ai-research-desk v1.0.2

Fixes from three independent code reviews and a live Claude vs ChatGPT comparison. Not financial advice.

Every run is kept

Each run gets its own folder, so running the same stock twice in a day no longer replaces the earlier run. Earlier runs open from "Other runs" on the dashboard, PDFs included.

Reports are reused only when nothing changed

A team report from earlier the same day is reused only if it came from the same plan, price session, models, fund settings, data sources and code. Otherwise the team runs again.

Price levels are computed, not written by the model

For each memo level the model picks a computed reference (a moving average, a recent high or low, a fair value, a Street target, optionally offset by ATR) and code works out the price. A level nothing computed supports is withheld, and so are levels in contradictory order (for example a stop inside the entry zone).

Cancel stops everything

Cancelling a run stops every team and every model call it started, on Windows and on Mac or Linux.

Fewer lost answers

• ai-hedge-fund personas answer against a JSON schema on both plans, and a reply with a clear call is no longer thrown away as unreadable. On one stock this had turned an Underweight into a Hold.
• When one analyst sits in two strategies, a call with a view wins over an abstention.
• Edge Desk keeps a section minus the sentences with an unsourced figure, instead of dropping the whole section.

Tested on a fresh install

A clean clone on Windows: setup in about 2.5 minutes, a full three-team run with a debate on the Claude plan without Alpaca keys (Yahoo prices), and a run on the ChatGPT plan with Alpaca.

Install

Download the zip below or clone with git clone --recurse-submodules. On Windows run setup.bat, then launch.bat; elsewhere python3.12 install.py, then scripts/run-dashboard.sh. Requirements: Python 3.12, git, and the claude or codex CLI for your plan. A new install starts empty: no runs and no demo data.

Upgrading

Your runs (memos/) and your settings (~/.hedge-desk/.env) are not touched. Runs saved before this version still open; their price levels show as "unchecked".

If you cloned with git: git pull, git submodule update, then run the setup again.

If you downloaded the ZIP: unzip this one into a new folder, copy your memos/ folder across, and run the setup once.

ai-research-desk v1.0.1

A fix for ZIP installs and a security update. Not financial advice.

The ZIP download now installs TradingAgents

GitHub leaves submodules out of a ZIP, so the v1.0.0 ZIP had an empty engines/TradingAgents folder and the setup could not fetch it. The setup now clones TradingAgents from the original project and checks out the exact commit this release is tested on. A git clone still uses the submodule.

Security updates

All open Dependabot alerts are cleared. ai-hedge-fund moves to the langchain 1.x line (langchain-core, langchain-anthropic and langchain-openai fixes), with newer python-dotenv, urllib3, pytest and black. The desk now needs anthropic 1.11 or later. Every test passes and nothing about how a run works has changed.

Install

Download the zip below or clone with git clone --recurse-submodules. On Windows run setup.bat, then launch.bat; elsewhere python3.12 install.py, then scripts/run-dashboard.sh. Requirements: Python 3.12, git, and the claude or codex CLI for your plan. A new install starts empty: no runs and no demo data.

Upgrading

Your runs (memos/) and your settings (~/.hedge-desk/.env) are not touched.

If you cloned with git: git pull, git submodule update, then run the setup again so the new library versions install.

If you downloaded the ZIP: unzip this one into a new folder, copy your memos/ folder across, and run the setup once.

ai-research-desk v1.0.0

The first public release: three AI research teams argue a stock into a swing and a long-term call, on your own Claude or ChatGPT plan.

Not financial advice. Not a signal service. The ratings have not been shown to beat the market. Every screenshot is synthetic demo data.

Three research teams and a desk manager

Type a ticker and three independent teams study the same company: TradingAgents (analysts, a bull and bear debate, a trader and risk managers), ai-hedge-fund (investor personas over SEC filings with a research manager) and Edge Desk (a point-in-time evidence package, a deterministic business case and a written analysis). A desk manager compares their calls for two holding periods, swing (2 days to 8 weeks) and long term (1 year or more), sends disagreements to a debate where a manager defends or concedes, checks which claims survived, and writes a short memo with a rating, the action for your position and price levels.

It runs on the AI plan you already pay for: your Claude plan through the claude CLI, or your ChatGPT plan through the codex CLI, picked per run in the dashboard. There are no per-call API charges, and every model call runs with no tools, no web search and no file access.

What is in this release

• A local dashboard (localhost only): start a run, watch the teams work live, reopen any past run, a weekly review of the latest call per symbol, and a two-page PDF summary.
• Data checks done in code: one settled close for every team, a stale-filing rule that withholds the long-term rating for the whole desk, and missing figures kept unknown rather than counted as zero.
• Market data from Alpaca (SIP, then IEX) when you have keys, Yahoo when you do not, and SEC EDGAR for filings.
• Tests for every part, run in CI on each pull request.

Install

Download the zip below or clone with git clone --recurse-submodules. On Windows run setup.bat, then launch.bat; elsewhere python3.12 install.py, then scripts/run-dashboard.sh. Requirements: Python 3.12, git, and the claude or codex CLI for your plan. Set SEC_USER_AGENT in ~/.hedge-desk/.env. A new install starts empty: no runs and no demo data.

Upgrading

Your runs (memos/) and your settings (~/.hedge-desk/.env) are not part of a release, so updating never touches them.

If you cloned with git: git pull and git submodule update, then run the setup again if dependencies changed.

If you downloaded the ZIP: unzip this one into a new folder, copy your memos/ folder across, and run the setup once.