Predict Analysis Skill

> Previously known as swarm_intelligence. The skill id > swarm_intelligence is retained as an alias for backward > compatibility. See docs/PREDICT_ANALYSIS.md for the full technical > walkthrough.

The team

This skill uses the largest team of any skill — 50 agents in total — orchestrated via the shared Agent Swarm Service (core/engine/agent_swarm.py).

| Role(s) | Count | What they do | |---|---|---| | Asset classifier | 1 | Identifies the asset + its price drivers | | Context analyser | 1 | Extracts regime + key levels from bars | | Intelligence gatherer | 1 | Web-searches news / analysis / regulation / indicators | | Personas (bull/bear/neutral/observer) | 50 | Debate the asset for 30 rounds | | Cross-examiner | 1 | Probes divergent personas with targeted questions | | Reporter | 1 | Synthesises final research note |

All coordination — parallelism, timeouts, retries, event recording — is handled by the Agent Swarm Service, not this skill.

Pipeline (5 stages, ~10-30 minutes total)

1. Context Analysis — classify asset + extract market context + build 6 specialisation data feeds
2. Intelligence Gathering — 4 web searches → synthesise bull/bear briefing
3. Persona Generation — 50 personas with distinct backgrounds, biases, influence weights, specialisations, tool access
4. Iterative Research — each persona plans its own research queries (min 3, max 8) using their assigned tools
5. Multi-Round Debate — 30 rounds × 15 speakers with per-agent memory + selective thread routing
6. Cross-Examination — press the 6-8 most divergent personas with targeted questions
7. ReACT Report — synthesise + apply influence-weighted consensus math

Multi-chart (portfolio) mode

When the Canvas has multiple chart windows, the focused chart is the primary asset (drives the full pipeline); siblings are summarised into the intel briefing as portfolio context. Personas reference them naturally in their arguments.

See docs/PREDICT_ANALYSIS.md § 5 for the processor-level normalisation (focused → index 0, missing-dataset warnings, etc.).

Tool calls emitted

| Tool | When | Purpose | |---|---|---| | simulation.set_debate | On completion | Push full debate payload to the store | | bottom_panel.activate_tab | On completion | Switch to DAG Graph tab | | notify.toast | On completion | Toast with consensus summary |

Output tabs

| Tab | Shows | |---|---| | DAG Graph | React Flow pipeline visualisation | | Personalities | 50 persona cards; click → full profile + research trail + live /interview chat | | Debate Thread | Flat list of all messages with sentiment bars + tool chips + agreement references | | Run Stats | Consensus + briefing + market context + data feeds + cross-exams + convergence chart + PDF export + Run Warnings banner |

Input

• Natural language: "run a swarm debate on BTC", "predict direction for CL=F", "what does the committee think about AAPL?"
• Requires at least one dataset loaded on the Canvas.
• Optional: message text is passed through as additional context into every persona's prompt.

Known limitations

See docs/PREDICT_ANALYSIS.md § 13. Summary: no streaming (user waits for full 10-30 min run), no persona caching (every run regenerates), global DDG rate limiter serialises web searches, no cross-session memory.