@mindflight/mindbrain-personal-studio 0.6.2 → 0.6.3

package/examples/immeuble/scripts/verify-immeuble-live-artifacts.mjs

@@ -0,0 +1,140 @@
+#!/usr/bin/env node
+/**
+ * Verify live projection artifacts on a running MindBrain backend.
+ *
+ * Usage:
+ *   node examples/immeuble/scripts/verify-immeuble-live-artifacts.mjs
+ *     [--workspace immeuble] [--url http://127.0.0.1:8091]
+ */
+
+import { readFileSync, mkdirSync, writeFileSync } from "node:fs";
+import { dirname, join, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+import { parseYaml } from "./yaml-lite.mjs";
+import { spawnSync } from "node:child_process";
+
+const immeubleRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
+const pkgRoot = resolve(immeubleRoot, "..", "..");
+const gcp = join(pkgRoot, "bin", "gcp.mjs");
+const reportsDir = join(immeubleRoot, "reports");
+const acceptancePath = join(immeubleRoot, "ACCEPTANCE.yaml");
+
+const args = process.argv.slice(2);
+const workspaceId = parseFlag(args, "--workspace-id", "immeuble");
+const backendUrl = parseFlag(args, "--url", process.env.GHOSTCRAB_MINDBRAIN_URL ?? "http://127.0.0.1:8091");
+
+const acceptance = parseYaml(readFileSync(acceptancePath, "utf8"));
+const requiredArtifacts = acceptance?.projections?.live_answer_view ?? [];
+const result = { ok: true, workspace_id: workspaceId, checks: [] };
+
+try {
+  const list = runArtifactCmd(["artifact", "list", "--url", backendUrl, "--workspace-id", workspaceId, "--kind", "live_answer_view"]);
+  result.checks.push({ name: "artifact.list", ok: true, detail: `status=${list.status}` });
+
+  const listPayload = parseJson(list.stdout);
+  const artifacts = Array.isArray(listPayload?.artifacts) ? listPayload.artifacts : [];
+  const present = new Set(artifacts.map((a) => a?.artifact_id).filter(Boolean));
+
+  const missing = requiredArtifacts.filter((artifactId) => !present.has(artifactId));
+  result.checks.push({
+    name: "required_artifacts_present",
+    ok: missing.length === 0,
+    detail: missing.length ? `missing: ${missing.join(", ")}` : "all required ids listed"
+  });
+  if (missing.length > 0) {
+    throw new Error("missing required live_answer_view artifacts");
+  }
+
+  const refreshChecks = [];
+  for (const artifactId of requiredArtifacts) {
+    const refresh = runArtifactCmd(["artifact", "refresh", "--url", backendUrl, artifactId]);
+    const refreshPayload = parseJson(refresh.stdout);
+    const refreshArtifactId = refreshPayload?.artifact_id || refreshPayload?.artifact?.artifact_id;
+    if (typeof refreshArtifactId === "string" && refreshArtifactId.length > 0) {
+      result.checks.push({ name: `artifact.refresh.${artifactId}`, ok: true, detail: `artifact_id=${refreshArtifactId}` });
+    } else {
+      result.checks.push({ name: `artifact.refresh.${artifactId}`, ok: true, detail: "refresh returned JSON payload" });
+    }
+    refreshChecks.push(refreshPayload);
+
+    const get = runArtifactCmd(["artifact", "get", "--url", backendUrl, artifactId]);
+    const getPayload = parseJson(get.stdout);
+    const payload = extractPayload(getPayload) ?? extractPayload(refreshPayload);
+    const artifactSummary = typeof payload?.summary === "string" && payload.summary.trim().length > 0;
+    const artifactRefreshChecks = Array.isArray(payload?.refresh_checks) && payload.refresh_checks.length > 0;
+    const refreshedOperational =
+      refreshPayload?.ok === true &&
+      payload &&
+      (typeof payload.graph_entities === "number" ||
+        typeof payload.facts === "number" ||
+        typeof payload.graph_relations === "number");
+    const payloadOk = (artifactSummary && artifactRefreshChecks) || refreshedOperational;
+    result.checks.push({
+      name: `artifact.payload.${artifactId}`,
+      ok: payloadOk,
+      detail: JSON.stringify({
+        has_summary: artifactSummary,
+        has_refresh_checks: artifactRefreshChecks,
+        refreshed_operational: refreshedOperational
+      })
+    });
+    if (!payloadOk) {
+      throw new Error(`artifact payload invalid for ${artifactId}`);
+    }
+  }
+
+  result.checks.push({ name: "refresh_count", ok: refreshChecks.length === requiredArtifacts.length, detail: `${refreshChecks.length}` });
+  result.ok = result.checks.every((item) => item.ok);
+} catch (err) {
+  result.ok = false;
+  result.error = err instanceof Error ? err.message : String(err);
+}
+
+mkdirSync(reportsDir, { recursive: true });
+writeFileSync(
+  join(reportsDir, "live-artifacts-refresh.validation.json"),
+  JSON.stringify(result, null, 2) + "\n",
+  "utf8"
+);
+console.log(JSON.stringify(result, null, 2));
+process.exit(result.ok ? 0 : 1);
+
+function runArtifactCmd(argsList) {
+  const res = spawnSync(process.execPath, [gcp, "brain", ...argsList], {
+    cwd: pkgRoot,
+    env: {
+      ...process.env,
+      GHOSTCRAB_MINDBRAIN_URL: backendUrl
+    },
+    encoding: "utf8"
+  });
+  if (res.status !== 0) {
+    throw new Error(`gcp brain ${argsList.join(" ")} failed (${res.status}): ${(res.stderr || res.stdout || "").trim()}`);
+  }
+  return { status: res.status, stdout: res.stdout || "", stderr: res.stderr || "" };
+}
+
+function parseJson(text) {
+  return JSON.parse(text || "{}");
+}
+
+function extractPayload(getPayload) {
+  const payloadJson = getPayload?.payload_json;
+  if (typeof payloadJson === "string") {
+    try {
+      return JSON.parse(payloadJson);
+    } catch {
+      return null;
+    }
+  }
+  if (getPayload?.payload && typeof getPayload.payload === "object") {
+    return getPayload.payload;
+  }
+  return null;
+}
+
+function parseFlag(argv, name, defaultValue) {
+  const index = argv.indexOf(name);
+  if (index === -1) return defaultValue;
+  return argv[index + 1] ?? defaultValue;
+}

