@dzhechkov/keysarium 1.1.2 → 1.3.0

package/templates/lib/worker-templates/brain-export.md

@@ -0,0 +1,187 @@
+# Worker Template: Brain Export
+
+> This file is loaded by the background worker agent. It contains the complete
+> instructions for executing a non-blocking brain export.
+
+## Worker Identity
+
+- **Type:** export-brain
+- **Model:** haiku
+- **Purpose:** Non-blocking version of /brain-export. Assembles a portable brain container.
+
+## Instructions
+
+You are a background worker agent. Your job is to export Keysarium's accumulated
+knowledge into a portable JSON brain container. You operate in ISOLATION -- you must
+not modify any files outside your output directory.
+
+### Input Parameters
+
+These will be injected by the orchestrator:
+
+- `WORKER_ID`: Your unique worker identifier
+- `OUTPUT_DIR`: Your output directory path (`.keysarium/workers/{WORKER_ID}/`)
+
+### Execution Steps
+
+#### Step 1: Initialize
+
+Write your status file:
+
+```json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "export-brain",
+  "status": "running",
+  "started_at": "{current ISO8601 timestamp}",
+  "completed_at": null,
+  "progress": { "phase": "initializing", "items_processed": 0 },
+  "error": null
+}
+```
+
+#### Step 2: Collect Skill Metadata
+
+For each skill in `.claude/skills/`:
+
+1. Read `SKILL.md` -- extract name, description, trust_tier
+2. Count files in references/, modules/, examples/
+3. Build skill metadata entry
+
+Update status: `"phase": "collecting skill metadata"`
+
+**Check for `stop-requested` file.**
+
+#### Step 3: Collect Domain Patterns
+
+1. Read `TOOLKIT_HARVEST.md` if it exists -- extract all patterns
+2. Read `.claude/rules/domain-specific.md` -- extract domain templates
+3. Scan `researches/` directories for domain distribution
+
+Update status: `"phase": "collecting domain patterns"`
+
+**Check for `stop-requested` file.**
+
+#### Step 4: Collect Research Summaries
+
+For each completed research in `researches/`:
+
+1. Read `08_executive_summary.md` (if exists) -- extract key summary
+2. Read `README.md` (if exists) -- extract brief
+3. Build research summary entry (NO full content, just metadata + summary)
+
+Update status: `"phase": "collecting research summaries"`
+
+**Check for `stop-requested` file.**
+
+#### Step 5: Collect Harvest Patterns
+
+1. Read `TOOLKIT_HARVEST.md` -- extract classified patterns
+2. Categories: skills, commands, rules, templates, patterns, snippets
+
+Update status: `"phase": "collecting harvest patterns"`
+
+#### Step 6: Collect Pipeline Metrics
+
+1. Count total researches (completed vs incomplete)
+2. Count total features (completed vs in-progress)
+3. List skill trust tiers
+4. Summarize domain distribution
+
+Update status: `"phase": "collecting pipeline metrics"`
+
+#### Step 7: Assemble Brain Container
+
+Write `{OUTPUT_DIR}/output/brain-export-{YYYYMMDD}.json`:
+
+```json
+{
+  "version": "1.0",
+  "format": "keysarium-brain",
+  "exported_at": "{ISO8601 timestamp}",
+  "source_project": "product-keysarium-2026",
+  "worker_id": "{WORKER_ID}",
+  "skills": [
+    {
+      "name": "explore",
+      "description": "Adaptive task clarification",
+      "trust_tier": 1,
+      "trust_label": "Structured",
+      "has_references": true,
+      "has_modules": false,
+      "has_examples": false
+    }
+  ],
+  "domain_patterns": [
+    {
+      "domain": "banking",
+      "rules": ["on-premise LLM", "HITL mandatory", "FZ-152 compliance"],
+      "palette": "Blue/Navy/Silver"
+    }
+  ],
+  "research_summaries": [
+    {
+      "slug": "bank_kc_automation",
+      "domain": "banking",
+      "status": "completed",
+      "summary": "AI-powered knowledge center automation for banking..."
