Checked with plain curl too, so it is not the CLI normalising anything. A stale REPLACE(col, ',', '.') is a no-op, which is the bad kind of wrong: it hides the CAST it was protecting. What does vary, and was not documented, is the quoting: numeric fields are quoted in the national prg and bare in the regional files

• cup-cig: dossier returns six of the seven MOP families — no localizzazione — so it answers everything about a work except where it is. mop does map the seventh, --famiglia localizzazione works; it is the flag's help text that lists six values, and --regione hides the row because Localizzazione is national. How dossier picks the region is not what the output suggests: it matches a region name inside Descrizione Titolare, so REGIONE AUTONOMA VALLE D'AOSTA resolves and COMUNE DI PALERMO does not — regione comes back null and it falls back to querying every partition. It never reads the Localizzazione dataset, so a null region says nothing about where the work is

• cup-cig: compatibility was 520 characters against the 500 the spec allows, and had been for a while. The macOS coreutils hint moved into the body next to the other cig-fetch.sh requirements, where it is more useful anyway: the field is now 473. No other skill in the repo is over its frontmatter limits

• cup-cig: the portal's RICERCA PER CUP form returns an Excel with a Dettaglio Indicatori sheet that has no counterpart in the open data. Of the 29 datasets matching indicator in the catalogue, all are budget Note Integrative, none is MOP — so that sheet, and the 300-CUP cap that comes with it, is the one reason left to use the browser

• cup-cig: the CLI author shipped fixes for everything the report raised, same day, and the reference is realigned to them. Verified on a build of that day's HEAD: dossier has a localizzazione section with Codice ISTAT Comune already concatenated and takes the region from Codice Regione instead of guessing it from the owner's name — a CUP of the Comune di Palermo now resolves to Sicilia and the call drops from 34,7s to 9,4s, because the guess used to fail and it queried all 21 partitions; righe carries meta.nota with the real total when a result touches the limit (troncato a --limite 50 di 347 righe) and results stays an array; archivio_vuoto: true separates an unsynced archive from a genuine miss, and --home now moves the archive; scarica converts to UTF-8, byte-identical to iconv -f ISO-8859-1, with --raw for the portal's own bytes; cig reads the schema once per family and runs twelve at a time, 18,8s, and 4,1s with --rate-limit 0

• cup-cig: on the one suggestion turned down, the author is right and the reasoning is worth keeping: an exit code other than zero reads as a failed command, and an unsynced archive is not a failure — the command ran and answered. Hence the boolean field rather than an exit status

• cup-cig: the reference now states the date it was checked against, because a build older than that differs on four points and every one of them fails silently. The public library copy was still being published when this was written

2026-08-31

• cup-cig: tested the OpenCoesione API with the credentials now available in the environment. They are plain HTTP Basic and buy throughput, not access: wrong ones give 401, which is what tells a credentials problem apart from a quota one. Anonymous cannot sustain one request per second — 24 of 70 got through, first 429 on the twelfth; authenticated the same pace ran 70 out of 70. A burst cap sits well below the per-minute quota either way, so requests must be paced rather than batched
• cup-cig: the reference had the throttling failure mode backwards. It said a throttled reply is a JSON body with a detail key and not an HTTP error, hence the advice to branch on .detail. It is 429, anonymous and authenticated alike, so the rule is to branch on the HTTP status before parsing: .count // 0 on a throttled body reads as "project not found" and hands back a false negative on a project that is there. A genuine miss is 200 with count: 0. No script does OpenCoesione lookups, so nothing but prose needed fixing
• cup-cig: new trap, the mirror image of one already documented — filter values are slugs and an unknown one answers 500, not an empty result (territorio=sicilia fails, territorio=sicilia-regione returns 118.846 projects). A wrong filter name fails silently and looks like data; a wrong filter value fails loudly. Slugs come from the temi, nature, territori and programmi endpoints
• cup-cig: confirmed unchanged — limit and per_page are ignored (limit=3 downloads 26 MB where page_size=3 downloads 278 KB), search= and q= are accepted and ignored and return the whole catalogue's count
• cup-cig: the router line in SKILL.md said "REST API without authentication". True, but it is the only line loaded by default, and as written it gave an agent no reason to open the reference looking for credentials it does not know exist. Reworded, and the optional env vars added to compatibility. Realigned PLAN.md, which still carried the 12 req/min / 60 req/min figures dropped from the references in v0.13

