@mindrian_os/install 1.13.0-beta.11

package/lib/core/chat-context-builder.cjs

@@ -0,0 +1,253 @@
+'use strict';
+
+/**
+ * chat-context-builder.cjs -- SQL-targeted context injection for the BYO
+ * chat panel (Phase 87-09). Produces a {systemPrompt, userPrompt,
+ * tokenEstimate} triple sized to stay under 5000 tokens per call -- a
+ * ~57x reduction vs. slurping the whole room.
+ *
+ * BSL 1.1. Copyright (c) Mindrian 2026.
+ * (Business Source License 1.1; SPDX BUSL-1.1; see LICENSE.)
+ *
+ * Five intent patterns (all return tokenEstimate on every path):
+ *   1. contradicts  -- top CONTRADICTS edges from room.db
+ *   2. converges    -- top CONVERGES edges from room.db
+ *   3. stakeholders -- graceful early-return "no data yet" when the
+ *                       stakeholders table is empty (R6 / 84-05 Pattern 3)
+ *   4. gaps         -- sections with the fewest indexed artifacts
+ *   5. briefing     -- fallback: top-confidence edges + artifact excerpts
+ *
+ * Schema notes (real room.db, from lib/core/lazygraph-ops.cjs):
+ *   - edges(source, target, type, properties)  -- NOT source_id/target_id
+ *   - nodes(id, type, properties)  -- Artifact nodes have id = "section/slug"
+ *     and properties JSON carries title + section; this IS the "artifacts"
+ *     view the plan's pseudocode referenced.
+ *   - stakeholders(id, type, name, canonical_ref, notes, metadata,
+ *     created_at, updated_at)
+ *
+ * The 5K budget is enforced by capping the JSON-stringified edge list and
+ * per-artifact excerpt size. Every return path computes tokenEstimate =
+ * Math.ceil((systemPrompt.length + userPrompt.length) / 4).
+ */
+
+const path = require('node:path');
+const fs = require('node:fs');
+const { DatabaseSync } = require('node:sqlite');
+
+const CTX_CHAR_BUDGET = 18000;         // ~4500 tokens at 4 char/token
+const PER_ARTIFACT_CHAR_CAP = 1500;    // cap each artifact excerpt
+const EDGE_JSON_CAP = 3000;            // cap JSON.stringify(edges, null, 2)
+
+/**
+ * Classify a free-text user message into one of the five intent buckets.
+ * Default fallback: briefing.
+ * @param {string} msg
+ * @returns {'contradicts'|'converges'|'stakeholders'|'gaps'|'briefing'}
+ */
+function classifyIntent(msg) {
+  const m = (msg || '').toLowerCase();
+  if (/contradic|conflict|inconsist/.test(m)) return 'contradicts';
+  if (/converg|align|pattern|theme/.test(m)) return 'converges';
+  if (/said|who|attribut|stakeholder/.test(m)) return 'stakeholders';
+  if (/gap|fill|next|where should/.test(m)) return 'gaps';
+  if (/briefing|intelligence|summary|overview/.test(m)) return 'briefing';
+  return 'briefing';
+}
+
+/**
+ * Read an artifact file with a hard size cap.
+ * @param {string} filePath
+ * @returns {string}
+ */
+function readArtifactSafe(filePath) {
+  try {
+    const s = fs.readFileSync(filePath, 'utf8');
+    return s.length > PER_ARTIFACT_CHAR_CAP
+      ? (s.slice(0, PER_ARTIFACT_CHAR_CAP) + '\n[truncated]')
+      : s;
+  } catch (_) {
+    return '';
+  }
+}
+
+/**
+ * Approximate token count from char length (4 chars per token heuristic).
+ * @param {string} systemPrompt
+ * @param {string} userPrompt
+ * @returns {number}
+ */
+function estimateTokens(systemPrompt, userPrompt) {
+  return Math.ceil(((systemPrompt || '').length + (userPrompt || '').length) / 4);
+}
+
+/**
+ * Derive a best-effort section name from an Artifact node id of the form
+ * "section/slug". Returns null if the id has no slash.
+ * @param {string} id
+ * @returns {string|null}
+ */
+function sectionOfId(id) {
+  if (!id || typeof id !== 'string') return null;
+  const i = id.indexOf('/');
+  return i > 0 ? id.slice(0, i) : null;
+}
+
+/**
+ * Parse a node properties JSON string, defensively.
+ * @param {string} s
+ * @returns {object}
+ */
+function safeProps(s) {
+  if (!s) return {};
+  try { return JSON.parse(s); } catch (_) { return {}; }
+}
+
+/**
+ * Build a chat context for `userMessage` scoped to `roomDir`. Returns
+ *   { systemPrompt, userPrompt, tokenEstimate, earlyReturn, intent }
+ * where tokenEstimate is a number on EVERY code path (including fallbacks
+ * and early-returns). Never throws on missing db; it returns a minimal
+ * context instead.
+ *
+ * @param {string} roomDir Absolute path to the active room.
+ * @param {string} userMessage Raw user input (post-trim).
+ * @param {object} [options] Reserved for future filters.
