What changed in 1.0.350 and 1.0.351

For whoever maintains a client of the OHLCX MCP server, of the in-app assistant, or of the Light relay. The first section lists what changes after 1.0.353; everything below it is in package version 1.0.351.

The MCP server went from 144 tools (1.0.349) to 145 (1.0.350) to 147 (1.0.351). The server version string is unchanged: 1.1.0.

After 1.0.353

Not released yet. Every MCP tool whose input or output changes for a client is listed here, one line each. No tool was added or removed: the server still has 147 tools.

What every list tool now has

• page (1 to 100000) on every tool whose API pages, and per_page where the API reads one. The answer carries meta: current_page, per_page and has_more, with last_page and total when the API gives them. Where the API sends a bare page of rows, has_more is true when the page came back full.
• limit on every tool whose API does not page. more says how many rows were left out.
• A size budget of 256 KB for every answer. Rows are dropped from the end (the oldest rows of a series), truncated is set and more counts them. An answer with no rows to drop is refused with "The result is too large to return. Ask for fewer rows or a narrower range."
• offset where a caller cannot lower a page length. While rows were left out the answer carries next_offset: ask again with it to read on. A single row too large for any answer is stepped over and row_too_large says so.
• Bounds in the schema. limit, page, per_page, offset, days, max_results and round_trips_limit declare minimum, maximum and default. A value that is not a whole number in plain digits within them (also true, "+2", 1.5) is refused with "<name> must be a whole number from <least> to <most>." and nothing is asked.
• Nothing was renamed or moved. meta, more, truncated and next_offset are added beside what a tool already answered.

A different shape

Tool	Change
get-account-growth	Was a bare list. Now {account_id, account, growth}.
get-account-pnl	Was a bare list. Now {account_id, account, pnl}.
get-sessions	Was a bare list. Now {sessions}. An empty list used to be answered as a failure and is now {sessions: []}.
get-sidebar-conversations	Was a bare list. Now {conversations}.
get-strategy-timeline	Answers the most recent whole days that fit one answer (about three), with returned (days, from, to) and more. days is 1 to 365 and now defaults to 3 (was 7, which no longer fitted). New logged_only leaves out the minutes with no log, so that a year of days usually fits. A day too busy to fit comes with its latest logged minutes and more_minutes.
get-account-pnl-symbols	symbols come largest move first, winners and losers alike (was best first). round_trips are bounded: new round_trips_limit (1 to 200, default 50) and more_round_trips.
get-strategy, list-strategies	The routing block is given as the routing tools give it: reason is a sentence (was a code), set_aside_reason and unknown keys are gone, a label is always masked, at most 50 accounts with more.
get-support-ticket	A conversation over 256 KB keeps its newest messages; more counts the older ones left out.

Fewer rows by default (limit 1 to 200, default 50; offset to read on)

search-knowledge-base, get-accounts-balances (accounts), get-account-pnl-symbols (symbols), list-watchlists (the user's own lists), screen-market-gaps, screen-asset-gaps, get-sessions, list-pending-join-requests, get-sidebar-conversations.

get-account-pnl-history: the 366 most recent days by default (limit 1 to 1000); narrow from and to for older days.

limit up to what the API itself gives

These got limit only to ask for fewer. The API stops where it always did, and the description now says so: list-accounts (100 accounts), get-account-balance (360 rows), get-account-growth and get-account-pnl (120 rows), get-strategy-activities (10), get-transaction-history (100 of each list), get-unread-messages (50).

New page, and offset within a page

Tool	Page length	New arguments
list-signals, list-conditions, get-strategy-conditions	100	page, offset
list-strategy-signals, list-strategy-trades, get-strategy-flags, list-trades	100	offset (page existed)
search-strategies	30	page, offset
list-backtests, list-backtest-jobs	25	offset (page existed)
list-markets	10	page, offset
get-ticker-analysis	10	offset (page existed)
get-ticker without a symbol	1000	page, offset. A page of 1000 is usually over the size budget: follow next_offset.
browse-users, discover-groups	20, 10	page, offset
list-support-tickets	20	page, offset

New or changed page and per_page

Tool	Change
list-news, list-crypto-news, list-popular-news	per_page now takes effect (it was ignored and every page had 10 items) and is 1 to 50.
get-analysis	New page and per_page (1 to 50, default 10). The description now says what the tool returns: the Analysis feed.
list-sectors	New page. per_page is 1 to 100, default 10.
list-activities	New page and per_page (1 to 50, default 30).
list-invites	New page and per_page (1 to 50, default 10).
list-users	per_page is 1 to 100; meta beside total, per_page, current_page.
list-strategies	meta gains has_more.
list-account-cash-transfers, list-credit-holds, screen-technicals, get-older-messages	meta added or gains has_more.