2026-08-17

• typst-cards: new references/charts.md — cards can now embed real data charts, drawn with gribouille (grammar of graphics for Typst, same author as the full-bleed pattern already credited). Verified on gribouille 0.6.0 + typst 0.14.2
• typst-cards: the sub reference pins only what is card-specific and delegates the API surface to the upstream llms.txt / .llms.md, regenerated on every release — a copy of the ~50-geom list here would rot within a release
• typst-cards: plot(width:) rejects ratios, so the usable width is arithmetic the writer must do — 6.4in margined, 6.05in full-bleed on a 1:1 card. The plot is one rigid unbreakable block, so it overflows onto an extra page in the margined pattern and clips silently in the full-bleed one
• typst-cards: brand mapping is theme-minimal(ink:, paper:), with paper: rgb(0,0,0,0) when the chart sits on a colour band; fonts are inherited from the deck's #set text. Default tick labels read small at 1080px — raise them with text: element-text(size: 13pt)
• typst-cards: new references/design-md.md — the skill now reads a DESIGN.md as the source of the theme instead of guessing a palette. Two version numbers not to confuse: the format is at version: alpha, the @google/design.md tooling has its own release line
• typst-cards: the two conversions that bite when a DESIGN.md meets a card. fontSize px→pt is ×0.75, but a literal translation makes a screen-sized scale look timid on 1080px — keep the ratios between levels, raise the absolute values. And CSS lineHeight is the whole line box while Typst leading is the gap added on top, so (L−1) collides: subtract ~0.75 (verified, 87px pitch against the 88px prescribed by lineHeight: 1.1)
• typst-cards: {token.references} must be resolved before reaching theme.typ, and named CSS colours must be converted to hex — rgb() rejects both loudly. oklch() is native in typst 0.14.2. Fonts are the quiet failure: a DESIGN.md names brand families that are typically absent, and Typst falls back with a warning, not an error
• typst-cards: components and elevation largely do not map onto a static card (no hover state, shadows read poorly at 1080px) — stated as a boundary so the mapping stays honest instead of inventing one for text inputs
• typst-cards: three silent failure modes found by compiling, not by reading docs — geom-text draws nothing when label maps to a number instead of a string, horizontal bars need coord-flip() rather than swapped aesthetics (swapping renders axes with no bars), and category order is alphabetical unless fixed with scale-discrete(limits:)
• cup-cig v0.14: fixed the dangling pointer in bdap-mop.md — two references to references/bdap-mop-datasets.md, a file that does not exist because the content was inlined into the same file. Found independently by three reviewers, which is what made it credible. Also removed the merge scar it left behind: two consecutive identical ## headings and a paragraph declaring itself "companion to the BDAP section" while sitting inside that very section. The second half of the file now opens with an explicit Reference heading saying it is material to consult, not to read
• cup-cig: the HEAD → 404 quirk on ItaliaDomani is real and was nearly deleted by mistake. A first check — one request per file — returned 200 on all four, which looked like proof the documented behaviour was gone. Repeating it five times per file showed the truth: the response is not deterministic. PNRR_Progetti and PNRR_Localizzazione answer 404 most of the time with the occasional 200; PNRR_Gare and PNRR_Soggetti answer 200 every time. The CDN edge nodes are not configured alike, so which node serves the request decides the answer
• cup-cig: the instruction «never probe these URLs with curl -I» therefore stays, with the non-determinism now documented — a 404 on HEAD proves neither that the file is missing nor that it exists. Added the methodological corollary: a single probe per file is not a measurement here. Confirmed separately that Range is genuinely ignored (a ranged request returns 200, not 206)
• cup-cig: reworded the Akamai block as conditional on the caller's network rather than as a property of the server — from consumer networks a plain GET works with no headers at all, from datacenter egress it needs User-Agent and Range together, and from some egress (GitHub Actions on Azure) nothing works. Stated the consequence: never conclude anything about this source from a single network
