strategic-insight-longform

# Strategic Insight Longform v4.2.1 S-T-D Cube + Multi-Dimensional Analysis Matrix + GoT Adaptive Path Optimization. Slimmed from 513→~250 lines; agent pipeline and roster moved to references. ## Modes | Mode | Agents | Time | Trigger | |------|--------|------|---------| | **Deep** (default) | 16 full + GoT | 12-20 min | Default execution | | **Standard** | 9 core | 7-10 min | User explicitly requests | | **Quick** | 5 core | 3-5 min | User explicitly requests | ## Architecture (High Level) ``` Leader (main process) ├─ Serial: topic-preprocessor → knowledge-enricher → framework-builder → got-controller(D) ├─ Parallel: 3-5 researchers (spatial/temporal/domain/stakeholder/causal) ├─ Serial: source-manager(D) → insight-synthesizer → longform-writer → output-finalizer │ └─ Stage 6.5: obsidian-md-ac formatting (emoji titles, callouts, Mermaid, YAML) ├─ Parallel(D): memory-curator + pattern-crystallizer └─ Copy final file to Obsidian inbox ``` **Full 16-agent pipeline**: see `references/agent-pipeline.md` **Agent roster**: see `references/agent-roster.md` ## Framework Building (Stage 1) framework-builder selects frameworks via two-tier selection: 1. **3 mandatory frameworks**: S-T-D Cube (always) + analysis_type specific (see `references/framework-library.md` for ~75 frameworks across 10 categories) 2. **3-5 supplementary frameworks**: Auto-recommended by analysis_type ## Knowledge Enhancement (Stage 0.5) 1. **qmd vsearch**: Search Obsidian knowledge base (relevance > 0.5) 2. **Historical analysis**: Match from memory-context.json 3. **Exa multi-layer search**: web_search_exa → company_research_exa → crawling_exa 4. **Degradation**: qmd unavailable → skip vector search; Exa unavailable → fallback to web_search Output: `knowledge-context.json` for all subsequent agents. ## Output Quality ### insight-synthesizer - Second-order inference: "What does this lead to?" - Third-order inference: chain effects - Cross-dimensional matrix: S*T, S*D, T*D combos - Counterfactual analysis: "What if X didn't happen / went the opposite way?" ### longform-writer - No hard word limit → content completeness check - Chain-of-thought embedded: main CoT (after holographic summary) + per-section callouts - Source index appendix integrated at end - Insight hierarchy fidelity: never compress or omit multi-layer inference chains - Wikilinks naturally embedded from knowledge-context - Appendices: data summary, causal chain diagram, game matrix, scenario hypothesis, source index ### output-finalizer - Single file output: `战略洞察-{title}.md` - Revision loop: score < 4.0 → max 1 revision round - Argument quality: logical gaps, implicit assumptions, counter-evidence, confidence annotation - Auto-save to Obsidian inbox ### source-manager (CoV anti-hallucination, Deep only) - **Layer 1**: Citation completeness — scan all factual claims (target 100% coverage) - **Layer 2**: Source reachability — crawling_exa check URLs, flag dead links - **Layer 3**: Cross-validation — 5-10 key claims, web_search_exa for independent confirmation ## Memory System 6 JSON files in `memory/`: topics.json, sources.json, frameworks.json, sessions.json, preferences.json, patterns.json. Schema: `references/memory-schema.md`. ## Learning System 5 pattern types (framework_effectiveness, source_reliability, writing_optimization, analysis_depth, topic_association). Crystallization: confidence ≥0.85 + ≥3 occurrences → auto-activate; 0.70-0.85 → pending; <0.70 → discard. Daily decay 0.01; >30 days unused → auto-degrade. ## Scheduling Rules All agents dispatched via `delegate_task` (Hermes) or `Task` (Claude Code). Agent `.md` `tools:` lists are documentation only. See `references/agent-pipeline.md` for full dispatch logic. ## References | File | Content | |------|---------| | `references/agent-pipeline.md` | Full 16-agent dispatch pipeline with pseudo-code | | `references/agent-roster.md` | 16 agent definitions, stages, modes | | `references/framework-library.md` | ~75 analysis frameworks, 10 categories | | `references/memory-schema.md` | Memory JSON schemas | | `references/std-cube-methodology.md` | S-T-D methodology | | `config.json` | Execution modes, agents, memory, learning, CoT, output, quality thresholds | ## Verification Checklist - [ ] Mode selected (Deep/Standard/Quick) based on user input? - [ ] framework-builder selected 3 mandatory + 3-5 supplementary frameworks? - [ ] All researcher agents completed before insight-synthesizer starts? - [ ] Source verification (CoV) completed in Deep mode? - [ ] output-finalizer score ≥4.0? If not, revision loop completed? - [ ] Final file saved to Obsidian inbox? *Strategic Insight Longform v4.2.1 — slimmed for Hermes compliance* --- ## Deployment & Sync After ANY update: `cd ~/code/jz-skills && ./deploy/sync-back.sh && git commit -am "sync: strategic-insight-longform" && git push`