package/examples/immeuble/scripts/yaml-lite.mjs

@@ -0,0 +1,96 @@
+export function parseYaml(text) {
+  const root = {};
+  const stack = [{ indent: -1, value: root }];
+  const lines = String(text)
+    .split(/\r?\n/)
+    .map((raw) => raw.replace(/\s+#.*$/, ""))
+    .filter((raw) => raw.trim() && !raw.trim().startsWith("#"));
+
+  for (const raw of lines) {
+    const indent = raw.match(/^ */)?.[0].length ?? 0;
+    const line = raw.trim();
+    while (stack.length > 1 && indent <= stack[stack.length - 1].indent) stack.pop();
+    const parent = stack[stack.length - 1].value;
+
+    if (line.startsWith("- ")) {
+      if (!Array.isArray(parent)) continue;
+      const body = line.slice(2).trim();
+      if (body.includes(":")) {
+        const [key, rest] = splitKeyValue(body);
+        const item = {};
+        item[key] = parseScalar(rest);
+        parent.push(item);
+        stack.push({ indent, value: item });
+      } else {
+        parent.push(parseScalar(body));
+      }
+      continue;
+    }
+
+    const [key, rest] = splitKeyValue(line);
+    if (!key) continue;
+    if (rest === "") {
+      const next = nextMeaningfulLine(lines, raw);
+      const value = next?.trim().startsWith("- ") ? [] : {};
+      parent[key] = value;
+      stack.push({ indent, value });
+    } else {
+      parent[key] = parseScalar(rest);
+    }
+  }
+  return root;
+}
+
+function nextMeaningfulLine(lines, currentRaw) {
+  const currentIndex = lines.indexOf(currentRaw);
+  return lines.slice(currentIndex + 1).find((line) => line.trim());
+}
+
+function splitKeyValue(line) {
+  const index = line.indexOf(":");
+  if (index === -1) return [line.trim(), ""];
+  return [line.slice(0, index).trim(), line.slice(index + 1).trim()];
+}
+
+function parseScalar(raw) {
+  const value = raw.trim();
+  if (value === "") return "";
+  if (value === "true") return true;
+  if (value === "false") return false;
+  if (value === "null") return null;
+  if (/^-?\d+(\.\d+)?$/.test(value)) return Number(value);
+  if (value.startsWith("[") && value.endsWith("]")) {
+    const inner = value.slice(1, -1).trim();
+    if (!inner) return [];
+    return splitInlineList(inner).map(parseScalar);
+  }
+  if (
+    (value.startsWith("\"") && value.endsWith("\"")) ||
+    (value.startsWith("'") && value.endsWith("'"))
+  ) {
+    try {
+      return JSON.parse(value);
+    } catch {
+      return value.slice(1, -1);
+    }
+  }
+  return value;
+}
+
+function splitInlineList(value) {
+  const parts = [];
+  let current = "";
+  let quote = "";
+  for (const char of value) {
+    if ((char === "\"" || char === "'") && !quote) quote = char;
+    else if (char === quote) quote = "";
+    if (char === "," && !quote) {
+      parts.push(current.trim());
+      current = "";
+    } else {
+      current += char;
+    }
+  }
+  if (current.trim()) parts.push(current.trim());
+  return parts;
+}

package/examples/immeuble/sources/agent-prompts/prompts/00-prerequisites-immo-mcp.md

