@yeaft/webchat-agent 0.1.408 → 0.1.410

package/unify/memory/dream-prompt.js

@@ -0,0 +1,272 @@
+/**
+ * dream-prompt.js — Dream prompt templates for each phase
+ *
+ * Dream has 5 phases:
+ *   Phase 1: Orient  — assess current memory state
+ *   Phase 2: Gather  — collect recent context
+ *   Phase 3: Merge   — combine duplicates, update outdated
+ *   Phase 4: Prune   — remove stale/low-value entries
+ *   Phase 5: Promote — extract patterns, update profile
+ *
+ * Reference: yeaft-unify-core-systems.md §3.3
+ */
+
+/**
+ * Build the Orient phase prompt (Phase 1).
+ * The LLM assesses the current memory state and identifies issues.
+ *
+ * @param {{ memorySummary: string, profileContent: string, entryCount: number }} context
+ * @returns {string}
+ */
+export function buildOrientPrompt({ memorySummary, profileContent, entryCount }) {
+  return `You are in Dream Mode — Phase 1: Orient.
+
+Your task is to assess the current state of the memory store and identify what needs attention.
+
+## Current Memory State
+
+${memorySummary}
+
+## MEMORY.md (User Profile)
+
+${profileContent || '(empty)'}
+
+## Assessment Instructions
+
+Review the memory state and provide:
+1. **Redundancies**: Are there entries that overlap or say the same thing?
+2. **Outdated info**: Are there entries that might be stale or no longer relevant?
+3. **Gaps**: Is there important context missing from MEMORY.md?
+4. **Quality**: Are entries well-categorized (kind, scope, tags)?
+
+Return your assessment as JSON:
+{
+  "redundantGroups": [["entry-a", "entry-b"]],
+  "potentiallyStale": ["entry-name-1"],
+  "profileGaps": ["missing X context"],
+  "qualityIssues": ["entry-y has wrong kind"],
+  "overallHealth": "good" | "needs-attention" | "poor",
+  "suggestedActions": ["merge entries about X", "prune stale context entries"]
+}
+
+Return ONLY valid JSON, no other text.`;
+}
+
+/**
+ * Build the Gather phase prompt (Phase 2).
+ * Collects recent compact summaries and completed task summaries.
+ *
+ * @param {{ recentCompact: string, completedTasks: object[], orientResult: object }} context
+ * @returns {string}
+ */
+export function buildGatherPrompt({ recentCompact, completedTasks, orientResult }) {
+  const taskSummaries = completedTasks.length > 0
+    ? completedTasks.map(t => `- [${t.id}] ${t.description}: ${t.summary || '(no summary)'}`).join('\n')
+    : '(no recently completed tasks)';
+
+  return `You are in Dream Mode — Phase 2: Gather.
+
+Your task is to identify what new information should be incorporated into long-term memory.
+
+## Recent Conversation Summary (compact.md)
+
+${recentCompact || '(no recent summaries)'}
+
+## Recently Completed Tasks
+
+${taskSummaries}
+
+## Orient Assessment
+
+${JSON.stringify(orientResult, null, 2)}
+
+## Instructions
+
+From the recent conversations and tasks, identify:
+1. **New facts** worth remembering (project structure, tech decisions)
+2. **New preferences** expressed by the user
+3. **New skills/lessons** learned during tasks
+4. **Context updates** (project progress, status changes)
+
+Return as JSON:
+{
+  "newEntries": [
+    { "name": "slug-name", "kind": "fact|preference|skill|lesson|context|relation", "scope": "path", "tags": ["tag1", "tag2"], "importance": "high|normal|low", "content": "description" }
+  ],
+  "updatesToExisting": [
+    { "entryName": "existing-slug", "updates": { "content": "updated text", "tags": ["new-tag"] } }
+  ]
+}
+
+Return ONLY valid JSON, no other text.`;
+}
+
+/**
+ * Build the Merge phase prompt (Phase 3).
+ *
+ * @param {{ duplicateGroups: object[][], gatherResult: object }} context
+ * @returns {string}
+ */
+export function buildMergePrompt({ duplicateGroups, gatherResult }) {
+  const groupDescriptions = duplicateGroups.map((group, i) => {
+    const entries = group.map(e =>
+      `  - [${e.name}] kind=${e.kind}, scope=${e.scope}, tags=[${(e.tags || []).join(', ')}]\n    ${(e.content || '').slice(0, 200)}`
+    ).join('\n');
+    return `Group ${i + 1}:\n${entries}`;
+  }).join('\n\n');
+
+  return `You are in Dream Mode — Phase 3: Merge.
+
+Your task is to merge duplicate/overlapping entries into single, richer entries.
+
+## Potentially Duplicate Groups
+
+${groupDescriptions || '(no duplicates detected)'}
+
+## New Entries from Gather Phase
+
+${JSON.stringify(gatherResult?.newEntries || [], null, 2)}
+
+## Instructions
+
+For each duplicate group:
+1. Decide if they should be merged (combine info) or kept separate (different enough)
+2. For merges, create a single entry that preserves all important info from both
+3. List which old entries should be deleted after merge
+
+Also process the new entries from Gather — check if any overlap with existing entries.