Other arguments

Tool	Change
get-accounts-balances	days is refused outside 1 to 365 with the same sentence as the other bounds.
list-orders	max_results (1 to 500) is refused the same way.
get-ticker-bars	limit (1 to 1000) is refused the same way. When the size budget cuts bars, meta.next_cursor is null: ask again with a smaller limit.
get-sector	sector_id must be plain digits. An undocumented second name for it, id, is no longer read.
list-markets, get-ticker	market_id must be plain digits.
get-news	id must be plain digits.
deploy-strategy, duplicate-strategy	strategy_id must be a whole number of 1 or more.
get-account-pnl-history, get-account-pnl-symbols, list-account-cash-transfers	An account_id that is not one of the caller's accounts answers "No linked account of yours has that id.", like the other account tools.
get-order, list-orders	The caller's own account that is not linked for orders answers "That account of yours is not linked with the broker for orders. Link it again in OHLCX."
run-trading-agent	No longer marked read-only: its agent can change the user's settings and preferences and, for an admin, a user's credits and billing package. The description says so. It also lists what the agent now reads: linked accounts with their realized profit and loss and balance over time, watchlists and the Screener's screens.
run-support-agent	Description corrected: it reads the knowledge base and pages of ohlcx.com, as a guest.

Failures

• 63 tools no longer add the cause of a failure to their sentence. "Unable to fetch sector: ..." is now "Unable to fetch sector."
• Of those, 22 write tools say why the API refused the request (a 422), in the API's own validation lines after the tool's sentence, at most ten lines of 200 characters, no markup: update-password, update-profile, create-group, update-group, invite-to-group, join-group, send-invite, accept-invite, respond-join-request, send-message, save-settings, update-preferences, log-activity, create-knowledge-base-article, update-knowledge-base-article, submit-contact-form, report-issue, submit-support-request, adjust-user-billing, set-user-billing-package, change-user-role, block-unblock-user. The other 41 (the read tools, the two agent tools, the deletes) answer their sentence alone whatever happened.
• OHLCX Light. The tools that already passed the API's words on for a refusal on Pro now read a failure the same way on Light: the validation lines of a 422 after the tool's sentence, and "That was not found." for a 404 (both were the sentence alone). A 401 or a 403 on Light stays the tool's sentence alone, because there it is the server's own access that was refused, not the user's.