@@ -0,0 +1,161 @@
+# Prerequisites — MCP lab immeuble (workspace `immo-mcp`)
+
+**Phase 0 — read only, no writes.**
+
+Variant of [`00-prerequisites.md`](00-prerequisites.md) with fixed IDs for workspace `immo-mcp`.
+
+## Human pre-flight (before opening the agent)
+
+1. **GhostCrab backend** reachable (`ghostcrab_status` must respond).
+2. **MCP `ghostcrab-personal-mcp`** enabled in Cursor — same SQLite as this lab (see below).
+3. **Single SQLite file** — do **not** use a separate `immo-mcp.sqlite`. Cursor MCP (`.cursor/mcp.json`) starts:
+
+```text
+gcp brain up --db /home/dlamotte/Documents/ghostcrab-personal-mcp/data/ghostcrab.sqlite
+```
+
+Lab process data lives in workspace **`immo-mcp`** inside that file. Golden reference uses workspace **`immeuble`** in the same file.
+
+For CLI commands outside Cursor, align explicitly:
+
+```bash
+export GHOSTCRAB_SQLITE_PATH="/home/dlamotte/Documents/ghostcrab-personal-mcp/data/ghostcrab.sqlite"
+# optional: gcp brain db-who  → must show the same path
+```
+
+4. **Golden workspace** (phase 06 comparison only — never load into `immo-mcp`):
+
+```bash
+export GHOSTCRAB_SQLITE_PATH="/home/dlamotte/Documents/ghostcrab-personal-mcp/data/ghostcrab.sqlite"
+node bin/gcp.mjs load examples/immeuble/bundle/immeuble.bundle.json \
+  --workspace immeuble --reindex all \
+  --db "$GHOSTCRAB_SQLITE_PATH"
+```
+
+MCP and CLI must target the **same** `--db` / `GHOSTCRAB_SQLITE_PATH` for the whole lab.
+
+## Two workspaces, one SQLite — comparison model
+
+| Workspace | Role |
+|-----------|------|
+| **`immo-mcp`** | What GhostCrab MCP **built** from corpus (phases 2–5) |
+| **`immeuble`** | What we **expect** — golden from `bundle/immeuble.bundle.json` (load in phase 6 only) |
+
+Both live in **`ghostcrab.sqlite`**. Phase 6 compares `immo-mcp` (process) vs `immeuble` (reference) vs [`success-criteria.yaml`](../success-criteria.yaml). See [`06-validate-and-compare-immo-mcp.md`](06-validate-and-compare-immo-mcp.md).
+
+Never load the bundle into `immo-mcp`.
+
+## Run identifiers
+
+| Field | Value |
+|-------|-------|
+| `workspace_id` | `immo-mcp` |
+| `collection_id` | `immo-mcp::docs` |
+| `ontology_id` | `immeuble::core` |
+| `golden_workspace_id` | `immeuble` |
+| `golden_bundle` | `examples/immeuble/bundle/immeuble.bundle.json` |
+| `corpus` | `examples/immeuble/sources/documents/` (8 md files) |
+| `criteria` | `examples/immeuble/success-criteria.yaml` |
+
+## Tools
+
+1. `ghostcrab_status`
+2. `ghostcrab_modeling_guidance`
+
+## Agent prompt (copy-paste)
+
+```
+Tu es un agent GhostCrab MCP en session vierge. Ta mission : reconstruire le domaine
+syndic belge (Résidence Les Tilleuls + Les Érables) depuis le corpus brut, en suivant
+strictement le MCP lab documenté.
+
+## Documentation de référence (lire en premier)
+
+1. docs/explanation/mcp-lab-context.md — contexte, phases 00→06, critères
+2. examples/immeuble/README.md — point d'entrée opérationnel
+3. examples/immeuble/sources/agent-prompts/prompts/00-prerequisites-immo-mcp.md — démarrage (ce fichier)
+
+## Identifiants de cette run
+
+| Champ | Valeur |
+|-------|--------|
+| workspace_id | immo-mcp |