• cup-cig: ported v0.11 into the repo. The skill had been developed for months in the installed copy at ~/.agents/skills/cup-cig/, which is not a git clone, so v0.1 → v0.11 existed only there, unversioned and unbacked
• cup-cig: decision tree rewritten — OpenCUP is now the mandatory entry point (resolves 3.457 of 3.481 CUPs measured), and its DESC_TIPO_COPERTURA / DESC_STRUMENTO fields are read as a routing hint to pick the monitoring system (PNRR → ReGiS, cohesion → OpenCoesione, national → BDAP MOP). The CIG branch now splits on what you need, not on the shape of the code
• cup-cig: the Z prefix no longer identifies below-threshold CIGs — it was the convention until 31/12/2023 only; measured at 96.5% of smartcig records in December 2023 and 0% from January 2024 onwards
• cup-cig: eight new reference files — opencup.md, bdap-mop.md, anac-datasets.md, anac-lifecycle.md, anac-ocds.md, anac-pvl.md, opencoesione.md, regis-italiadomani.md; scp-mit-api.md replaced by scp-mit.md (SCP-MIT is empty from 2024 on, superseded by ANAC PVL via the anac-pl CLI)
• cup-cig: dropped the old scripts/cig-fetch.sh — the browser-scraping route to ANAC dettaglio-gara is superseded by the ANAC bulk datasets and CLIs. Recoverable from history at bf37822 if ever needed
• cup-cig v0.12: new scripts/cig-fetch.sh, a rewrite of the dropped one — batch input from file or arguments, deduplication, one retry, idempotent skip of valid existing files, --force, --stdout. The download is accepted only if .bando.CIG matches the request, so a wrong file is never left on disk. Translated from Italian on the way in; Risultati della Ricerca stays as is, it is ANAC page text
• cup-cig: cig-fetch.sh returns 20 top-level sections in one call — what would otherwise take four or five bulk datasets — and bando.CUP makes it a CIG→CUP resolver. Measured: ~10s per CIG, and 3 CIGs also ~10s total since the browser session is reused; a non-existent CIG costs ~56s (two attempts) and reports the misleading starting the search; SmartCIG yield ~6 populated sections against ~10 for ordinary CIGs
• cup-cig: new scripts/test-cig-fetch.sh — offline contract test driving the script with a fake agent-browser, no network, ~1s. Run it after touching the fetcher
• cup-cig v0.13: stripped the one-off measurements out of SKILL.md and all nine reference files. A skill carries the concepts needed to use a source well, not a technical datasheet of it — the test for a number is does it change a decision?. Kept: operational limits (the 10.000 totcount cap, 12 req/min, the bare 50 rows that signal a silent truncation, exit codes, timeouts), the 31/12/2023 regime cutoff, break-even thresholds, and every ISTAT/CUP/CIG code used as a worked example. Dropped: file sizes in MB, row counts, distinct-value counts, percentages, and the provenance of each sample
• cup-cig: the worst offenders were figures hard-coded from a single ad-hoc measurement — «99,3% of CUPs resolve (3.457 of 3.481)», «49 of 100 CUPs appear in MOP», «250 of 257 pairs coincide». They freeze one sample as if it were a property of the source, and they age silently. Each was rewritten as the claim it was evidence for, keeping the operational consequence (a miss in MOP is expected, not an anomaly; the two CIG↔CUP bridges disagree, use the union)
• cup-cig: three measurements were load-bearing and were kept as claims rather than cut — the Z prefix appearing on no record from January 2024 (what makes "never infer the regime from the code shape" absolute), multi-municipality CUPs spanning dozens of territories, and the API gateway serialising per credential so parallel calls buy nothing
• cup-cig: added two operational facts the mirror's README had and the skill lacked — the cost of a query is the number of round-trips rather than the bytes, and wildcards are refused over plain HTTP (_manifest.json is the way to enumerate the files). Fixed a broken reference to references/scp-mit-api.md, deleted in the v0.11 port, now pointing at the WADL
• cup-cig: three CUP↔CIG sources exist (ANAC cup, BDAP MOP Gare, ReGiS PNRR_Gare) and they do not fully agree — use the union when completeness matters

