Skip to content

Releases

The notes for each tagged release, newest first. Each release on GitHub uses the same text. A release gives a zip with the app wheel, the TradingAgents wheel, constraints.txt and .env.example, and a Docker image at ghcr.io. For how to use them, see Run it yourself.

For each change by date, see the changelog and the journey.

v0.5.1 — 2026-10-01

  • A pass that a stop fill starts now asks what the freed cash does next. The wake reason says the sale's money is cash again, and asks the agent to put it into a position, pay for research on a replacement, or keep it as cash by choice. On one deployment cash went from 1% to 81% of the book as three stops fired and no pass replaced them. In probes, research on a replacement went from 15 of 21 samples to 21 of 21.
  • Research is now the agent's main tool in the prompt, and its price heads the candidate list. "Paying for research" says what research is for and that a thousand analyses cost $50. The research rule no longer says that ignored research "is how the money leaves this account". The candidates result opens with the price and says research is how a listed name becomes a trade. In probes, research went from 7 of 21 samples to 16 of 21. The $0.05 charge did not change.
  • A market buy may carry its own stop and target. A level the agent leaves out is still the signal's, or a 2×ATR stop when the signal has none. Python refuses, and never moves, a stop at or above the price, a target at or below it, and any level on a limit buy.
  • The signals legend says Entry, Stop and Target are reference points, not limits. One deployment waited three days in cash for INTC to come back to its Entry while the price ran 10% past it.
  • The rule "Doing nothing is a valid answer, and often the right one" and the "Your recent wakeups" section are gone. Both were added to stop wakeups at the minimum interval. The agent quoted them as its reason to hold cash, and on most days it asked for the next morning.
  • The Decisions page on the live dashboard shows a pass while it runs. The card shows when the pass started, what the agent was told on waking, the turn, what the pass is doing now and the research it waits on. It reads the new read-only route GET /api/agent/pass. The public site does not show it.
  • Each prompt change has a probe record in .claude/rules/agent-probes.md, with 7 to 21 samples an arm on a copy of the book and the clock set to an open market.

v0.5.0 — 2026-09-30

  • Alpaca paper is a second broker, selected with BROKER=alpaca. Alpaca paper keys need only an email address, where Webull OpenAPI needs a funded account and an access application. Set ALPACA_API_KEY, ALPACA_API_SECRET and ALPACA_ACCOUNT_NUMBER, and ALPACA_ACCOUNT_CLASS (cash by default). webull stays the default, so an existing deployment changes nothing. The four guards apply in Alpaca's terms: the paper host only, a PA account number, the class from the account's margin multiplier with shorting off, and no fallback account. Fills settle on the 15-minute poll, because the Webull trade stream does not start. Fills and an OCO on a real position are not proved yet, so run backend/scripts/alpaca_paper_check.py during a session first.
  • An Alpaca deployment gets a candidate menu and 5-minute charts without a Webull client. The candidate screen uses Alpaca's most-actives and movers screens, priced on the delayed_sip feed, with delisted symbols and ETFs removed. Before, the menu was empty and the agent had no way to order research. The chart's 1-minute bars come from Alpaca's sip feed and are about 16 minutes behind. To fill the chart back to the experiment's start, run python -m backend.scripts.backfill_intraday_bars once in the container.
  • The news and sentiment analysts read more sources. get_news now adds Google News, Finnhub company news (with FINNHUB_API_KEY) and the company's SEC 8-K filings to Yahoo's headlines, which alone gave thin news. With LAYA_URL set, each item and post starts with a relevance grade and a tone. Nothing is dropped unless LAYA_DROP_BELOW is set, and it stays unset until a person has read the grades.
  • The agent can ask one question about an analysis with the new ask_analyst fetch. The answer comes only from that analysis's stored reports, about 20,000 characters the agent never saw before, and says so when the reports do not hold the answer. It costs one fetch and charges nothing. Tool channel only. Not probed in an open session yet.
  • A read now ends with every other analysis of the ticker, up to eight, newest first: the date, the verdict, the price, the app's levels, the chance and the sentiment score. Before, the agent could see that the newest analysis said Hold but not that the five before it said Buy.
  • The final pass runs at 3:45 PM ET, or 12:45 PM on a half-day. At 3:55 the agent read "closes in 4m" and decided it had no time to order research, read it and trade.
  • A pass that the model never answers is now recorded. On 2026-09-28 Google answered eight passes in a row with 503, and the record showed a quiet open. The row has skipped set and the new column unanswered true, and the migration runs at startup. Such a row does not count as the previous pass, so it does not move the alert window or drop the agent's note. A research that a shutdown stops is now reported as that, labelled analysis, not as a broker refusal.
  • The ticker page shows the latest analysis beside "What happened": when it ran, the call, the entry, the horizon, why it ran, the model and the rationale, with a link to the full analysis. Each timeline entry now shows its time as well as its date.
  • The Decisions page shows the agent's memory notes. A new panel, "What it remembers now", lists each note with when and where it was written, and each turn shows the memory orders it sent. GET /api/agent/memory serves the list, and the static site reads agent_memory.json.
  • The compose example uses the published image, ghcr.io/nandyalu/ten-acre:latest. A Docker deployment now needs compose.example.yaml and .env, and no clone or build. The image is x86-64 only, and "Build the image yourself" in docs/deploying.md keeps the steps for ARM and for your own changes.
  • Tooling: backend/scripts/fetch_audit.py counts which fetch results the agent's reasoning uses, and backend/scripts/prompt_cost.py shows where a pass spends its prompt tokens. probe_prompt.py now authenticates to a hosted OpenAI-compatible provider. A test now imports each of four modules first in a fresh process, after the backfill script failed on an import cycle between bars and watchdog.