+| collection_id | immo-mcp::docs |
+| ontology_id | immeuble::core |
+| golden_workspace_id | immeuble |
+| golden_bundle | examples/immeuble/bundle/immeuble.bundle.json |
+| corpus | examples/immeuble/sources/documents/ (8 fichiers md) |
+| critères | examples/immeuble/success-criteria.yaml |
+
+## Règles absolues
+
+- Appeler ghostcrab_status en PREMIER.
+- Lire avant d'écrire : count → search → pack (ou graph_search pour le graphe).
+- NE PAS charger le bundle golden dans immo-mcp — le golden (immeuble) sert
+  uniquement à comparer en phase 06.
+- Phases 00–01 : lecture seule — aucune écriture.
+- Avant toute écriture (phases 02–05) : produire un Model Proposal (entités, arêtes,
+  facettes, questions de compétence) et ATTENDRE ma confirmation explicite
+  (ONBOARDING_CONTRACT §9).
+- Ingestion documentaire : CLI gcp brain document (pas streaming MCP unitaire).
+- Outils graph/gap extended : découvrir via ghostcrab_tool_search si absents du défaut.
+
+## Workflow — exécuter dans l'ordre
+
+Phase 0 — examples/immeuble/sources/agent-prompts/prompts/00-prerequisites-immo-mcp.md
+  → ghostcrab_status, ghostcrab_modeling_guidance
+  → lire corpus/manifest.json, success-criteria.yaml, contracts/scenarios.yaml
+
+Phase 1 — prompts/01-discovery-and-model-proposal.md
+  → affiner le Model Proposal depuis le corpus (toujours sans écriture)
+
+GATE : stop et présente le Model Proposal. N'écris rien tant que je n'ai pas confirmé.
+
+Phase 2 — prompts/02-ontology-register.md (après confirmation)
+  → gcp brain ontology compile sur ontologies/immeuble/core.yaml
+
+Phase 3 — prompts/03-gap-rules-design.md
+  → ghostcrab_graph_gap_rules_import (réf. training/reference gap-rules)
+
+Phase 4 — prompts/04-document-ingest.md
+  → gcp brain document : collection-create, ingest 8 corpus, profile, qualify
+
+Phase 5 — prompts/05-graph-extraction.md
+  → ghostcrab_learn (incrémental) OU document-business-extract (batch)
+
+Phase 6 — prompts/06-validate-and-compare-immo-mcp.md
+  → ghostcrab_graph_search, ghostcrab_graph_diagnostics vs success-criteria.yaml
+  → comparer immo-mcp (process) vs immeuble (bundle golden), même SQLite
+
+## Livrables attendus par phase
+
+- Phase 0–1 : Model Proposal textuel
+- Phase 2–5 : rapport court (ce qui a été écrit, outils utilisés, counts)
+- Phase 6 : tableau écart vs success-criteria.yaml + recommandations
+
+## Commence maintenant
+
+1. ghostcrab_status
+2. Résume mcp-lab-context.md en 5 lignes
+3. Lis corpus/manifest.json et success-criteria.yaml
+4. Propose un premier Model Proposal (sans écrire)
+```
+
+## Required reading
+
+- `corpus/manifest.json`
+- `success-criteria.yaml`
+- `../contracts/scenarios.yaml`
+
+## Deliverable
+
+Model Proposal textuel (entités, arêtes, facettes, competency questions couvertes).
+
+## Gate
+
+**Confirmation humaine requise** — reply **« je confirme »** (ou corrections) before `01-discovery-and-model-proposal.md` refinement and before any write phases (ONBOARDING_CONTRACT §9).
+
+## Docs if blocked
+
+- [`01-reference-to-graph`](../../../docs/explanation/01-reference-to-graph.md)
+- [`mcp-lab-context`](../../../docs/explanation/mcp-lab-context.md)

