Local macOS research dashboard prototype: staged runs, evidence checks and consolidated reports

[MetaVenturesCompany]

I have been building a local macOS dashboard around TradingAgents, with AI coding assistance, to make the research workflow easier to operate and its outputs easier to review.

My starting point was a local installation around v0.7.1. The dashboard lives separately from the original installation. It is a personal prototype, not an official TradingAgents release or a published, installable extension.

I’m sharing the approach and lessons learned to see which parts would be useful to other users or worth preparing as focused contributions.

1. A simpler local interface

The dashboard opens through a Mac app or Apple Shortcut, without having to start TradingAgents manually in Terminal.

It provides ticker/date selection, analyst selection, research depth, progress, cancellation and access to saved reports. Opening the interface does not start research.

2. Stage-by-stage research controls

The research workflow has four selectable groups:

1. Analyst reports: market, fundamentals, news and sentiment, as selected.
2. Bull/bear debate and research manager.
3. Trader proposal.
4. Risk committee and portfolio manager.

Users can select the full workflow, run selected stages or pause for review between stages.

A preview shows which stages will run, which saved outputs will be reused and which prerequisites are missing. Later stages require compatible upstream results; missing stages are not silently added.

Rerunning an upstream stage invalidates its old downstream conclusions. Saved state uses JSON snapshots and hashes, with compatibility checks covering relevant engine/dashboard code, configuration and dependency metadata. Changing research settings clears both the preview and usage approval.

3. Consolidated HTML research reports

A separate offline step combines saved TradingAgents research with manually supplied Perplexity research into one HTML report.

The report includes executive summary, main-body analyst summaries, fundamentals, valuation, news, sentiment, thesis, scenarios, risk assessment and an evidence register. Original research text is retained in expandable appendices, with source snapshots and hashes.

This is deterministic consolidation and validation—not another independent model analyst. Perplexity research is pasted or imported manually; no Perplexity API integration is enabled. Consolidation itself makes no model or financial-data calls.

4. Financial consistency and evidence checks

The validation layer checks supported inputs for P/E, EPS, growth, enterprise-value definitions, scenario arithmetic and ATR/stop distance.

One useful regression example from saved research was:

• Reference price: $206.69
• Claimed EPS: $5.84
• Claimed P/E: 26.13×

The first two imply approximately 35.39×, so the stated P/E must be reconciled.

Other safeguards include:

• Distinguishing arithmetic agreement from verified inputs.
• Checking eligible local accounting evidence against the requested as-of date.
• Flagging unsupported consensus forecasts and unresolved earnings bases.
• Treating missing news or sentiment as unknown, rather than neutral evidence.
• Labelling FY2028 exit-multiple valuations as future-year, undiscounted figures rather than present-day targets.
• Keeping incomplete workflows explicitly labelled as partial research.

Evidence coverage is limited and currently concentrated on the NVDA example. Hashes establish file identity, not financial truth.

5. Dashboard access to existing engine capabilities

The interface also exposes manually supplied holdings and cash, historical evaluation, shared research memory, checkpoint recovery and ChatGPT session management.

These underlying capabilities already exist in TradingAgents. My work adds dashboard controls, task isolation, approval steps and clearer presentation around them.

A saved-report comparison view supports side-by-side analysis and transparent single-metric sorting. It does not discover sector members or claim to identify the best investments.

6. Manual storage management

Storage controls preview items older than 30 days by default, with other ages or a custom cutoff available.

Archiving is reversible and keeps whole artifacts together. Permanent deletion applies only to archived items and requires a separate confirmation. Active/unfinished work, linked inputs, credentials and shared memory are protected.

Archiving on the same drive does not free disk space.

Testing and limitations

The latest recorded local regression suite passed 188 tests, including cleanup safety and HTTP tests using disposable fixtures. Earlier component checks exercised installed graph routing and checkpoint behaviour with deterministic substitutes.

Fresh provider-backed workflows were not exercised during development. Local test results do not establish live provider compatibility, research accuracy or investment performance.

There is no brokerage connection, order execution or paper trading in this dashboard. New research can still consume provider usage; the offline reporting steps do not make the entire system free to run.

The code has not yet been packaged for public installation.

Feedback requested

Which would be most useful as a first focused contribution?

• Financial arithmetic regression tests and clearer evidence labels.
• Stage selection, pause/review and saved-input dependency handling.
• A documented local dashboard/launcher example.

I would also welcome advice on how best to align these additions with the current upstream architecture before preparing anything for submission.