The in-app assistant

• list_brokerage_accounts: an account's label is five asterisks and then up to four digits of a number the broker stores masked, or the last three characters of any other number (was four asterisks and up to four characters). It is the label every other tool shows for the same account.
• A failed turn has a code. The stream's error event is now {"type":"error","code":...,"error":...,"retryable":...}. The codes are timeout, busy, too_long, step_limit, unavailable and failed, each with one fixed sentence under error, which is kept for clients that know no codes. The table is in in-app-assistant.md. A stream that ends after tool calls and no text is an error only when the step limit was reached.
• A request with stream: false that fails answers that same object as JSON, with status 504, 503, 413, 502, 402 or 502 by code. The old body, {"text": "AI provider unavailable (insufficient credits or quota)."}, is gone.
• New limits. The three agents take at most 12 steps in a turn (before: 1.5 times the agent's tool count, which is more than 12 for a signed-in user) and each request to the model may take 120 seconds (before: 60).
• Failover. The next provider is now tried when the first is rate limited or overloaded as well as out of credits, and only while nothing of the answer has been sent. A tool that started is never run on a second provider.
• Two new assistant tools (OHLCX Pro, with Ask OHLCX on), 47 tool names in all. get_backtest_run_trades reads single trades of one saved run: filtered, sorted, at most 50 rows, with the totals of every match; it is listed only where the app reaches the strategies data as the signed-in user. show_backtest_trades shows trades of the open run as a card the page draws from the trades it holds; it reads nothing and returns no trade.
• New page-context key trades_card (kind backtest_run, the boolean true only): the page saying it can draw the trades card. The server offers show_backtest_trades only then, and removes the key from what the model reads. The front end sends it only with the build switch VITE_RUN_ASSISTANT_TRADES=true. Switch that on only on a host whose package has the tool: an older package passes the key into the run's data with no tool behind it.
• One shared trade filter. Both tools take filters, a list of {"field", "value"} entries, beside sort, order and limit. A field is one of direction, option_type, result, exit_reason, entry_hour, weekday, from, to, min_pnl, max_pnl, contract; a value is always text; a field is given at most once; an entry with an empty value asks for nothing and is left out. A filter given as an argument of its own is refused, with a sentence that points to filters. A profit bound at the outer limit (min_pnl of -1000000000, max_pnl of 1000000000) lets every trade through and is left out of the view. What the card tool returns is unchanged: a view whose filter is a plain object. The browser applies the same rules to the card. They are pinned by tradeFilterCases.json, held byte for byte in the package and in the host app with the same hash.
• Two more assistant tools that show a card (OHLCX Pro, with Ask OHLCX on), 49 tool names in all. show_backtest_breakdown shows how the trades of the open run split by entry hour, weekday, exit kind, side, calls or puts, or month; show_backtest_stats shows one to six headline figures as tiles, out of 23 names. Both read nothing and return a checked view with no figure; the page draws the card from the run it holds. An answer may show trades, a breakdown and figures and carry one proposal, one of each.
• New page-context key cards (kind backtest_run): the page's list of the cards it can draw, "trades", "breakdown", "stats". The server offers each card's tool only for a name in the list and removes the key from what the model reads. A name it does not know is ignored on its own; a value that is not a list of text is ignored as a whole. trades_card: true still works alone and means "trades". The front end sends the list only with VITE_RUN_ASSISTANT_BREAKDOWN=true or VITE_RUN_ASSISTANT_STATS=true (both off by default, Pro), and lists "breakdown" only for a run whose every trade the page holds. Install the package release with the two tools first, then set the switches and build: an older package passes the unknown key on to the model, and no card comes.
• Two more shared fixtures. The grouping of a breakdown and the names of the stats card are written on both sides and pinned by tradeBreakdownCases.json (SHA-256 75ed9bfb2ed94f3a0e788e4e267bf3589bb440c898f8dc84539c5a5d8c205730) and statsCardCases.json (SHA-256 177766e4c70040f2386f62a03da0c784be9d9049981fc08ee639dd2061264d26), each held byte for byte in both repositories.
• One rounding on the backtest run page. The results header ("Net profit", "Win rate"), the Metrics tab's own rows ("Net profit", "Gross profit", "Gross loss", "Win rate", and the bar above them) and its breakdowns (by signal, exit, weekday and hour) now write money as floor(sum * 100 + 0.5) / 100 and a win rate as floor(wins * 1000 / trades + 0.5) / 10, in floating point as the server does, and the breakdowns mark the best and the worst group by the rounded result. The breakdown and stats cards use the same rule, so no two places show two numbers for the same trades. With profits in whole cents and trade counts that are not multiples of 80 nothing reads differently. Otherwise a figure can change by its last digit, in either direction: 23 winners of 80 trades reads 28.8% (was 28.7%); a sum of 8.575 reads $8.57 (was $8.58); a sum of 1400.5149999999999 reads $1,400.52 (was $1,400.51); a sum of -0.125 reads -$0.12 (was -$0.13); a loss under half a cent reads "+$0.00" (was "-$0.00" in red); groups less than a cent apart are no longer marked best and worst. A run saved without its longest losing streak shows a dash there (was 0). This is in the front end and does not depend on either switch.