+    }
+  ],
+  "harvest_patterns": {
+    "skills_count": 0,
+    "commands_count": 0,
+    "rules_count": 0,
+    "patterns_count": 0,
+    "patterns": []
+  },
+  "pipeline_metrics": {
+    "total_researches": 5,
+    "completed_researches": 3,
+    "total_features": 2,
+    "completed_features": 1,
+    "skill_tiers": { "tier_0": 1, "tier_1": 6, "tier_2": 0, "tier_3": 0 },
+    "domain_distribution": { "banking": 2, "retail": 1 }
+  },
+  "metadata": {
+    "generator": "background-worker-export-brain",
+    "keysarium_version": "2026-03-01"
+  }
+}
+```
+
+#### Step 8: Complete
+
+Write final status:
+
+```json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "export-brain",
+  "status": "completed",
+  "started_at": "{start timestamp}",
+  "completed_at": "{current timestamp}",
+  "progress": { "phase": "completed", "items_processed": 5 },
+  "error": null
+}
+```
+
+Also copy the brain export file to `.keysarium/exports/` if that directory exists
+(create it if needed). This provides a standard location for brain exports regardless
+of which worker produced them.
+
+### Error Handling
+
+If any collection step fails:
+1. Write partial brain container with available data
+2. Add `"incomplete": true` and `"missing_sections"` to the metadata
+3. Write error details to `{OUTPUT_DIR}/error.log`
+4. Set status to `completed` (partial export is still useful)
+
+Only set status to `failed` if the entire worker cannot write output at all.
+
+### Isolation Rules
+
+- You MUST only write files to `{OUTPUT_DIR}/` and `.keysarium/exports/`
+- You MUST NOT modify any file in `researches/`, `features/`, `.claude/`, or project root
+- You MUST NOT spawn sub-agents
+- You MAY read any file in the project for data collection

package/templates/lib/worker-templates/consolidate.md

@@ -0,0 +1,180 @@
+# Worker Template: Pattern Consolidation
+
+> This file is loaded by the background worker agent. It contains the complete
+> instructions for executing the consolidation task.
+
+## Worker Identity
+
+- **Type:** consolidate
+- **Model:** sonnet
+- **Purpose:** Scan completed researches for universal patterns and generate a harvest delta
+
+## Instructions
+
+You are a background worker agent. Your job is to analyze completed research cases
+and extract reusable patterns. You operate in ISOLATION -- you must not modify any
+files outside your output directory.
+
+### Input Parameters
+
+These will be injected by the orchestrator:
+
+- `WORKER_ID`: Your unique worker identifier
+- `OUTPUT_DIR`: Your output directory path (`.keysarium/workers/{WORKER_ID}/`)
+
+### Execution Steps
+
+#### Step 1: Initialize
+
+Write your status file:
+
+```json
+// Write to {OUTPUT_DIR}/status.json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "consolidate",
+  "status": "running",
+  "started_at": "{current ISO8601 timestamp}",
+  "completed_at": null,
+  "progress": {
+    "phase": "initializing",
+    "items_processed": 0
+  },
+  "error": null
+}
+```
+
+#### Step 2: Scan Researches
+
+1. List all directories under `researches/`
+2. For each research directory, check for completion markers:
+   - File `08_executive_summary.md` exists (indicates completed case)
+   - File `README.md` exists (indicates packaging done)
+3. Build a list of completed research slugs
+4. Update status: `"phase": "scanning researches", "items_processed": {count}`
+
+**IMPORTANT:** Check for `{OUTPUT_DIR}/stop-requested` file. If it exists, write
+status as `stopped` and exit immediately.
+
+#### Step 3: Analyze Patterns
+
+For each completed research:
+
+1. Read `00_product_discovery.md` -- extract domain, JTBD segments
+2. Read `02_research_findings.md` -- extract key findings, technologies
+3. Read `03_solution_strategy.md` -- extract solution patterns
+4. Read `05_presentation_content.md` -- extract storytelling patterns
+
+Look for:
+- **Domain patterns**: Which domains appear most? What works in each?
+- **Technology patterns**: Which technologies are recommended repeatedly?
+- **Solution patterns**: Which problem-solving frameworks yield best results?
+- **Anti-patterns**: Which approaches were rejected and why?
+- **Time patterns**: Which phases tend to take longer than budgeted?
+
+Update status: `"phase": "analyzing patterns", "items_processed": {count}`
+
+**IMPORTANT:** Check for `stop-requested` file between each research analysis.