+ * @returns {Promise<{systemPrompt:string, userPrompt:string, tokenEstimate:number, earlyReturn:boolean, intent:string}>}
+ */
+async function buildContext(roomDir, userMessage, options = {}) {
+  const intent = classifyIntent(userMessage);
+  const dbPath = path.join(roomDir, '.mindrian', 'room.db');
+
+  if (!fs.existsSync(dbPath)) {
+    const systemPrompt =
+      'You are Larry -- the PWS methodology teaching partner. ' +
+      'The active room has no graph data yet (room.db missing).';
+    return {
+      systemPrompt,
+      userPrompt: userMessage,
+      tokenEstimate: estimateTokens(systemPrompt, userMessage),
+      earlyReturn: false,
+      intent,
+    };
+  }
+
+  const db = new DatabaseSync(dbPath, { readOnly: true });
+  try {
+    // Fetch intent-specific rows. Every query returns an array of plain
+    // objects with bounded shape; JSON.stringify cap below ensures budget.
+    let rows = [];
+
+    if (intent === 'contradicts') {
+      rows = db.prepare(
+        "SELECT source, target, type, properties FROM edges WHERE type = 'CONTRADICTS' LIMIT 5"
+      ).all();
+    } else if (intent === 'converges') {
+      rows = db.prepare(
+        "SELECT source, target, type, properties FROM edges WHERE type = 'CONVERGES' LIMIT 5"
+      ).all();
+    } else if (intent === 'stakeholders') {
+      const n = db.prepare('SELECT COUNT(*) AS n FROM stakeholders').get().n;
+      if (n === 0) {
+        // R6 / 84-05 Pattern 3: graceful fallback. Do NOT fabricate
+        // attribution when the stakeholders table is empty.
+        const sp = 'You are Larry -- the PWS methodology teaching partner.';
+        const up =
+          'Stakeholder attribution has not been populated for this room yet. ' +
+          'File meetings via /mos:file-meeting to populate.';
+        return {
+          systemPrompt: sp,
+          userPrompt: up,
+          tokenEstimate: estimateTokens(sp, up),
+          earlyReturn: true,
+          intent,
+        };
+      }
+      rows = db.prepare(
+        'SELECT id, type, name, canonical_ref, notes FROM stakeholders LIMIT 10'
+      ).all();
+    } else if (intent === 'gaps') {
+      // Sections with the fewest Artifact nodes = candidate gaps. Derive
+      // the section from each Artifact node's id prefix (schema stores
+      // section in the node id path).
+      const artifactRows = db.prepare(
+        "SELECT id FROM nodes WHERE type = 'Artifact'"
+      ).all();
+      const perSection = new Map();
+      for (const r of artifactRows) {
+        const sec = sectionOfId(r.id);
+        if (!sec) continue;
+        perSection.set(sec, (perSection.get(sec) || 0) + 1);
+      }
+      rows = [...perSection.entries()]
+        .map(([section, entry_count]) => ({ section, entry_count }))
+        .sort((a, b) => a.entry_count - b.entry_count)
+        .slice(0, 5);
+    } else {
+      // briefing: top edges as a general overview
+      rows = db.prepare(
+        'SELECT source, target, type, properties FROM edges LIMIT 10'
+      ).all();
+    }
+
+    // Fetch a few referenced Artifact nodes by id (best-effort).
+    const referencedIds = [...new Set(
+      rows.flatMap((r) => [r.source, r.target]).filter(Boolean)
+    )].slice(0, 3);
+
+    let artifactExcerpts = '';
+    for (const aid of referencedIds) {
+      const a = db.prepare(
+        "SELECT id, properties FROM nodes WHERE id = ? AND type = 'Artifact'"
+      ).get(aid);
+      if (!a) continue;
+      const props = safeProps(a.properties);
+      const title = props.title || a.id;
+      // The real on-disk file path is derived from the node id:
+      //   id "section/slug"  ->  roomDir/section/slug.md (current layout)
+      //   id may also reference a section-level node (skipped above).
+      const candidatePaths = [
+        path.join(roomDir, aid + '.md'),
+        path.join(roomDir, aid, path.basename(aid) + '.md'),
+      ];
+      let body = '';
+      for (const p of candidatePaths) {
+        body = readArtifactSafe(p);
+        if (body) break;
+      }
+      if (body) {
+        artifactExcerpts += '\n\n--- ' + title + ' ---\n' + body;
+      }
+    }
+
+    const systemPrompt = [
+      'You are Larry -- the PWS methodology teaching partner for MindrianOS.',
+      'Voice: conversational, provocative, concise. 3-8 sentences default.',
+      '',
+      'ACTIVE ROOM: ' + path.basename(roomDir),
+      'INTENT: ' + intent,
+      '',
+      'RELEVANT GRAPH DATA (from SQL query matching user question):',
+      JSON.stringify(rows, null, 2).slice(0, EDGE_JSON_CAP),
+      '',
+      'ARTIFACT EXCERPTS (only the ones referenced by the edges above):',
+      artifactExcerpts.slice(0, CTX_CHAR_BUDGET - EDGE_JSON_CAP - 500),
+    ].join('\n');
+
+    return {
+      systemPrompt,
+      userPrompt: userMessage,
+      tokenEstimate: estimateTokens(systemPrompt, userMessage),
+      earlyReturn: false,
+      intent,
+    };
+  } finally {
+    try { db.close(); } catch (_) { /* ignore */ }
+  }
+}
+
+module.exports = {
+  buildContext,
+  classifyIntent,
+  estimateTokens,
+  CTX_CHAR_BUDGET,
+};