package/examples/immeuble/sources/agent-prompts/prompts/00-prerequisites.md

@@ -0,0 +1,39 @@
+# Prerequisites — MCP lab immeuble
+
+**Phase 0 — read only, no writes.**
+
+## Tools
+
+1. `ghostcrab_status`
+2. `ghostcrab_modeling_guidance`
+
+## Agent prompt (copy-paste)
+
+```
+Je veux reconstruire le domaine syndic belge "Résidence Les Tilleuls + Les Érables"
+depuis le corpus brut dans examples/immeuble/sources/documents/manifest.json.
+Workspace cible : immeuble.
+Référence de validation : examples/immeuble/bundle/immeuble.bundle.json (bundle golden immeuble).
+
+Ne charge pas le bundle golden dans le workspace LLM — compare seulement à la fin.
+Propose un Model Proposal (entités, relations, facettes documentaires) avant toute écriture.
+```
+
+## Required reading
+
+- `manifest.json`
+- `success-criteria.yaml`
+- `../contracts/scenarios.yaml`
+- `workspace.json`
+
+## Deliverable
+
+Model Proposal textuel (entités, arêtes, facettes, competency questions couvertes).
+
+## Gate
+
+**Confirmation humaine requise** avant de passer à `01-discovery-and-model-proposal.md` puis aux phases d'écriture (ONBOARDING_CONTRACT §9).
+
+## Next
+
+→ [`01-discovery-and-model-proposal.md`](01-discovery-and-model-proposal.md)

package/examples/immeuble/sources/agent-prompts/prompts/01-discovery-and-model-proposal.md