• The prompts and the cards. The first line of "Response formatting" in the main prompt now says that what a show_ tool is offered for is shown by calling it, with no table and no block, and that trades of a backtest run are never written as a table. The backtests prompt has a new block, "Cards of a backtest run (Pro)". The rules for each card are also in the run's own section, which does not depend on the prompt files.
• The prompts of the three agents have a rule for turns that failed and say the assistant cannot create or change a strategy or run a backtest. A host reads its own published copies of the prompts, so it gets the rule when it republishes them.
• Eight new assistant tools that only read, 57 tool names in all: screen_market_gaps, screen_symbol_gaps and screen_technicals (the Screener page's three screens), list_watchlists and get_watchlist (symbols only, no prices), and get_account_pnl, get_account_pnl_by_symbol and get_account_growth (realized profit and loss, by symbol, and the balance over time of one linked account). The Trading and the Unified agents have them for a signed-in user, on OHLCX Light and on OHLCX Pro, with nothing switched on. Tool counts for a signed-in user: Trading 30 on Light and 41 on Pro (were 22 and 33); Unified 33 on Light, 44 on Pro and 47 on Pro with Ask OHLCX on (were 25, 36 and 39); Support is unchanged at 21. No tool saves a screener, changes a watchlist, starts a profit and loss sync, or reads single closed trades or cash transfers; the assistant has no saved screeners to read, because there are none. No MCP tool changed.
• One more assistant tool that shows a card (OHLCX Pro, with Ask OHLCX on), 58 tool names in all. show_backtest_equity shows the equity and drawdown chart of the open run, drawn by the page from the curve it holds. It takes no arguments and always shows the whole run; any argument but the run's ids is refused. It reads nothing and returns { "note", "type": "backtest_equity_view", "view": {} }: no figure and no point of the curve. Tool counts without an open run are unchanged.
• "equity" is a fourth name in the page-context list cards. The server offers the tool only for a run whose page lists it. The front end lists it only with VITE_RUN_ASSISTANT_EQUITY=true (off by default, Pro) and only for a run whose equity curve it holds; with the switch on it sends cards for this name alone too. Install the package release with the tool first, then set the switch and build: a package that reads cards but does not know the name ignores it and keeps the other cards.
• One more shared fixture. The rule of the equity card's view (exactly {}) is written on both sides and pinned by equityCardCases.json (SHA-256 278764a95026507bb2ab647b9b512a8fd19b294d7a97d4bd27899df1f6fbad10), held byte for byte in both repositories.
• The prompts and the equity card. The first line of "Response formatting" in the main prompt now names the equity and drawdown chart among what a show_ tool is offered for; its rule for labels and data blocks is for get_market_balance and get_sector_balance only, as it was, and is unchanged. The backtests prompt's block "Cards of a backtest run (Pro)" has a line for the card: never describe the curve or state a value or a date from it, and no labels or data block for it; and, where the tool is not offered, describe nothing of the curve from memory and name the "Equity and drawdown" view of the results chart. The same rule for the card is in the run's own section when the page lists the card, which does not depend on the prompt files.
• Tools by page. On an Ask OHLCX panel the Unified agent is now offered only the tools that fit that page, not its whole list (PageToolScope). For a signed-in Pro user with Ask OHLCX on, where the app reads the strategies data as the user: a backtest run with every card and both proposals has 17 tools (was 53); a run with nothing to draw or propose, and the panel under a sweep's results, 11 (was 47); the Strategies page 9 (was 48); a strategy's order accounts 7 (was 48). The instructions of such a page end with a short section, "Limits of this page", and their first line points at it: for a part of a question the page's tools do not cover, the assistant says in one short sentence that this page cannot do that part, never offers to look it up itself, and answers the rest. The two tools that change Settings and Preferences and the admin's billing tools are on none of these pages. The drawer is unchanged but for the new tool below.
• New server setting TRADING_APP_ASSISTANT_SCOPED_TOOLS (trading-app.assistant_scoped_tools), on by default. Off gives every page the whole list and the instructions it had before. A blank value reads as off.
• The OHLCX website reader on the panels. fetch_ohlcx_webpage is kept on every page that keeps the knowledge base, so pricing and other product questions are answered on a panel as in the drawer. On a strategy's order accounts the page's section has one more sentence: propose first, and do not search the knowledge base or fetch a web page in that answer, because reading one makes the proposal refuse.
• New assistant tool get_ticker_bars (OHLCX Pro; the Trading and the Unified agent, and the run and Strategies panels), 59 tool names in all. It reads a symbol's daily or weekly price bars: symbol, with period (last_5, last_20, last_60; last_5 when left out) and timeframe (1Day, 1Week), or day (YYYY-MM-DD) for one day's bar. A window returns at most 60 bars, newest first, with totals: the change measured from the close before the window, the high, the low and the volume. One day returns that bar and the change from the close before; a day with no bar returns "bar": null and latest_before. There are no intraday bars. Tool counts for a signed-in Pro user: Trading 42 (was 41); Unified 45, or 48 with Ask OHLCX on (were 44 and 47). Counts on OHLCX Light and for the Support agent are unchanged. The MCP tool get-ticker-bars is untouched.