+
+Return as JSON:
+{
+  "merges": [
+    {
+      "merged": { "name": "new-slug", "kind": "...", "scope": "...", "tags": [], "importance": "...", "content": "..." },
+      "deleteOriginals": ["old-entry-1", "old-entry-2"]
+    }
+  ],
+  "newEntries": [
+    { "name": "...", "kind": "...", "scope": "...", "tags": [], "importance": "...", "content": "..." }
+  ],
+  "updates": [
+    { "entryName": "existing-slug", "updates": { "content": "updated text" } }
+  ]
+}
+
+Return ONLY valid JSON, no other text.`;
+}
+
+/**
+ * Build the Prune phase prompt (Phase 4).
+ *
+ * @param {{ staleEntries: object[], entryCount: number, maxEntries: number }} context
+ * @returns {string}
+ */
+export function buildPrunePrompt({ staleEntries, entryCount, maxEntries }) {
+  const staleDescriptions = staleEntries.map(e =>
+    `- [${e.name}] kind=${e.kind}, scope=${e.scope}, freq=${e.frequency || 1}, days_since_update=${e._daysSinceUpdate}\n  ${(e.content || '').slice(0, 150)}`
+  ).join('\n');
+
+  return `You are in Dream Mode — Phase 4: Prune.
+
+Your task is to remove stale, low-value, or redundant entries.
+
+## Potentially Stale Entries (${staleEntries.length} found)
+
+${staleDescriptions || '(none detected)'}
+
+## Capacity
+
+Current entries: ${entryCount}
+Maximum allowed: ${maxEntries}
+${entryCount > maxEntries ? `⚠️ OVER CAPACITY by ${entryCount - maxEntries} entries — must prune aggressively` : 'Within capacity'}
+
+## Prune Guidelines
+
+Delete entries that are:
+- **Outdated context**: Project status from weeks ago
+- **Never recalled**: frequency=1 and old — nobody needs it
+- **Too vague**: "user mentioned something about X" without useful detail
+- **Redundant with profile**: If MEMORY.md already captures it
+- **Re-derivable**: Info that can be obtained by running a command (e.g., "Node version is 20")
+
+KEEP entries that are:
+- High importance or high frequency
+- Recent preferences or lessons
+- Facts about project structure (hard to re-discover)
+
+Return as JSON:
+{
+  "toDelete": ["entry-name-1", "entry-name-2"],
+  "reasoning": {
+    "entry-name-1": "outdated context from 45 days ago",
+    "entry-name-2": "never recalled, too vague"
+  }
+}
+
+Return ONLY valid JSON, no other text.`;
+}
+
+/**
+ * Build the Promote phase prompt (Phase 5).
+ *
+ * @param {{ entries: object[], profileContent: string, scopesSummary: string }} context
+ * @returns {string}
+ */
+export function buildPromotePrompt({ entries, profileContent, scopesSummary }) {
+  // Find entries that might form patterns
+  const highFreq = entries
+    .filter(e => (e.frequency || 1) >= 3)
+    .map(e => `- [${e.name}] kind=${e.kind}, freq=${e.frequency}, scope=${e.scope}: ${(e.content || '').slice(0, 150)}`)
+    .join('\n');
+
+  const lessons = entries
+    .filter(e => e.kind === 'lesson')
+    .map(e => `- [${e.name}] scope=${e.scope}: ${(e.content || '').slice(0, 150)}`)
+    .join('\n');
+
+  return `You are in Dream Mode — Phase 5: Promote.
+
+Your task is to identify patterns and update the user profile.
+
+## High-Frequency Entries (recalled ≥3 times)
+
+${highFreq || '(none)'}
+
+## All Lessons
+
+${lessons || '(none)'}
+
+## Current MEMORY.md Profile
+
+${profileContent || '(empty)'}
+
+## Scopes
+
+${scopesSummary}
+
+## Instructions
+
+1. **Pattern promotion**: If multiple entries share a pattern, create a higher-level insight
+   - Example: 3 entries about "user corrects indentation" → 1 preference: "default to 2-space indent"
+2. **Profile update**: Update MEMORY.md sections based on accumulated knowledge
+   - Keep MEMORY.md under 200 lines
+   - Sections: Facts, Preferences, Project Context, Skills, Lessons
+3. **Scope promotion**: If a lesson applies across projects, promote scope to parent or global
+
+Return as JSON:
+{
+  "profileUpdates": {
+    "Facts": ["- New fact line 1"],
+    "Preferences": ["- New preference line"],
+    "Project Context": [],
+    "Skills": [],
+    "Lessons": []
+  },
+  "promotedEntries": [
+    { "name": "...", "kind": "...", "scope": "global", "tags": [], "importance": "high", "content": "..." }
+  ],
+  "entriesToDelete": ["entry-that-was-promoted-to-profile"]
+}
+
+Return ONLY valid JSON, no other text.`;
+}