@@ -0,0 +1,42 @@
+# Discovery and model proposal
+
+**Phase 1 — read corpus, propose model (no writes yet).**
+
+## Tools
+
+- File reads: `corpus/*.md`, `corpus/manifest.json`
+- Optional: `ghostcrab_tool_search` if you need specialized tools
+
+## Agent prompt
+
+```
+Lis les 8 fichiers listés dans examples/immeuble/sources/documents/manifest.json.
+
+Produis une table avec :
+1. Entity types (building, block, unit, person, household, lease_contract, cellar, parking_space, …)
+2. Edge types (contains, owns, occupies, leases, assigned_cellar, assigned_garage, …)
+3. Facet dimensions documentaires (source.document_type, domain.building, domain.unit, domain.role, …)
+4. Mapping doc_id → document_type → entités/relations attendues
+5. Competency questions couvertes (réf. examples/immeuble/contracts/scenarios.yaml)
+
+Compare à ../checklists/ontology-checklist.md. Si tu t'écartes, explique pourquoi.
+Ne crée rien dans GhostCrab tant que le Model Proposal n'est pas validé.
+```
+
+## Deliverable
+
+| Column | Content |
+|--------|---------|
+| Entity types | List with brief definition |
+| Edge types | List with source/target |
+| Facet dimensions | Namespace.dimension + value examples |
+| Doc mapping | 8 rows from manifest |
+| Scenarios | Which scenario IDs each doc supports |
+
+## Gate
+
+Human validation of Model Proposal before `02-ontology-register.md`.
+
+## Next
+
+→ [`02-ontology-register.md`](02-ontology-register.md)

package/examples/immeuble/sources/agent-prompts/prompts/02-ontology-register.md

@@ -0,0 +1,44 @@
+# Ontology register
+
+**Phase 2 — create workspace and register ontology.**
+
+## Tools
+
+- `ghostcrab_workspace_create`
+- `ghostcrab_schema_register` or `ghostcrab_schema_inspect` / `ghostcrab_schema_list`
+- CLI (documented): `gcp brain ontology compile …`
+
+## Agent prompt
+
+```
+Crée le workspace immeuble (domain_profile: syndic).
+Enregistre l'ontology immeuble::core alignée sur le Model Proposal validé :
+
+Option A — LinkML :
+  gcp brain ontology compile \
+    --workspace-id immeuble \
+    --ontology-id immeuble::core \
+    --input ontologies/immeuble/core.yaml \
+    --profile syndic \
+    --import-db --force
+
+Option B — MCP schema :
+  ghostcrab_schema_register avec facet dimensions du Model Proposal
+
+Vérifie avec ghostcrab_schema_list / ghostcrab_schema_inspect sur immeuble.
+Ne charge pas examples/immeuble/bundle/immeuble.bundle.json dans ce workspace.
+```
+
+## Reference (read-only)
+
+Canonical LinkML: `ontologies/immeuble/core.yaml`
+Checklist: `reference/ontology-checklist.md`
+
+## Deliverable
+
+- Workspace `immeuble` exists
+- Ontology `immeuble::core` visible in workspace
+
+## Next
+
+→ [`03-gap-rules-design.md`](03-gap-rules-design.md)

package/examples/immeuble/sources/agent-prompts/prompts/03-gap-rules-design.md

@@ -0,0 +1,44 @@
+# Gap rules design
+
+**Phase 3 — design and import closed-world gap rules.**
+
+## Tools
+
+- `ghostcrab_graph_gap_rules_import`
+- `ghostcrab_graph_gap_rules`
+- Optional: `ghostcrab_graph_diagnostics` (smoke on empty graph)
+
+## Agent prompt
+
+```
+Conçois les graph_gap_rules closed-world pour workspace immeuble, ontology immeuble::core.
+
+Niveau 1 — Patrimoine (cf. examples/immeuble/gap-rules/L0-patrimoine.json) :
+  - unit-one-cellar, unit-in-building, garage-at-most-one-unit
+
+Niveau 2 — Syndic filtré (cf. `../gap-rules/L2-syndic-filtered.json`) :
+  - unit-has-owner, occupied-unit-has-occupant, tenant-occupied-has-lease
+  - entity_filter sur usage_status (exclure vacant / vacant_works)
+
+Importe avec replace:true via ghostcrab_graph_gap_rules_import.
+Liste les rules actives via ghostcrab_graph_gap_rules.
+Explique chaque rule_id en langage métier (cf. `../gap-rules/closed-world-contract.md`).
+
+Pédagogie optionnelle : importer d'abord L1-syndic-naive.json, diagnostiquer le FP sur
+Érables Appartement A4 (vacant_works), puis passer à L2.
+```
+
+## Reference files
+
+- `../gap-rules/demo.json` — patrimoine demo
+- `../gap-rules/syndic.json` — syndic L2 for immeuble
+- `../gap-rules/` — curriculum L0–L3
+- `../checklists/gap-rules-checklist.md`
+
+## Deliverable
+
+JSON gap-rules imported; list of active rule_ids with business labels.
+
+## Next
+
+→ [`04-document-ingest.md`](04-document-ingest.md)