• New assistant tool suggest_main_assistant and new card type main_assistant_handoff (OHLCX Pro, with Ask OHLCX on), 60 tool names in all. On a page with its own set of tools that lists "handoff" in the page-context list cards, the tool shows a button that opens the main assistant with the question the user just asked. It takes no arguments, reads nothing and returns { "note", "type": "main_assistant_handoff", "view": {} }; whatever a model sends is dropped. One per answer; it takes no proposal's place. Where it is not offered the page's limits tell the assistant to say in words that the question can be asked in the main AI Assistant. Never offered in the drawer.
• cards is accepted on every kind of page. A strategy_builder and a strategy_routing context may now carry cards; the only name read there is "handoff". It is read as strictly as a run's list (a JSON list of text only) and never shown to the model. A package from before this is not disturbed by the key: for those two kinds it builds what the model reads from the parts it checked, so the key is dropped; for a run it ignores a name it does not know.
• The hand-off card in the front end. With VITE_ASSISTANT_HANDOFF=true (off by default, Pro) the three panels, and the panel under a sweep's results, list "handoff" where the main assistant's drawer can be opened (Pro, signed in, Embedded AI on). The page draws a card under the answer, "This page can't do that. Ask the main AI Assistant.", with a button, "Ask in the AI Assistant". It opens the drawer and asks the user's own question of that turn there. If the drawer already holds a conversation it starts a new chat first, as its New chat button does; the earlier conversation stays in the history list. What the user was typing in the drawer's message box is kept; pictures chosen there are dropped with the conversation left. If the drawer is busy with an answer or the microphone, or the question is over 16,000 characters, the question is only put in its message box, and the card's button then opens the drawer and puts it there again without sending. No text written by the model is passed on. With the switch off the contexts are exactly what they were.
• One more shared fixture. The rule of the hand-off card's view (exactly {}) is written on both sides and pinned by handoffCardCases.json (SHA-256 61759574b5e2582fe4218e8bf4e7cdd57b4b0516b6c1defe1280a237002d5cf1), held byte for byte in both repositories.
• openMainAssistant({ question, send }) (front end, helpers/assistant/mainAssistant.js): one way to open the main AI Assistant from anywhere in the app, with a question. It rides on the event the header's button already opens the drawer with. With send: true the question is asked at once; otherwise it is put in the drawer's message box. It returns false, and does nothing, where there is no drawer to open. The drawer's own send path is unchanged for everything else, in both states of VITE_ASSISTANT_SHARED_THREAD.
• The knowledge base search matches a question by its words. Before, the text searched for had to appear in an article as one exact phrase, so a whole question found nothing. Now an article matches when it holds every word of it (common words such as "how" or "the" are left out, and at most 8 words are used), the best match first: a word in the title counts most, then the slug and the area, then the excerpt, then the content. When no article holds every word, those that hold any are given. This is one shared search for search_knowledge_base_articles, the MCP tool search-knowledge-base and the knowledge base API, so all three changed together. The arguments and the shape of the result are unchanged.
• Three tool descriptions changed. search_knowledge_base_articles says how q is matched and to leave area out unless asked. get_ticker points at get_ticker_bars for price history, when it is among the tools. get_trading_strategy no longer points at the execution log.
• The prompts of the Unified and the Trading agent have one new line, for get_ticker_bars. A host that published its own copies gets it when it publishes them again; the tool's description carries the same rules. This host's copies in resources/prompts were brought in line.
• Fixes from the live check (package 1.0.364). get_ticker_bars: a day given beside a period is the window that ends on that day (the model filled in today's date and then ran out of steps asking day by day), and the description says how long each period is in weekly bars. The hand-off: the rule in "Limits of this page" is one line that says to call the tool before writing, and Support/HandoffNet adds the tool's result to the stream when a finished answer names the main assistant without it, so the button is always drawn. The Unified and Trading prompts' get_ticker_bars line says the same about day.
• "Hide Account Balance" is honoured by the account tools. While the user has it on, the results of get_account_pnl, get_account_pnl_by_symbol, get_account_growth and list_brokerage_accounts hold no amount of money (profit and loss, fees, balances, buying power), only percentages, counts and dates, and the model is told so in a fixed note. If the preference cannot be read it counts as on. The MCP server's own account tools do not follow this preference.
• list_brokerage_accounts result changed. Every result now has balances_hidden (true or false), and note when true. When true, the five balance keys (liquidation_value, cash_balance, buying_power, day_trading_buying_power, balances_as_of) are absent from every row, not null. A client that draws the accounts list from this result must handle rows without figures.