2026-08-10

• typst-cards: new full-bleed layout pattern — margin: 0 + a slide() helper that adds the padding back via inset, with rail, page counter and URL drawn through place (out of flow, so content cannot push them). Enables backgrounds, colour bands and images reaching every edge, which the margined pattern cannot do. New references/slides-fullbleed-starter.typ (5 cards, 1:1); the two existing starters are unchanged. New "Two layout patterns" section with the rule for choosing. Adapted from Mickaël Canouil's post on Typst LinkedIn carousels (credited in SKILL.md); geometry kept in inches — 7.5in × 144ppi = exactly 1080px, while his 21cm square gives 1191px
• typst-cards: PDF output alongside PNG — a LinkedIn carousel is uploaded as a single PDF document post, so typst compile slides.typ output/deck.pdf is now part of the flow; frontmatter description updated accordingly; optional ImageMagick one-liner for an animated GIF preview
• typst-cards: new pitfall — block(width: 100%, height: 100%, inset: ...) clips silently on overflow (one page, text cut at the edge, no warning, exit 0); the same content without height: 100% spills into 6 real pages. Diagnostic recipe + page-count check added to Phase 4. Verified on typst 0.14.2
• typst-cards: new pitfall — inside body, width: 100% is the padded column, not the page, so "full-bleed" decoration drawn there comes out framed; the slide() helper takes a bleed argument for content that must be measured against the page
• typst-cards: font availability documented as two tiers — only Libertinus Serif, New Computer Modern (+ Math) and DejaVu Sans Mono are embedded in the binary (typst fonts --ignore-system-fonts); DejaVu Sans, previously the skill's primary body font, is system-only and can silently fall back elsewhere. Stated plainly: there is no embedded sans-serif
• typst-cards: neutral placeholder palette in theme-starter.typ — the old default was recognisably GitHub (#0d1117 / #58a6ff) and looked like a decision already made, so it got shipped unchanged. Now neutral greys for surfaces and text plus a single accent marked as the first value to replace, with ACC/ACC-L documented as the same hue at two lightnesses (not two brand colours). Every pairing measured: 16.5 / 15.9 / 6.6 / 8.4 / 5.8 / 7.0 : 1, all above 4.5:1, and the numbers are in the file. All three starters recompiled and reviewed
• typst-cards: fix visible bug in slides-16x9-starter.typ — [short\ndescription] rendered as shortndescription, since a backslash escapes the next character and a line break is a trailing backslash at end of line. Compiled cleanly, output was wrong. New pitfall documented
• typst-cards: new pitfall sub-case — smart quotes never reach string values, so #helper("l'x") keeps a straight apostrophe while #helper[l'x] and plain markup get ’; interpolating the string into content does not rescue it. Silent, no error, and a deck mixing prose with helper calls ends up with two kinds of apostrophe. Surfaced by the first real deck built with the new pattern; mechanism isolated with a 5-case compile
• typst-cards: fix broken cp recipe — all three slide starters imported theme-starter.typ while Phase 3 tells you to copy the theme to theme.typ, so following the documented recipe failed with "file not found". Imports now point at theme.typ; verified end-to-end on all three starters (5 / 3 / 5 pages + PDF). Reported by the Copilot review on PR #27
• typst-cards: #set document(title:, author:) for PDF metadata, #set text(lang: "it") for Italian hyphenation, and deck-level guidance (5–7 cards, cover → content → CTA, one inverted card, one visual idea per deck)

2026-07-28

• scrivi-chiaro-pa: new skill — write, rewrite, or review Italian PA content in plain language, based on the Designers Italia Guida al linguaggio della Pubblica Amministrazione (CC-BY 4.0). Three modes (draft / rewrite / review), tone-of-voice scenarios, final glossary pass + checklist. References: style/grammar/numbers, structure/usability/SEO/accessibility, tone of voice (9 scenarios), social/newsletter/images/editorial management, A–Z glossary (~90 terms in two tables). Content harvested from the published "bozza" on docs.italia.it (suggerimenti + tono di voce sections exist only there) and from the italia/writing-toolkit repo (glossary RST sources)

2026-06-14

• datawrapper: document arrow markers in locator maps (new Datawrapper feature). Tested end-to-end via pure API: arrows are writable (line + flow types, triangle/lines heads, bidirectional, gradient, taper, curve). Full JSON schema + per-field table added to references/locator-map.md
• datawrapper: fix wrong marker storage in references/locator-map.md — there is NO /markers endpoint (404); points, areas and arrows are all stored in the chart data via PUT /v3/charts/<ID>/data as {"markers":[...]} (full replace each PUT). Corrected point (needs icon block) and area (colors in top-level properties, GeoJSON-style keys) examples
• datawrapper: document counterintuitive curve.angle behaviour — values near 0 produce huge looping arcs; gentle curves need larger magnitudes that scale with arrow length (|angle| ≈ km/2.5); recommend tuning in GUI then reading back via API (verified by hand-edit round-trip)