package/examples/immeuble/sources/agent-prompts/prompts/04-document-ingest.md

@@ -0,0 +1,40 @@
+# Document ingest
+
+**Phase 4 — enqueue, profile, and qualify source documents.**
+
+## Tools
+
+MCP: `ghostcrab_workspace_create` (if needed)
+CLI: `gcp brain document …` (profile + qualify pipeline)
+
+## Agent prompt
+
+```
+Pour chaque entrée de examples/immeuble/sources/documents/ :
+
+1. collection-create --workspace-id immeuble --collection-id immeuble::docs
+2. ontology-attach --ontology-id immeuble::core --role taxonomy
+3. document-profile-enqueue --content-file examples/immeuble/sources/documents/<filename> \
+     --doc-id <doc_id> --language fr
+4. document-profile-worker --limit <n>
+5. qualification-vocab-list
+6. document-qualify --taxonomies immeuble::core \
+     --facets source.document_type,domain.building,domain.unit,domain.role,domain.scenario,finance.payment_status
+
+Commence par 1 document si itération prompts (--limit-docs 1).
+Vérifie les qualifications par doc_id avant de continuer.
+```
+
+## Mock CI reference
+
+```bash
+node examples/immeuble/scripts/run-immeuble-import.mjs --apply --skip-provenance-validation --skip-preflight
+```
+
+## Deliverable
+
+8 documents profiled and qualified in `immeuble::docs`.
+
+## Next
+
+→ [`05-graph-extraction.md`](05-graph-extraction.md)

package/examples/immeuble/sources/agent-prompts/prompts/05-graph-extraction.md

@@ -0,0 +1,44 @@
+# Graph extraction
+
+**Phase 5 — extract business entities and relations from qualified documents.**
+
+## Tools
+
+- `ghostcrab_remember` / `ghostcrab_upsert`
+- `ghostcrab_graph_search`
+- Reindex: `gcp load` partial or workspace reindex CLI
+
+## Agent prompt
+
+```
+À partir des documents qualifiés dans `immeuble`, extrais le graphe métier syndic :
+
+Structure :
+  - 2 immeubles (Tilleuls, Érables), 13 lots, personnes, ménages
+  - 13 caves, garages sélectifs, jardins privatifs RDC
+  - 5 contrats de bail, 3 écritures CODA
+
+Relations minimales :
+  contains, owns, occupies, household_member, leases, rented_to,
+  assigned_cellar, assigned_garage, uses_exclusive, matched_to
+
+Utilise ghostcrab_remember / ghostcrab_upsert ou l'extraction LLM documentée
+dans le flux d'ingestion `run-immeuble-import.mjs`.
+Reindex le graphe après écriture.
+
+Vérifie :
+  ghostcrab_graph_search query "appartement" → ≥ 13 unités
+  ghostcrab_graph_search query "Dupont" → résultats pertinents
+```
+
+## Success thresholds
+
+See `success-criteria.yaml` → `entity_counts`, `relation_edges`.
+
+## Deliverable
+
+Populated `graph_entity` for workspace `immeuble`; counts near success criteria.
+
+## Next
+
+→ [`06-validate-and-compare.md`](06-validate-and-compare.md)