Update to LibreJustice

{
  "name": "recherche-jurisprudence",
  "description": "Search and cite French case law and statutes with the LibreJustice MCP tools. Use whenever the user asks about French or European court decisions (Cour de cassation, Conseil d'État, cours d'appel, TJ/TA, CNDA, Conseil constitutionnel, CNIL, CEDH, CJUE), needs authorities for a brief, requête, référé or consultation, wants to check how courts rule on an issue, needs counts of decisions on a point, or asks what a code article says at a given date, even if they never mention LibreJustice.",
  "included_files": [],
  "skill_md_contents": "---\nname: recherche-jurisprudence\ndescription: Search and cite French case law and statutes with the LibreJustice MCP tools. Use whenever the user asks about French or European court decisions (Cour de cassation, Conseil d'État, cours d'appel, TJ/TA, CNDA, Conseil constitutionnel, CNIL, CEDH, CJUE), needs authorities for a brief, requête, référé or consultation, wants to check how courts rule on an issue, needs counts of decisions on a point, or asks what a code article says at a given date, even if they never mention LibreJustice.\n---\n\n# French and European case-law research (LibreJustice)\n\nThe user is typically a litigator: they need decisions whose reasoning\nstates a precise proposition, applicable to their configuration, with\nan excerpt quotable in a brief.\n\nFive rules dominate everything:\n\n1. **No decision appears in the answer — as support, contrary\n   authority, analogy, or lead — unless you read its full text with\n   `get_decision`.** An `aiSummary` or a `snippet` is not reading.\n2. **Quotation marks quote a text you fetched, nothing else.** A\n   hit's `aiSummary` is a machine paraphrase — never the court's\n   words, however fluent it sounds; a `snippet` is verbatim but torn\n   from context. Every quoted string in the answer is copied\n   character-for-character from a `get_decision` or\n   `get_legal_text` text of this session — not opened means not\n   quotable. A quotation is **one continuous span of one single\n   text**, cuts marked « […] » : merging passages — or decisions —\n   into one pair of quotation marks is fabrication, one source per\n   quote, two passages are two quotes. A sentence reporting a\n   decision in your own voice (« la cour retient que… » followed by\n   your summary or a parenthesized list) is prose, never a\n   quotation.\n3. **Every decision named anywhere is a markdown link** to its `url`\n   verbatim — `[CA Paris, 10 janv. 2024, n° 21/22203](https://librejustice.fr/decision/…)`.\n   Never reconstruct or shorten a URL.\n4. **The exact object of the question is a hard filter.** The most\n   quotable sentence in the corpus routinely concerns a near-twin\n   object one word apart — another mention, another délai, another\n   clause. A decision whose decisive sentence names a different\n   object is an analogy, never an authority, however perfectly its\n   words match. When the decisive sentence leaves the object\n   unqualified (« l'heure », « ce délai », « cette mention »), its\n   object is the one in the moyen it answers — find the party\n   critique above it; the sentence inherits *that* object.\n5. **The first and last lines of every `get_decision` text are the\n   decision's fate on appeal.** A `[SORT DE CETTE DÉCISION SUR\n   RECOURS : …]` banner (also served as the `appellateFate` field)\n   states what became of it on review. Copy it into the answer. It\n   overrides your own docket sweep: a sweep that found nothing while\n   the text carries a banner means the sweep missed the arrêt, never\n   that none exists.\n\nA `get_decision` response may also carry `commentaires` — the court's\nown analysis served inline (`body`) and outbound links (`url`) to the\nrapporteur public's conclusions or related court documents. They are\ncontext and cite as commentary, never as the ruling: only the decision\ntext quotes as the court's words.\n\n## Protocol — run the steps in order\n\n### 1. Frame\n\nPin down (ask when the request leaves them open): the proposition in\none sentence; who demands what and what winning means; the legally\nrelevant date; the target courts; the exclusions; the proximity axes\nthat decide transposability. **A legal assertion inside the question\n(« c'est bien uniquement devant telle juridiction ? », « cet acte est\nnul, non ? ») is a claim to research, never an input**: the\ncomplement of the assertion gets its own step-2 sweeps — its own\nqueries and filters, including the courts or outcomes the assertion\nexcludes — and the answer's opening sentence states your verified\nverdict, never an echo of the premise. An answer built on the user's\nfalse premise fails whole. Follow-ups refine, they never reset:\n« d'autres ? » = new decisions, same bar; « tu peux vérifier ? » =\nreopen the source, never restate with more confidence.