01 / Real cloud sandbox
Real financial server
Not a simulated terminal, but a real cloud financial server.
Each user receives an independent Sandbox. Tasks can pause and resume while the Workspace persists. The Agent runs under a low-privilege account, with financial credentials isolated by a separate operating-system identity and file permissions.
persistent · Asia/Shanghai
Preinstalled research and delivery environment
Python pandas NumPy SciPy statsmodels PyArrow
Plotly Matplotlib WeasyPrint openpyxl python-docx PDF tools
ripgrep jq Noto CJK
fin and vl sync by content digest at startup and run through controlled launchers.
02 / Workspace
Memory that keeps growing
Memory is a directory tree.
Workspace stores organized long-term memory. Earlier conversations remain in session storage and are recalled only when needed.
/home/user/workspace
workspace/
├── AGENTS.md
├── user.yaml # identity and structured preferences
├── portfolio/ # accounts, holdings, trades, watchlists
├── stocks/ # research organized by security
├── topics/ # sectors, macro, strategies, themes
├── briefs/ # recurring briefs
├── inbox/ # user inputs, read only
├── evidence/ # data evidence, read only
└── skills/ # user-installed capabilities
User model Profile, preferences, constraints, and output habits update only through explicit requests.
Investment state Portfolio files stay structured, with dedicated write rules loaded before updates.
Research output Single-security work goes to stocks; cross-security and thematic work goes to topics.
Procedural memory User Skills keep methods, scripts, and assets inside their own system.
Evidence boundary The data plane writes to evidence; the Agent can read it but cannot overwrite the source.
Cross-session recall When earlier work must continue, real exchanges from prior sessions are retrieved on demand.
Context freeze The first run freezes only the map, profile, and Skill index—not the entire history.
Separate records from facts Old sessions prove what was said then; prices, filings, and policy still require fresh data.
Files are the source of truth; context is only a cache for the current step.
03 / fin CLI · Data plane
Professional financial data
Financial data and evidence archives.
Connect current multi-asset quotes, price history and technical indicators, financial news, corporate events, and institutional research through one interface. Every call returns a concise summary and writes the complete result to evidence for later review.
fin · financial data
Market data quote · technical
Financials financials · call
Information event · news · research
General websearch
Manifest driven Data-source definitions stay centralized and consistent
Evidence coverage coverage / degraded / as_of
Deterministic failures Exit codes tell the Agent what to do next
Credential isolation Sensitive credentials never enter the Sandbox
04 / Skills
Professional research workflows
Indexes stay resident; full instructions load on demand.
A Skill can define evidence checklists, reasoning order, output templates, executable scripts, and acceptance rules. Analysts write the evidence route, methods, scripts, templates and acceptance into a Skill, and the Agent follows that professional method.
Research listed-company-analysis Analyze listed-company fundamentals, earnings, valuation, and competitive position.
Decision trade-decision Evaluate adding, trimming, chasing, stopping out, or holding a single security.
Deliverable Company one-pager A buy-side one-pager with evidence, timestamps, and visible coverage gaps.
05 / KOL Agent
Financial KOL Agents
An expert Agent is more than a scripted character.
A financial expert Agent is a complete Skill Package. Stable methods, investment principles, calculation scripts, evaluation rules, and searchable source material determine how it researches and communicates.
Skill Package skills/kol-agent/
kol-agent/
├── SKILL.md # principles and research roadmap
├── references/ # principles and historical views
├── scripts/ # formulas, data retrieval, calculation, validation
├── evals/ # decision consistency and style evaluation
└── .data/ # searchable source material
01 Master index and research roadmap SKILL.md summarizes the core principles and routes tasks to the right references, scripts, data, and evaluations.
02 Traceable investment principles references/ stores specific principles, historical calls, and the conditions under which they apply.
03 Compute and retrieve scripts/ carries formulas and data retrieval; .data/ supports progressive search through underlying material.
04 Continuously evaluated evals/ checks factual quality, decision consistency, style fidelity, and capability boundaries.
06 / Scheduled research
Always-on research schedules
Start research automatically on schedule.
Set the task, run time, time zone, and model. Even after the page is closed, the system creates a new conversation at the scheduled time and runs the complete task.
Daily / weekdays Morning market briefing Summarize overnight markets, important events, portfolio risks, and the day’s watchpoints before the open.
Daily 08:00 · Asia/ShanghaiWeekly Investment review and strategy check Run portfolio reviews, sector tracking, or a fixed stock-selection workflow on a chosen day.
Sunday 20:00 · Selected modelOne time Automatic post-event follow-up Schedule one research run after earnings, a rate decision, or another known event, then end automatically.
Selected date · Local time zone
current · available now
Every run is a traceable new conversation
The task uses the latest Workspace in its space and can call fin and Skills. The full process and result appear in conversation history, and reports, data, and charts can be written back to the workspace.
Schedule → New conversation → Workspace + fin + Skills → Results / files