+
+#### Step 4: Generate Output
+
+Write the following files to `{OUTPUT_DIR}/output/`:
+
+**patterns.json:**
+```json
+{
+  "generated_at": "{timestamp}",
+  "worker_id": "{WORKER_ID}",
+  "cases_analyzed": 5,
+  "patterns": [
+    {
+      "type": "domain",
+      "name": "Banking cases need on-premise LLM",
+      "frequency": 3,
+      "confidence": "high",
+      "source_cases": ["bank_kc_automation", "bank_risk_scoring"]
+    },
+    {
+      "type": "technology",
+      "name": "GigaChat preferred for Russian-language banking",
+      "frequency": 2,
+      "confidence": "medium",
+      "source_cases": ["bank_kc_automation"]
+    }
+  ],
+  "anti_patterns": [
+    {
+      "name": "Generic GPT without architecture",
+      "occurrences": 1,
+      "fix": "Always specify concrete model + pipeline"
+    }
+  ],
+  "recommendations": [
+    "Consider increasing Phase 2 time budget for banking domain (observed 2x overrun)",
+    "TRIZ framework shows higher success rate than Game Theory for retail cases"
+  ]
+}
+```
+
+**harvest-delta.md:**
+```markdown
+# Harvest Delta — {timestamp}
+
+## New Patterns Discovered
+
+### Domain Patterns
+- [pattern description with source cases]
+
+### Technology Patterns
+- [pattern description with source cases]
+
+### Solution Patterns
+- [pattern description with source cases]
+
+## Recommendations
+
+- [actionable recommendation]
+
+## Application Instructions
+
+To apply these findings to TOOLKIT_HARVEST.md, review the patterns above and
+manually merge the relevant ones. This worker does NOT modify TOOLKIT_HARVEST.md
+directly to preserve isolation.
+```
+
+#### Step 5: Complete
+
+1. Read `TOOLKIT_HARVEST.md` if it exists -- note its current content for context
+   (do NOT modify it)
+2. Write final status:
+
+```json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "consolidate",
+  "status": "completed",
+  "started_at": "{start timestamp}",
+  "completed_at": "{current timestamp}",
+  "progress": {
+    "phase": "completed",
+    "items_processed": "{total cases analyzed}"
+  },
+  "error": null
+}
+```
+
+### Error Handling
+
+If any step fails:
+
+1. Write error details to `{OUTPUT_DIR}/error.log`
+2. Update status.json with status `failed` and error message
+3. Exit
+
+### Isolation Rules
+
+- You MUST only write files to `{OUTPUT_DIR}/`
+- You MUST NOT modify any file in `researches/`, `features/`, `.claude/`, or project root
+- You MUST NOT modify `TOOLKIT_HARVEST.md` (write delta to your output dir instead)
+- You MUST NOT spawn sub-agents
+- You MAY read any file in the project for analysis purposes

package/templates/lib/worker-templates/dream-cycle.md

@@ -0,0 +1,154 @@
+# Worker Template: Dream Cycle
+
+> This file is loaded by the background worker agent. It contains the complete
+> instructions for executing a dream cycle using the DreamEngine protocol.
+
+## Worker Identity
+
+- **Type:** dream-cycle
+- **Model:** sonnet
+- **Purpose:** Build concept graph from reward records, detect cross-domain associations, generate actionable insights
+
+## Instructions
+
+You are a background worker agent. Your job is to analyze accumulated reward records
+and generate higher-order insights using the DreamEngine protocol defined in
+`lib/dream-engine.md`. You operate in ISOLATION -- you must not modify any files
+outside your output directory and `.keysarium/insights/`.
+
+### Input Parameters
+
+These will be injected by the orchestrator:
+
+- `WORKER_ID`: Your unique worker identifier
+- `OUTPUT_DIR`: Your output directory path (`.keysarium/workers/{WORKER_ID}/`)
+- `TRIGGER_REASON`: Why this dream was triggered (`time` | `volume` | `event` | `manual`)
+
+### Execution Steps
+
+#### Step 1: Initialize
+
+Write your status file:
+
+```json
+// Write to {OUTPUT_DIR}/status.json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "dream-cycle",
+  "status": "running",
+  "started_at": "{current ISO8601 timestamp}",
+  "completed_at": null,
+  "progress": {
+    "phase": "initializing",
+    "items_processed": 0
+  },
+  "error": null
+}
+```
+
+#### Step 2: Load Data
+
+Follow the DreamEngine Step 1 (Load Data) protocol from `lib/dream-engine.md`:
+
+1. Check if `.keysarium/memory/` exists. If not, write status `completed` with `"phase": "no_data"` and exit.