• The prompts of the Trading and the Unified agent have a subsection for these tools and for balances_hidden, and the Unified prompt's "Response formatting" says they are not written as tables. A host that published its own copies gets them when it republishes.

• Fuller drafts on the Strategies page (front end; needs the package release that reads the three keys). Three build switches, each with VITE_STRATEGY_ASSISTANT and off by default: VITE_STRATEGY_ASSISTANT_SETTINGS sends settable (the settings a draft may set, built from the settings forms' fields, plus the three of an unfilled entry's cancelling), VITE_STRATEGY_ASSISTANT_MULTI sends max_drafts: 4, VITE_STRATEGY_ASSISTANT_BACKTEST sends backtest: true. With all three off the strategy_builder context and the card are unchanged. With them the draft card lists the settings in the forms' words, what is left at its default and what the draft does not build; after "Create this strategy" it is a summary of what was done, each setting compared with the strategy the API answered with ("could not be confirmed: open it and check" where it differs); an answer may carry four cards, created one by one; and "Backtest it" opens the strategy's Backtest tab with the draft's period filled in and, where the draft asked, the run started once. The create call is the four choices, the checked settings, the schedule's time zone and the nine switches last, all off. A symbol may be a pair written with one /.

• The assistant before a backtest run (front end; needs the package release that has propose_backtest_setup). VITE_BACKTEST_SETUP_ASSISTANT=true (with VITE_RUN_ASSISTANT, off by default) adds "Set up this backtest" under the backtest form while no run is shown. It sends a page of kind backtest_setup and draws the proposal as a card with "Use these settings" and "Run". Both read the proposal again against the form as it is and go through validateRunSettings; a proposal that no longer fits says so in the form's own sentence and does nothing. The conversation is carried under the run's result. A package that does not know the kind answers such a page with status 422: install the release first.
• Four more shared fixtures, held byte for byte in both repositories: draftSettingsCases.json (SHA-256 bc3ee96bc398ed62803e2a654c97dc45d8e09a7687b6ddefe62afeaf60f8342b), draftGaps.json (4a4a49d4a30cd7a73e6269f3c022abb49189ed9047a1b3ecea372f71dd962a18), periodCases.json (c956053f163252a848322b086d7879a955ea8b39ec950f3480fc69af25739e94) and backtestSetupCases.json (897dba5900a61876051857514f99b05b6cfcfff65fbc08efc8f77c55b9213971).
• useAssistantConversation: several. A proposal kind marked several: n may come up to n times in one answer; remember(card, all) is handed the list. many still means nothing on a proposal kind.

The relay (OHLCX Light)

• GET /api/support/tickets passes page on to the support desk when it is a whole number of 1 or more.
• Whether the relay routes exist is read from the configuration only, and the data source setting is read without regard to case.

At a glance

Area	1.0.350	1.0.351
New MCP tools	retain-strategy	list-strategy-accounts, set-strategy-accounts
MCP tools whose input or output changed	list-strategies; the strategy and condition write tools (refusals)	the account and order tools; list-credit-holds, list-activities, get-user-billing; retain-strategy
Assistant	Ask OHLCX; bounded results on every tool; three strategy tools	get_strategy_routing, propose_strategy_accounts
Relay (Light)	Strategy and condition writes keep a refusal's status and never send its cause	no change

New MCP tools

retain-strategy (1.0.350)

Takes a strategy out of evaluation. Arguments: strategy_id (integer, required), stop (boolean, optional).

• A plain retain leaves the strategy's switches as they were, so deploying it again resumes them.
• stop true also switches orders, trades and signals off.
• Since 1.0.351 it also refreshes list-strategies at once.

A deployed strategy refuses changes. The workflow is: retain-strategy, make the change, deploy-strategy.

list-strategy-accounts (1.0.351, OHLCX Pro only)

No arguments. Lists the user's linked accounts as a strategy's orders can be routed to them.

{
  "connected": true,
  "accounts": [
    { "id": 4, "label": "*****678", "broker": "schwab", "routable": true, "reason": null, "is_default": true },
    { "id": 5, "label": "*****321", "broker": "alpaca", "routable": false, "reason": "Accounts at this broker cannot receive strategy orders yet." }
  ]
}

At most 50 accounts; more says how many were left out. reason is a sentence, not a code.

set-strategy-accounts (1.0.351, OHLCX Pro only)

Chooses which linked accounts receive a strategy's orders.

Argument	Type	Notes
strategy_id	integer, required
mode	selected or all, required	all: every linked account that can receive orders when an order is sent, including accounts linked later.
account_ids	array of integers	Required with selected: ids from list-strategy-accounts, at most 50, each once. JSON integers only. An empty list removes every account. Not used with all.