2026-04-26

• typst-cards: refactor starter templates to native page-level footer — #set page(footer: footer-block(URL, dark: ...)) with Typst's counter(page).display() / .final().first(); replaces fragile #v(1fr) + #ctr(n,total) last-row pattern that produced silent ghost slides on dense content; new footer-block helper in theme; new pitfall entry; sub-case on literal helper params under special-chars table; smoke-tested both starters (5+3 PNGs, footer fully visible)
• typst-cards: add "Footer/edge content never clips" design rule — footer, counter, source URL must be fully visible; fix via larger page bottom margin or smaller footer font; must-check in Phase 4
• typst-cards: expand "Special characters in content mode" pitfall — table covering $, //, _, *, @, <, ~ (replaces the previous single bullet on <); typical contexts where these characters break compilation with "unclosed delimiter" errors
• typst-cards: tighten Phase 4 — explicit read → evaluate → report+propose → wait flow; AI must look at PNGs (not just check existence), evaluate against design principles, propose concrete fixes, wait for user decision before iterating
• typst-cards: new skill — generates PNG images for online communication (social posts, carousels, infographics) using Typst; brand-materials interview, supported formats (1:1, 4:5, 9:16, 16:9, 1.91:1), three reference starters (theme-starter.typ, slides-1x1-starter.typ, slides-16x9-starter.typ), known pitfalls

2026-04-24

• openalex: strengthen API key security — added prominent SECURITY callout at top of SKILL.md; expanded Common Pitfalls rule to cover all output contexts (text responses, echoed commands, logs), not just verification

2026-03-03

• evals/openalex: add run_evals.sh — runs prompts via claude -p, saves outputs to runs/YYYY-MM-DD/, auto-checks 4 pitfall patterns (relevance_sort, title.search scope, multi-line curl, API key printed)
• openalex: never print API key to verify presence — use [[ -n "${OPENALEX_API_KEY:-}" ]]; added to Quick Start and Common Pitfalls in SKILL.md
• evals/openalex: add 20 synthetic test prompts (test-09–test-28) targeting 11 known pitfall zones (relevance_sort_no_search, title_search_standalone, entity_name_filter, sequential_batch, nested_select, missing_api_key, default_per_page, multi_line_curl, null_csv_field, pdf_fallback_chain, wrong_search_scope); add corresponding pitfall checks to checks.md

2026-03-01

• open-data-quality: ZIP-wrapped CSV support — if resource declared as CSV is actually a ZIP, extract largest CSV to /tmp and proceed with full analysis; report MINOR zip_wrapped_csv instead of BLOCKER; uses Python zipfile (no DuckDB/Polars ZIP support); score 40→86 on Liguria air quality dataset

• open-data-quality: fix _extras_value — also check top-level package fields; some harvesters (dati.gov.it from regional portals) promote holder_name/identifier to top-level instead of extras; fixes false-positive MAJORs; score 83→97 on test dataset

• open-data-quality: add qualitative assessment section to SKILL.md — 9 LLM-only checks (title discoverability, title↔description, description↔content, content↔update frequency, dataset usefulness); runs after scripts, requires data content; Good/Acceptable/Poor rating; added to report template

• open-data-quality: add outlier_values check (phase3_content) — IQR method on numeric columns (≥100 rows); severity MINOR, -2 pts; no fix suggestion (signal only); new fixture outlier_values.csv; 36/36 tests pass

• open-data-quality: add duplicate_rows check (phase3_content) — detects exact duplicate rows via DuckDB SELECT DISTINCT *; severity MAJOR, -3 pts on data content quality; new fixture duplicate_rows.csv; 35/35 tests pass

2026-02-26

• open-data-quality: fix IT holder label — dcatapit:datasetHolder → dct:rightsHolder in code and all reference docs (confirmed from real dati.gov.it data)

• open-data-quality: remove deprecated dcatapit:datasetHolder row from profiles table; dct:rightsHolder now marked M for IT

• open-data-quality: add portal_field_aliases.json — JSON vocabulary mapping standard DCAT-AP field names to portal-specific CKAN extras keys; UK profile maps issued/modified → dcat_issued/dcat_modified

• open-data-quality: fix UK date detection — metadata_validator.py now uses FIELD_ALIASES fallback for date fields per profile