\n\n### 2. Sweep — three passes, all mandatory\n\n1. **One `solution`-filtered sweep per side — your first two\n   `search_decisions` calls.** Map each direction of the holding to\n   outcomes, then run the same query twice with opposite `solution`\n   filters, e.g. `{\"query\": \"convention de forfait en jours privée\n   d'effet\", \"jurisdiction_code\": [\"ca_versailles\"], \"solution\":\n   [\"SATISFACTION_TOTALE\", \"SATISFACTION_PARTIELLE\"]}` then the same\n   with `[\"REJET\"]`. This is the only lever that separates the two\n   directions of a line.\n2. **The consecrated formula, quoted, in each phrasing.**\n3. **A plain full-sentence query** (no quotes).\n\nEngine facts:\n\n- **Ranking is direction-blind** — semantic matching ignores\n  negations, and summaries surface the side that won. Never read\n  direction or absence in an order, an `aiSummary` or a `snippet`.\n- **A lexical list (quotes/operators) is a set, not a ranking**: read\n  it whole or partition it, never skim its top.\n- **The window is `limit` (max 20), no pagination.** Read the\n  `date_lecture_year` facet on every sweep: a year newer than the\n  newest hit you opened, with a nonzero count, is unswept — re-run\n  with `date_from`/`date_to` on that year before concluding.\n- **`date_from` is the filter that silently hides a line's founding\n  arrêt** — often decades old. Bound dates only when the question\n  itself is time-bounded or a facet year needs re-sweeping, never by\n  default.\n- Filter tokens (« ca_paris », « REJET ») poison the query text —\n  filters carry constraints, the query carries only words a court\n  would write. After two queries with no new candidate, change one\n  real axis or start opening.\n\n### 3. Read everything you might cite\n\nOpen every hit of the target court that touches the issue. Open\nfirst, and always, every hit whose served `solution` sits on the\nuser's side: a directional question answered without opening a\nsingle same-side hit is the run's defining failure. Fewer than\neight decisions opened by the end of this step means the research\nhas not happened: go back to step 2 and change a real axis (court\nlevel, phrasing, dates) until eight full texts are read or the\ntarget courts' relevant hits are exhausted — the corpus almost\nalways holds more than one screen of them. Five verdicts per\ndecision:\n\n- **Who speaks.** Court's motifs, or a party's argument (« il\n  soutient que… ») ? Cross-check the dispositif and who succombe —\n  the loser's phrase is contrary authority under a favorable label.\n  In a cassation arrêt the moyen (« alors que… ») is the loser's\n  text even when it reads like a holding; the Cour speaks in « Mais\n  attendu » / « Réponse de la Cour », and « par les motifs reproduits\n  au moyen » attributes those motifs to the court below. Quote each\n  voice separately — one clause of the moyen imported into the\n  Cour's motif turns the quote into fabrication.\n- **Exact object.** Near-twins one word apart have opposite regimes\n  (forclusion/prescription, faute grave/lourde, nullité\n  relative/absolue) — and snippets truncate, summaries smooth over,\n  exactly the deciding word.\n- **Appellate fate** (dominant rule 5). `INFIRMATION` kills the\n  decision as support. `CONFIRMATION` = open that arrêt and cite the\n  pair; the arrêt holds what the judgment held, never the opposite\n  line's direction. No banner served = run the judgment's docket\n  number, quoted, filtered to the appellate court. **A sweep hit is\n  the same case only if its « Décision déférée » header names your\n  judgment — same court, same date** : RG numbers collide across\n  courts, and a hit from another ressort or another date is noise,\n  never the fate. No banner and no sweep = « aucun recours lié dans\n  le corpus » — that sentence describes the served data and is true\n  by construction; « balayé » or « vérifié » may only be written\n  when the docket-number query is in this session. Never\n  « définitif ».\n- **Legal basis.