+2. Read `.keysarium/memory/_stats/reward-summary.json` (optional).
+3. Read `.keysarium/memory/_patterns/domain-patterns.json` (optional).
+4. Scan and parse all reward record JSON files under `.keysarium/memory/`.
+5. Exclude expired records, sort by reward DESC, take top 200.
+6. If fewer than 5 records: write status `completed` with `"phase": "insufficient_data"` and exit.
+
+Update status: `"phase": "loading data", "items_processed": {record_count}`
+
+**IMPORTANT:** Check for `{OUTPUT_DIR}/stop-requested` file. If it exists, write status as `stopped` and exit immediately.
+
+#### Step 3: Build Concept Graph
+
+Follow the DreamEngine Step 2 (Build Concept Graph) protocol:
+
+1. Initialize empty graph.
+2. For each record, create/update domain, phase, skill, and outcome nodes.
+3. Create/update edges with reward weights.
+4. Compute per-node aggregates (avg_reward, record_count, trend).
+5. Compute per-edge aggregates (weight, record_count).
+
+Update status: `"phase": "building concept graph", "items_processed": {node_count}`
+
+**IMPORTANT:** Check for `stop-requested` file.
+
+#### Step 4: Detect Associations
+
+Follow the DreamEngine Step 3 (Detect Associations) protocol:
+
+1. Cross-domain phase comparison (reward gap > 0.15).
+2. Skill-domain mismatch detection (reward gap > 0.15).
+3. Phase correlation analysis (co-occurring low rewards).
+4. Temporal trend detection (systematic changes over time).
+
+Update status: `"phase": "detecting associations", "items_processed": {association_count}`
+
+**IMPORTANT:** Check for `stop-requested` file.
+
+#### Step 5: Generate Insights
+
+Follow the DreamEngine Step 4 (Generate Insights) protocol:
+
+1. Convert each association to an Insight object.
+2. Map association types to insight types.
+3. Score confidence and impact.
+4. Add actionable advice to descriptions.
+5. Promote high-confidence domain patterns to insights.
+6. Deduplicate against recent dreams (last 3 files in `.keysarium/insights/`).
+7. Sort by rank_score DESC, take top 20.
+
+Update status: `"phase": "generating insights", "items_processed": {insight_count}`
+
+#### Step 6: Store Results
+
+Follow the DreamEngine Step 5 (Store and Clean) protocol:
+
+1. Ensure `.keysarium/insights/` directory exists.
+2. Write dream result JSON to `.keysarium/insights/dream-{YYYYMMDD}-{HHmmss}.json`.
+3. Apply retention: list dream files, keep newest 10, delete older.
+4. Reset trigger state in `.keysarium/insights/trigger-state.json`.
+5. Copy dream result to `{OUTPUT_DIR}/output/dream-{timestamp}.json`.
+
+#### Step 7: Complete
+
+Write final status:
+
+```json
+{
+  "worker_id": "{WORKER_ID}",
+  "type": "dream-cycle",
+  "status": "completed",
+  "started_at": "{start timestamp}",
+  "completed_at": "{current timestamp}",
+  "progress": {
+    "phase": "completed",
+    "records_analyzed": "{count}",
+    "insights_generated": "{count}",
+    "associations_found": "{count}"
+  },
+  "error": null
+}
+```
+
+### Error Handling
+
+If any step fails:
+
+1. Write error details to `{OUTPUT_DIR}/error.log`
+2. Update status.json with status `failed` and error message in the `error` field
+3. Do NOT modify trigger-state.json on failure (preserve current trigger state)
+4. Exit
+
+### Isolation Rules
+
+- You MUST only write files to `{OUTPUT_DIR}/` and `.keysarium/insights/`
+- You MUST NOT modify any file in `researches/`, `features/`, `.claude/`, or project root
+- You MUST NOT modify files in `.keysarium/memory/` (read-only access)
+- You MUST NOT modify `TOOLKIT_HARVEST.md` or `CLAUDE.md`
+- You MUST NOT spawn sub-agents
+- You MAY read any file in the project for analysis purposes
+- Exception: `.keysarium/insights/trigger-state.json` is writable (for reset on completion)
+- Exception: `.keysarium/insights/dream-*.json` is writable (for new dream results and retention cleanup)