champollion-mcp-server 0.3.1 → 0.4.0

package/src/tools/forge.js

@@ -335,6 +335,7 @@ export function forgeFlagArg(flag) {
  */
 export const FORGE_FLAGS_NEVER_TOOLS = {
   '--show-text': 'prints a withheld (local-only / sealed) corpus\'s sentences — for a person at the terminal, never an agent',
+  '--holders-agreed-by': 'the data\'s holders\' own answer (a local-only model served off this machine) — an agent never gives it',
   '--json': 'every tool adds it',
   '--workspace': 'every tool takes `workspace`',
   '--help': 'not a step',
@@ -922,10 +923,11 @@ export function statusNextHint(r) {
   const terminal = tools.filter((t) => t.startsWith('terminal:')).map((t) => t.replace(/^terminal: nmt-forge /, ''));
   if (a.state === 'serving') {
     return 'the chosen model is served and answering (result.advice.ready names its URL) — forge has nothing '
-      + 'left to run. What is left is the app\'s: the api method config from DEPLOY.md (in the model '
-      + 'directory), then `npx champollion sync` and `npx champollion verify` in the app, in a terminal '
-      + '(or the translate tool, method api, endpoint <that URL>/translate). A speaker reviews the strings '
-      + 'before anyone relies on them.';
+      + 'left to run. What is left is the app\'s: `npx champollion connect <model dir> --app <the app folder>` '
+      + '(ask the person which app; for a model trained on local-only sentences it needs --holders-agreed-by, '
+      + 'the holders\' own answer), then `npx champollion sync` and `npx champollion verify` in the app, in a '
+      + 'terminal (or the translate tool, method api, endpoint <that URL>/translate). A speaker reviews the '
+      + 'strings before anyone relies on them.';
   }
   if (a.state === 'no-dev-set' && /get_training_guardrails/.test(cmd)) {
     // forge names the guardrails ONCE, here, before the split (Round 13):

package/src/tools/harness.js

@@ -647,11 +647,18 @@ export function buildCorpusArgv(params, {
   }
   const localModel = method === 'local-model'
     || (method != null && methodEntries?.[method]?.kind === 'local-model');
-  if (attest_local_transport === true && localModel) {
-    throw new Error('attest_local_transport does not apply to method "local-model": the harness runs that '
-      + 'model in this process, so nothing leaves the machine and no attestation is needed (or recorded). '
+  // the dictionary floor: word-for-word lookup in the file given as `model`
+  const lexicon = method === 'lexicon'
+    || (method != null && methodEntries?.[method]?.kind === 'lexicon');
+  if (attest_local_transport === true && (localModel || lexicon)) {
+    throw new Error(`attest_local_transport does not apply to method "${method}": the harness runs it `
+      + 'in this process, so nothing leaves the machine and no attestation is needed (or recorded). '
       + 'Drop attest_local_transport.');
   }
+  if (lexicon && !model) {
+    throw new Error('method "lexicon" needs model: the dictionary file to look words up in (a .tsv, '
+      + 'source<TAB>target per line) — the harness refuses a run without one (-m)');
+  }
   if (model != null) {
     // A method's model is a name in ITS world (a Hugging Face id, a local
     // model directory, a plugin's own model name); an LLM provider's is a
@@ -669,9 +676,9 @@ export function buildCorpusArgv(params, {
   // An OPUS-MT pair model names its own pair (opus-mt-en-fi writes Finnish):
   // the harness refuses one whose pair visibly differs from the run's unless
   // the user asks for it (a related-language baseline).
-  if (allow_model_pair_mismatch === true && !localModel) {
-    throw new Error('allow_model_pair_mismatch applies to method "local-model" only (an OPUS-MT pair model '
-      + 'run against another pair) — drop it.');
+  if (allow_model_pair_mismatch === true && !localModel && !lexicon) {
+    throw new Error('allow_model_pair_mismatch applies to methods "local-model" (an OPUS-MT pair model run '
+      + 'against another pair) and "lexicon" (a dictionary whose sidecar names another pair) only — drop it.');
   }
   if (base_url != null) {
     if (!['local', 'openai'].includes(provider)) throw new Error('base_url applies only to provider "local" or "openai"');

package/src/tools/language-card.js

@@ -691,7 +691,7 @@ export function formatLanguage(r) {
     p.publishedAt && `published ${String(p.publishedAt).slice(0, 10)}`,
     p.sourceCount != null && `${p.sourceCount} cited sources`,
     p.coverage && p.coverage.total && `${p.coverage.present}/${p.coverage.total} card components present`,
-  ].filter(Boolean).join(' · ') || 'see the card'}. Card spec: https://champollion.dev/docs/network/specifications/language-cards`);
+  ].filter(Boolean).join(' · ') || 'see the card'}. Card spec: https://champollion.dev/docs/reference/language-card-spec`);
   out.push(`Next: language_overview { "code": "${s.code}" } — benchmarks, results, methods, constraints and the next steps for building a model.`);
   return out.join('\n');
 }

package/src/tools/languages.js

@@ -318,28 +318,35 @@ export function uncitedReason(lang, it = 'it', { brief = false } = {}) {
 }
 