v0.4.0 — 2026-09-25

  • A wake that comes during a pass now gets a pass after it. Before, a watchdog alert, an earnings wake or a fill that came while a pass ran was dropped, and the agent heard of it only at its next own wakeup. Now it is kept, and it is dropped only when the running pass built its last prompt after the event. A pass that comes due during the evening review now waits for the review, where before it was skipped.
  • A stop that the trade stream catches now wakes the agent and reaches its prompt. The stream settled the fill and posted it to Discord only, so the agent got no stop_fill alert and no wake. It now goes through scheduler.announce_fills, the same path as the 15-minute poll. A limit buy that filled was missed the same way and is fixed too.
  • The account line gives the equity first. It started with "Your account is $10,000.00 in total", and pass 88 sized its risk on $10,000 when the book was worth $10,942.50. The stop-fill alert no longer says "No pass ordered this sale", which was false: a pass placed the stop.
  • The scheduler's jobs run on quiv's threads, and quiv's job history is now true. Every job used to hand its work to the main loop with run_on_main and return, so each one showed a one-millisecond success. With quiv 1.2.0 only notify and the analyses go to the main loop, through call_on_main. The alarm, the five-minute backstop and the final pass are now one task, agent_pass, so the jobs page lists 7 tasks.
  • Set stop_grace_period: 20s on the ten-acre service. At shutdown the app gives the scheduler 10 seconds to stop its jobs and 3 more for work queued on its loop, and Docker's default of 10 seconds would kill it first. compose.example.yaml has the line; a deployed compose file needs it added by hand. A pass that is running stops between turns, never between an order and its record.
  • The Decisions page can show the system prompt, the rules every pass gets, behind a "Show the system prompt" button. GET /api/agent/system-prompt serves it, and the static site reads agent_system-prompt.json.
  • The site's tables, the Journal and the Decisions transcript are redesigned. The trade log, the positions and the analysis list filter and page in the browser. The Journal shows each day's parts on their own, a closed day carries a one-line summary, and one "Show the transcript" button opens the prompt and the model's answer side by side.
  • A test runs the agent task against a real quiv and checks that the time a pass chooses is kept, and quiv_findings.md records how the app uses quiv.

v0.3.0 — 2026-09-24

  • The agent reviews its own day each evening. After the close and the grading, it is shown every pass since its last review, what its trades did, and the analyses it bought with what the price did since. It answers two forced calls in one conversation: report, notes for whoever maintains it, each naming the pass where a tool or a fact was missing; then revise, a new note for the next pass and changes to its memory notes. It can place no order, commission nothing, and cannot move its alarm. The review runs on the tool channel only, so a deployment on the JSON channel skips it with a log line.
  • The next pass sees the review's note in place of the last pass's own, labelled as the review's, and the fixed rules now say the review exists and may rewrite the note and the memory. A memory note now carries the day it was written and whether a pass or a review wrote it. A note from before this release reads as undated.
  • A new table, agentreflection, keeps each review: both prompts, both answers, the notes, the memory changes and what was applied. The migration runs at startup like every other. GET /api/agent/reflections serves it, and the static site gets agent_reflections.json.
  • The Decisions page shows each review in its day, above the passes, and the Notes page lists the review's notes with the pass each one names. Discord gets a post only when a review sent a note or changed memory. Most days it says nothing, and that is the expected answer.
  • probe_prompt.py --turn reflection probes the review out of the app: the real prompt from the database as it stands, both calls, nothing applied and nothing recorded.
  • The notes for each release live in docs/releases.md, and a release skill gives the steps to cut one.

v0.2.3 — 2026-09-24

  • Every Gemini request now has a time limit. A Google request fails after 300 seconds (llm_gemini.REQUEST_TIMEOUT_SECONDS), and TRADINGAGENTS_LLM_TIMEOUT can change this value. On 2026-09-24, one request that never returned stopped the agent for ten hours.
  • The agent now sees a stop or target that fills during a pass. Each turn of a pass settles pending orders before it builds the book, so the agent no longer sees shares that it sold already. A fill found during a pass is sent as an alert, and a limit buy that fills during a pass wakes the agent again.
  • A refused adjust now says when the exit already filled. The agent gets "that exit already filled at $X, so there is nothing left to move" and not the raw broker error OPENAPI_ORDER_CANT_NOT_BE_REPLACE. For all other refusals, the agent still gets the broker's own words.
  • Tests no longer read the developer's database for pass records. The prompt tests now keep pass records in memory, so they pass when the local data/trading.db has not had the latest migration.

v0.2.2 — 2026-09-23

This release puts the agent's prompt in four headed groups, tells the agent why each pass started, which stop or target filled and when the next pass comes, moves a stop on every lot of a position, lets the agent decide with its own Gemini model, and publishes the Docker image to ghcr.io.

v0.2.1 — 2026-09-19

No notes were written. See the commits from v0.2.0.

v0.2.0 — 2026-09-19

No notes were written. See the commits from v0.1.0.

v0.1.0 — 2026-09-18

The first release.