• open-data-quality: add 3 new tests — test_it_holder_present_ok, test_it_holder_missing_flagged, test_uk_dcat_prefixed_dates_accepted; 34/34 pass

• open-data-quality: added pytest test suite — 25 tests across phase0–3 + CLI integration; fixtures in scripts/tests/fixtures/; pytest added as dev dependency in pyproject.toml

• open-data-quality: fix fuzzy check — skip datetime/timestamp columns stored as VARCHAR (e.g. 2025-03-14T00:00:00 was triggering false positive near-duplicate alert)

• open-data-quality: file type detection in phase 0 — detect ZIP, HTML/XML, JSON, PDF, OLE2/Excel, UTF-16 via magic bytes/content sniffing; report specific type (e.g. "File is a ZIP archive") instead of generic binary/separator error

• open-data-quality: fuzzy false positive fix — add levenshtein/max_len < 10% ratio filter + raise JW threshold to 0.95 + minimum length > 5; eliminates NORD-ESTNORD-OVEST, MINISTERO DELLA DIFESASALUTE type false positives while preserving real typos (D'INTERESSE~DI INTERESSE caught at 4% ratio)

• open-data-quality: fix encoding false positive — normalize utf_8 → utf-8 before comparison; was marking valid UTF-8 files as MAJOR issue

• open-data-quality: split fuzzy check — trailing/leading whitespace now reported separately; fuzzy comparison works on trimmed values to avoid spurious matches

• open-data-quality: fix #11 — placeholder message now shows actual values found (e.g. NA) instead of full catalog n/a, n.d., -…; SQL uses list_distinct(list(...)) to collect found values per column

• open-data-quality: non-UTF8 encoding no longer a BLOCKER — charset_normalizer detects encoding, file converted to UTF-8 temp copy, full analysis runs; MAJOR finding reported (tested on Comune di Palermo CP1250 dataset: 33→73/100)

• open-data-quality: added fuzzy near-duplicate category check via jaro_winkler_similarity > 0.92 (DuckDB built-in) — found real issues in Palermo dataset

• open-data-quality: CRLF line endings no longer flagged (RFC 4180 prescribes CRLF)

• open-data-quality: replaced chardet with charset_normalizer for more accurate encoding detection

• open-data-quality: added developer notes to CONTRIBUTING.md (WSL/uvx cache, PYTHONUTF8, DuckDB lenient parsing)

• evals/open-data-quality/fixtures/palermo-edifici-pubblici-cp1250.csv: archived as encoding test fixture

• open-data-quality: fixed false BLOCKER on valid CSVs with quoted newlines in headers — retry with strict_mode=false; added _lenient flag + _rcsv() helper across all DuckDB queries (score 29→86 on Copertino dataset)

• open-data-quality: SKILL.md rewritten — uvx odq-csv/odq-ckan is now the single primary path; all bash inline phases removed (no double maintenance)

• SKILL.md reduced from ~450 to ~130 lines

• evals/checks.md updated to validate uvx + package usage

2026-02-25

• Added open-data-quality skill: CSV validator (odq-csv) and CKAN/DCAT-AP metadata validator (odq-ckan)
• Created missing __init__.py for package discovery
• Translated scripts/README.md from Italian to English
• Tested end-to-end on real dati.gov.it datasets (bilancio, popolazione)
• Created eval suite: evals/open-data-quality/ (8 prompts, 15 checks)

2026-02-24

• openalex SKILL.md v0.2: added filter syntax (OR/NOT/ranges), batch pipe lookup, two-step entity lookup, group_by/sample/seed params, per-page=200 default, error handling with backoff, endpoint costs table

2026-02-15

• Saved reference documents in docs/: testing-agent-skills-with-evals.md and agent-skills-specification.md
• Added "Core Idea" section to PRD.md (skill creation as an inclusive, non-technical activity)
• Created docs/prd.md with Alessio Cimarelli's eval proposal
• Created evals/ structure with _template/ and first eval for openalex
• First complete eval run on openalex: score 78/100 (7/9 checks), 7 improvements to the skill
  • single-line curl required (warning at top of SKILL.md)
  • title.search inside filter=, not standalone
  • display_name as title (Output Format section)
  • api_key explicitly documented as pitfall
  • Europe PMC fallback documented as recipe #6
  • Definition of Done added to SKILL.md
• Updated README.md: catalog with Eval column, Evals section, fix skill data-quality-csv → openalex
• Created project CLAUDE.md