 /**
- * The card's alternate names (ISO 639-3's list), as attributed facts —
- * capped, with the overflow counted. Codes and the language's own name are
- * not alternate names.
+ * The card's alternate names (ISO 639-3's name index, SIL langtags), as
+ * attributed facts — capped, with the overflow counted. A spelling several
+ * registries record is one fact citing every one of them (Montagnais: ISO
+ * 639-3 and SIL langtags), and the names more registries record come first:
+ * with dozens of names on some cards and four shown, the corroborated ones
+ * are the ones a reader should see. Codes and the language's own name are not
+ * alternate names.
  *
  * @returns {{facts: Array<{text: string, sources: string[]}>, more: number}}
  */
 function alternateNamesOf(card, ctx) {
-  const facts = [];
-  const seen = new Set([String(card?.name ?? '').toLowerCase()]);
-  let total = 0;
+  const byKey = new Map([[String(card?.name ?? '').toLowerCase(), null]]);
   for (const c of claimsFor(card, 'alternateNames', ['alternateNames'], ctx)) {
     for (const n of (Array.isArray(c.value) ? c.value : [c.value])) {
       // A bare lowercase 2–3 letter string is a code, not a name ("Ata" is a name).
       if (typeof n !== 'string' || !n.trim() || /^[a-z]{2,3}$/.test(n.trim())) continue;
       const key = n.trim().toLowerCase();
-      if (seen.has(key)) continue;
-      seen.add(key);
-      total += 1;
-      if (facts.length < ALT_NAME_CAP) facts.push({ text: n.trim(), sources: c.sources });
+      if (!byKey.has(key)) byKey.set(key, { text: n.trim(), sources: [] });
+      const fact = byKey.get(key);
+      if (!fact) continue; // the card's own name
+      for (const s of c.sources) if (!fact.sources.includes(s)) fact.sources.push(s);
     }
   }
-  return { facts, more: total - facts.length };
+  const all = [...byKey.values()].filter(Boolean)
+    // stable: equally attested names keep the card's (alphabetical) order
+    .map((f, i) => ({ f, i }))
+    .sort((a, b) => b.f.sources.length - a.f.sources.length || a.i - b.i)
+    .map(({ f }) => f);
+  return { facts: all.slice(0, ALT_NAME_CAP), more: all.length - Math.min(all.length, ALT_NAME_CAP) };
 }
 
 /**
@@ -368,11 +375,15 @@ export function renderAttributed(facts, { unstamped = 'source not stated on this
  * Every name a card records for a language, each with its source: the
  * registries' names (the `name` envelope — normalizeCard keeps only the
  * displayed one, so the claims are taken from the raw card BEFORE it is
- * normalized: `nameClaims`) and every endonym in the `endonym` envelope (a
+ * normalized: `nameClaims`), every endonym in the `endonym` envelope (a
  * community's own name is often not the one displayed: the Cree
- * macrolanguage's card shows ᓀᐦᐃᔭᐍᐏᐣ and also records nēhiyawēwin).
+ * macrolanguage's card shows ᓀᐦᐃᔭᐍᐏᐣ and also records nēhiyawēwin), and
+ * every alternate name (ISO 639-3's name index, SIL langtags) — one entry per
+ * spelling, citing every registry that records it, the shape the bundle's
+ * name-only entries carry (manifestSearchNames), so a match on one is said
+ * and cited the same way whichever index found it.
  *
- * @returns {Array<{text: string, field: 'name'|'endonym', source: string|null}>}
+ * @returns {Array<{text: string, field: 'name'|'endonym'|'alternate', source: string|null}>}
  */
 function recordedNames(card, ctx, nameClaims = []) {
   const out = [];
@@ -387,6 +398,17 @@ function recordedNames(card, ctx, nameClaims = []) {
   for (const c of nameClaims) add(c?.value, 'name', c?.source);
   for (const c of ctx.attributions?.(card?.endonym) ?? []) add(c?.value, 'endonym', c?.source);
   if (card?.nativeName) add(card.nativeName, 'endonym', null);
+  const alternate = new Map();
+  for (const c of claimsFor(card, 'alternateNames', ['alternateNames'], ctx)) {
+    for (const n of (Array.isArray(c.value) ? c.value : [c.value])) {
+      // A bare lowercase 2–3 letter string is a code, not a name.
+      if (typeof n !== 'string' || !n.trim() || /^[a-z]{2,3}$/.test(n.trim())) continue;
+      const sources = alternate.get(n.trim()) ?? [];
+      for (const s of c.sources) if (!sources.includes(s)) sources.push(s);
+      alternate.set(n.trim(), sources);
+    }
+  }
+  for (const [text, sources] of alternate) add(text, 'alternate', sources.join(', ') || null);
   return out;
 }
 