Answers { "strategy_id": 7, "routing": { "mode", "connected", "can_route", "accounts": [ { "id", "label", "broker", "routable", "reason", "set_aside" } ] } }.

• The choice replaces the previous one.
• It never switches orders on or off.
• Confirm with the user first, naming accounts by their masked labels.
• set_aside: true on an account means an order to it was refused: it receives no orders until the accounts are set again. Setting the same choice again brings it back.

Refusals, each a fixed sentence:

Sentence	When
Retain the strategy before changing it.	The strategy is deployed.
Choose accounts from your linked accounts.	An id is not one of the user's linked accounts.
Alpaca accounts cannot receive strategy orders yet.	An account at a broker that cannot be routed to yet.
Switch orders off before removing every account that orders can go to.	It would leave a real strategy with orders on without an account.
Choose at most 50 accounts.	Too many ids.
Unable to set this strategy's accounts. You are not allowed to change this strategy.	The strategy is not the caller's.
Order routing accounts are managed in OHLCX Pro.	Called where routing is not available.

Both tools are listed only on OHLCX Pro for a signed-in user.

MCP tools whose input or output changed

Accounts: no broker key, no full number (1.0.351)

No tool returns a broker's key for an account (account_hash, accountHash, hashValue, broker_account_hash) or a full account number any more. An account is its id and a masked label such as *****678, the same label list-strategy-accounts shows.

Tool	Input	Output
list-accounts	unchanged	The key is removed at every depth. account_number is the masked label. Every other field is unchanged.
get-account-balance	account_id: see below	New shape. Was an object keyed by the broker's key. Now { "account_id": 4, "account": "*****678", "history": [ ... ] }.
get-accounts-balances	unchanged (days)	New shape. Was an object keyed by the broker's key. Now { "accounts": [ { "account_id": 4, "account": "*****678", "history": [ ... ] } ] }. A history that names no account is left out.
get-account-growth, get-account-pnl	account_id: see below	Same shape. A key is removed and a number masked if present.
get-account-pnl-history, get-account-pnl-symbols, list-account-cash-transfers	account_id: see below	account_hash is removed; account_display is the masked label.
get-order, list-orders	account_id: see below	No key. accountNumber in an order is a masked string at every depth (it was the full number).
list-credit-holds	unchanged	broker_account_hash is removed from every hold.
list-activities	unchanged	Inside each activity, a key is removed and an account number masked.
get-user-billing (admin)	unchanged	A key is removed and an account number masked.

account_id on the eight account and order tools (1.0.351)

get-account-balance, get-account-growth, get-account-pnl, get-account-pnl-history, get-account-pnl-symbols, list-account-cash-transfers, get-order, list-orders.

• The schema declares "type": ["string", "integer"].
• Send the account's id from list-accounts, as a number (4) or in digits ("4").
• A client that still holds a broker key may send it, but only its own user's key is accepted.
• The value may hold only letters, digits, _ and -, at most 128 characters.
• get-order used to need the broker's key. It now takes the id. This is the change most likely to break an existing client, because the key is no longer available from list-accounts.
• get-account-balance, get-account-growth, get-account-pnl, get-order and list-orders answer "No linked account of yours has that id." for an id or a key that is not one of the caller's accounts. The other three answer that sentence for a key that is not the caller's.
• list-orders without account_id still lists every account.

Failure sentences (1.0.351)

These tools no longer append the cause of a failure:

Tool	Sentence
list-accounts	Unable to fetch accounts.
get-account-balance	Unable to fetch the account balance.
get-account-growth	Unable to fetch the account growth.
get-account-pnl	Unable to fetch the account profit and loss.
list-activities	Unable to fetch activities.
get-user-billing	Unable to fetch user billing.

Most other tools still answer their own sentence followed by the cause.

list-strategies (1.0.350)

New optional arguments page and per_page (1 to 50).

• Without them: the plain list, up to 50 strategies, as before.
• With either: { "strategies": [ ... ], "meta": { "total", "per_page", "current_page", "last_page" } }.

This is the only MCP list tool that gained paging in these releases.

Strategy and condition writes: refusals (1.0.350)

Twelve tools say in their description that they are refused while the strategy is deployed, and answer one sentence when it is: "Retain the strategy before changing it."

update-strategy, update-strategy-settings, set-strategy-schedule, set-strategy-status (switching on), set-strategy-flag (switching signals, trades or orders on, or toggling), delete-strategy, clear-strategy-conditions-readings, clear-strategy-data (flags), create-condition, update-condition, delete-condition, delete-signal.