** Decision texts carry inline `/texte/` links; open\n  them with `get_legal_text` at the facts' date. A failed tool call\n  (bad URL format, unknown code) is retried with the corrected\n  argument, never silently dropped: an article the question turns on\n  that you never managed to read voids every claim about it.\n- **The quotable sentence.** While the text is in front of you, copy\n  the exact sentence(s) of the motifs you would quote in a brief.\n  The writing stage may only put between quotation marks strings\n  collected this way — a quote reconstructed at writing time from\n  memory or from an `aiSummary` is where fabrication happens.\n\n### 4. Map the line\n\nRe-run the winning query with `sort: \"date_desc\"` — **once per\ndirection, with that side's `solution` filter, per target court** —\nand read the most recent decisions of each. When a court ruled both\nways, the deliverable is a dated timeline, never a flat « court X\nsays P »: either side of a flip, presented alone, misleads. A hit\nnewer than every decision the answer names, seen in *any* list this\nsession, is either opened and cited or excluded for a stated\nreason — never silently dropped.\n\n### 5. Absence claims and counts\n\n« No decision of court X states P » is falsifiable with one\ncitation. Before writing it: exact formula (both phrasings) + a\ndescriptive query under the court filter; a `solution`-filtered\nsweep on P's outcomes; `date_from` widened; facets read. Last:\nre-scan the hits you did NOT open across every list of the session —\none unopened hit whose `solution` sits on P's side voids the claim\nuntil opened. Counts: define the corpus; keep raw hits, deduplicated\ndecisions and verified holdings apart. A counting table is built\nfrom the facets of **one named query per column** — never merge\nfacets from different queries into one series. Facets count the same\ncandidate set as `total` — the engine's best few hundred matches —\nnot the whole corpus: for a corpus-wide count, narrow with filters\n(court, dates) until `total` itself is the count. A yearly series\nthat starts or jumps abruptly usually marks the edge of source\ncoverage, not the birth of the contentieux: say so instead of\nnarrating the jump.\n\n## The answer — always these four blocks, in this order\n\nExact French headings: « État du droit », « Autorités pour la\nposition recherchée », « Autorités contraires et risques »,\n« Périmètre de la recherche ».\n\n1. **« État du droit »**: the direct answer, then per target court\n   the most recent decision in *each* direction, named with its\n   date. Naming a court's latest word requires having run the\n   `date_desc` + `solution` query of step 4 for that court and\n   direction.\n2. **« Autorités pour la position recherchée »** — a table when\n   three or more, with exactly these columns: Décision | Extrait des\n   motifs | Proximité (objet exact) | Sort en appel | Limite.\n   - The citation cell is itself the markdown link — never a\n     separate « lien » column.\n   - The **excerpt** cell carries either a quotation collected at\n     read time (the quotable sentence of step 3) or — when none was\n     collected for that decision — a plain-prose description with NO\n     quotation marks, marked « (résumé — citation non collectée) ».\n     An honest unquoted summary always beats quotation marks around\n     anything not copied verbatim; a passage from the exposé of a\n     party's moyens is that party's argument, not the court's.\n   - The **proximity** cell quotes the object words of the decisive\n     sentence; when they name a different object than the user's,\n     the row moves to the analogies section, whatever its direction.\n   - The **fate** cell copies the served banner — « infirmé\n     par [linked arrêt] » / « confirmé par [linked arrêt] » — or, when\n     no banner was served, says « aucun recours lié dans le corpus » :\n     that sentence describes the served data and needs no further\n     proof. The words « balayé » / « vérifié » may appear only when\n     the docket-number query is in this session's searches — claiming\n     a check that never ran is the worst lie this answer can carry.\n     Any other wording (« à vérifier », « définitif »…) is a failed\n     check. An arrêt cited for its confirmation names and\n     links the judgment it confirms in the same row — the pair cites\n     together.