@@ -516,10 +538,7 @@ function entryFromCard(card, ctx, { nameClaims = [] } = {}) {
       : counts.length > 1 ? `${counts.join(' / ')} (sources differ)` : '');
   const fam = card.classification?.family;
   const code = card.code || card.iso639_3 || card.bcp47 || '';
-  const aliases = [
-    ...(Array.isArray(card.alternateNames) ? card.alternateNames : []),
-    ...(Array.isArray(card.aliases) ? card.aliases : []),
-  ];
+  const aliases = Array.isArray(card.aliases) ? card.aliases : [];
   const names = recordedNames(card, ctx, nameClaims);
   return {
     code,
@@ -534,13 +553,13 @@ function entryFromCard(card, ctx, { nameClaims = [] } = {}) {
     region: card.macroarea || '',
     typology: card.typologicalProfile?.verbSynthesis
       || card.typologicalProfile?.morphologicalSynthesis || '',
-    // Documented alternate names so a language is findable by ANY of its
-    // names — the same alias set the Atlas searches.
+    // Other CODES for the language. Its alternate NAMES are in `names`, each
+    // with its source — in `aliases` a match on one could not be cited.
     aliases,
     isolate: card.isIsolate === true,
-    // Every registry name and endonym, with its source: searchable, and a
-    // result found through one of them says which (Round 5 school persona:
-    // nêhiyawêwin found nothing).
+    // Every registry name, endonym and alternate name, with its source:
+    // searchable, and a result found through one of them says which (Round 5
+    // school persona: nêhiyawêwin found nothing).
     names,
     // Those names read in the working script of the language's registered
     // converter (Round 8: nêhiyawêwin found cre but never crk, whose card has
@@ -1147,8 +1166,11 @@ function matchTerms(index, qTokens, limit) {
         used.add(at);
       }
       if (used.size === words.length) {
+        // A name whose every word an earlier name already matched explains
+        // nothing new ("Cree, Plains" after "Plains Cree"), so it is not listed.
+        const adds = [...used].some((i) => !whole.has(i));
         for (const i of used) { whole.add(i); covered.add(i); }
-        if (!matched.some((m) => normalizeForMatch(m.text) === normalizeForMatch(text))) {
+        if (adds && !matched.some((m) => normalizeForMatch(m.text) === normalizeForMatch(text))) {
           matched.push({ text, field, ...(source ? { source } : {}), ...(derived ? { derived } : {}) });
         }
         if (field === 'name') ownName = 0;

package/src/tools/training.js

@@ -323,8 +323,8 @@ const COMMAND_ORDER = [
 // user cannot reach. The public page below is the same map, symptom first.
 const TAXONOMY_POINTER =
   'The failure map (what you see → the likely cause → the forge guard or lever '
-  + 'that fixes it): https://champollion.dev/docs/network/getting-started/'
-  + 'diagnosing-training. Every forge refusal also carries its own what/why/fix. '
+  + 'that fixes it): https://champollion.dev/docs/agents/training'
+  + '. Every forge refusal also carries its own what/why/fix. '
   + 'The forge_* MCP tools drive each guard; get_training_guardrails (this tool) '
   + 'is the rulebook behind them.';
 
@@ -399,9 +399,9 @@ export function formatTrainingGuardrails(answer) {
     + 'https://champollion.dev/docs/network/context/mt-training-concepts');
   parts.push(
     '- Step-by-step, agent-forward walkthrough: '
-    + 'https://champollion.dev/docs/network/tutorials/train-your-own-model');
+    + 'https://champollion.dev/docs/your-first-translator');
   parts.push(
     '- The guardrails on one page: '
-    + 'https://champollion.dev/docs/network/getting-started/training-honestly');
+    + 'https://champollion.dev/docs/agents/training');
   return parts.join('\n').trim();
 }

package/src/tools/translate.js

@@ -984,12 +984,12 @@ export async function translateTexts(args, deps = {}) {
       method,
       // The OpenRouter default slug only makes sense for the OpenRouter lane;
       // every other engine (direct APIs, a local server) applies its own
-      // default when model is unset — "google/gemini-3.8-flash" sent to Ollama
+      // default when model is unset — the OpenRouter default sent to Ollama
       // or to Anthropic's API is a guaranteed failure. A method that came from
       // the project brings the model sync would pair it with.
       model: model
         || (methodSource === 'project' ? projectWouldBe?.model : null)
-        || (method === 'llm' ? (champollion.DEFAULT_OPENROUTER_MODEL ?? 'google/gemini-3.8-flash') : undefined),
+        || (method === 'llm' ? champollion.DEFAULT_OPENROUTER_MODEL : undefined),
       batchSize: champollion.DEFAULT_BATCH_SIZE ?? 20,
       maxRetries: 2,
       ...reg,