Never refused for that reason: switching anything off, the notification switches, deploying, duplicating, backtests, deleting or clearing trades. Deleting a signal, or clearing all signals, is refused only while the strategy is deployed, real and has orders on.

deploy-strategy, duplicate-strategy and create-strategy joined them in answering two more refusals as one sentence each:

Sentence	When
Choose a linked account before switching orders on.	Switching orders on for a real strategy with no account that can receive them.
Choose a linked account, or switch orders off, before deploying.	Deploying a real strategy with orders on and no account.

For any other refusal these tools answer their own sentence. When the API refused the request as invalid, not signed in or not allowed, what the API said is added, shortened and cleaned. A strategy or condition that does not exist reads "That was not found."

Per-user caches (1.0.350)

list-strategies, list-signals and list-conditions keep an answer for one minute per user. A write through any strategy, signal or condition tool clears it, so a list read after a write is fresh.

Server instructions and the settings resource (1.0.351)

The server's instructions now also say that choosing a strategy's order routing accounts decides which brokerage accounts receive its orders, never switches orders on, and must be confirmed with the user. The ohlcx://strategy-settings resource gained a section on order routing accounts and on the deployed lock.

The in-app assistant

1.0.350

• Ask OHLCX (OHLCX Pro, behind a switch): the assistant reads the open backtest run from the page; list_backtest_runs and get_backtest_run; the proposing tools propose_backtest_change, propose_backtest_sweep and propose_strategy_draft, each of which returns a card the user must press.
• Three strategy tools on the unified assistant: get_trading_strategy, get_trading_strategy_statistics, get_trading_strategy_performance.
• Every tool's result is bounded. Lists are short rows with a row limit and say returned, count, total, truncated and, where the total is not known, more_may_exist. list_trading_strategies gained a page argument (25 to a page).
• Less personal data. list_brokerage_accounts returns a masked label and a few fields, never a number or a key. The community tools return names, never emails. Failures are one fixed sentence.

1.0.351

• get_strategy_routing (OHLCX Pro): which linked accounts a strategy's orders go to, as masked labels. Read-only. No account id.
• propose_strategy_accounts (OHLCX Pro, in a strategy's settings): offers a change of the strategy's order accounts as a card with a "Set these accounts" button. It calls nothing and sets nothing. One proposal per answer. It refuses in an answer that read a web page, news, an article or messages.

The relay (OHLCX Light)

Changed in 1.0.350. A Light app relays strategy requests to the hosted OHLCX API.

Strategy and condition writes

Creating, updating and deleting a strategy or a condition, every settings section, the schedule, retain, retain-stop, deploy, duplicate, every switch (status, signals, trades, orders and their notifications), the Slack test notification, the four clear actions, and deleting a signal now answer like this:

The API answered	The relay answers	message and error
422	422, with the validation errors	The server did not accept these settings.
422 with one of three named refusals	422, with errors and a code	that refusal's sentence
401	401	You are signed out. Sign in and try again.
403	403	This account is not allowed to do that.
404	404	That was not found.
429	429	Too many requests. Wait a moment and try again.
a server error, or no answer	502	Unable to save right now.
anything else	500	Unable to save right now.

The three codes: strategy_deployed ("Retain the strategy before changing it."), no_routable_account ("Choose a linked account before switching orders on."), deploy_needs_account ("Choose a linked account, or switch orders off, before deploying.").

{
  "message": "Choose a linked account before switching orders on.",
  "error": "Choose a linked account before switching orders on.",
  "code": "no_routable_account",
  "errors": { "orders_on": ["Choose a linked account before switching orders on."] }
}

What a client should change:

• Read the status. A 4xx from this table means nothing was stored. Before, most of these answered 500 with the cause as error.
• A 404 is now a 404, not a 500.
• message and error carry the same sentence. Show either.
• To offer "choose an account", branch on code.

The two account routes

GET /api/strategy-accounts and POST /api/strategy/{strategy}/settings/update-accounts are not relayed. On Light both answer 403 with "Order routing accounts are managed in OHLCX Pro." under message and error.

The paged strategies list

GET /api/strategies?paged=1 returns the paged list, with optional page and per_page (1 to 50). Without paged the two are ignored and the answer is the plain list, as before.

Unchanged

Reads, backtests, the screener and the market, ticker and signal-action routes answer as before.