\n   - A thin side stays thin: when a court offers a single decision,\n     present it alone and say so — padding the table with non-ruling\n     or off-direction rows is worse than a one-row table.\n3. **« Autorités contraires et risques »**: what opposing counsel\n   will plead — the recent contrary line, infirmations, reversals,\n   dated, and **per target court**: a first-instance court has its\n   own contrary decisions, and calling its line « constante » while\n   the opposite-`solution` sweep never ran for it is the check that\n   fails most.\n4. **« Périmètre de la recherche »**: a table with one row per\n   `search_decisions` call of this session — Requête | Filtres\n   (`solution`, dates, cours) | Tri | Hits ouverts. Build it the way\n   excerpts are built: append the row when the query is sent, then\n   paste the table into the answer — a table reconstructed from\n   memory at drafting time is where fabricated rows come from.\n   The block carries a **second table, « Décisions ouvertes »** : one\n   row per `get_decision` of this session, in call order — the linked\n   citation. Built the same way, one row appended per returned text.\n   This table is the answer's whitelist: every decision link anywhere\n   else in the answer is a copy of one of its rows, and a row for a\n   call that never returned a text is the same lie as a fabricated\n   quote. Every\n   cell copies a parameter actually sent or a uid actually opened; a\n   row whose query never ran is the same lie as a fabricated\n   citation, and a search that ran but is missing (including failed\n   ones) is a hole in the audit trail. A direction of the holding\n   with no `solution`-filtered row is unresearched: run it now, or\n   write « non recherché » under the table — that plain sentence is\n   always available and always true. Analogies in their own clearly\n   separated section.\n\nCite only decisions returned by the tools.\n\n## Final check — verify each item against the draft, fix before sending\n\n- Four blocks present, in order, exact headings; table carries its\n  five columns.\n- No row whose Proximité cell names a different object stays in the\n  authorities table — it moves to the analogies section.\n- Every Sort en appel cell is verbatim one of: the copied banner\n  (« infirmé par… » / « confirmé par… », pair linked in the row) or\n  « aucun recours lié dans le corpus ». Re-read the first line of\n  each cited decision's text now — a banner you saw and did not copy\n  is the worst chronology error. The words « balayé » / « vérifié »\n  survive only if you can point to the search of this session whose\n  query was that docket number; otherwise rewrite the cell to the\n  default formula.\n- Point each Périmètre row to the `search_decisions` call of this\n  session it copies — delete any row you cannot point to, add any\n  search you ran but did not list. Filters that appear in a row but\n  were never sent (a `solution`, a court, a date) are fabrication.\n- Cross the draft against the « Décisions ouvertes » table: every\n  decision link elsewhere in the answer has its row there; a link\n  without a row is opened now (row added) or deleted with every\n  claim resting on it. Never state a number of decisions read other\n  than that table's row count.\n- Search every quoted string of the draft, **as one continuous\n  block**, in the `get_decision` / `get_legal_text` texts of this\n  session. A string found only in separate pieces is a splice —\n  re-cut it into one quote per source, « […] » for internal cuts. A\n  string you can only find in a hit's `aiSummary` or `snippet`, or\n  nowhere, is rewritten from the fetched text or unquoted. Then\n  check who speaks: the motifs of the decision it is attributed\n  to — not a party's argument or moyen, not another decision.\n- Weight adjectives (« isolé », « constant ») consistent with the\n  decisions the answer itself lists; the answer states which way the\n  most recent decisions of each target court go; an infirmed\n  judgment never stands as support.\n- Every « no decision » claim earned under step 5.\n\nA failed item is fixed before sending, not flagged.\n\n## Statutes\n\nQuick lookups only — anything deeper (treaties, EU law, versions\nacross recodifications) is the recherche-normes skill's job.\n`get_legal_text` returns an article as it read on a given date:\npass `date` and say which version you quote. `list_my_activity`\n(signed-in account) lists recent searches, bookmarks and reading\nhistory — useful to resume ongoing work.\n"
}