Skip to content

0072 — A wake-up carries the market around it, and a take near its stop wakes the agent

Status: Accepted; amended by ADR-0073 (a symbol that is not trading answers an error, not a context) Edited 2026-11-15 to remove deployment details. Date: 2026-11-14 Amends: ADR-0061 (a wake-up’s body gains context; a third kind of event, setup.near_stop), ADR-0059 (a take’s record and log say when it went near its stop)

An agent these wake-ups are for spends four or five tool calls at the start of each woken session reading the same context: the price, the trend on a higher interval or two, funding, the long/short ratio and the Fear & Greed index. Putting them into the wake-up saves those calls.

An agent may also miss that a take filled. That wake-up exists: a status the server gives a take wakes the agent (ADR-0061), though a delivery during an outage of the receiver’s host can fail all its attempts. What does not exist is a warning before the stop: a take could go to 0.9 R against it and the agent would hear of it only at the stop. The only way round was an alert of its own, set by hand, on every take.

  • The wake-up contract stays backward compatible: fields are added, none renamed or moved. A receiver that ignores what it does not know keeps working.
  • Short and cheap. No new data source, and no slow external call in a wake-up’s way.
  • What the context says is what chartgen’s tools say: the same reads and indicators.
  • A warning comes once. A position that hovers near its stop must not wake the agent every minute.
  • The event log stays the one record of what could wake the agent (ADR-0061).

For the context:

  1. In the wake-up’s body, made when it is sent (chosen), held per symbol for a minute.
  2. Stored with the event when it is logged. The event would then hold what the market was when the event happened, but a body made an hour later would say the old price, and an event would no longer be a few numbers and words.
  3. Left to the agent, with a tool that answers it in one call. Saves calls, not a round trip. Offered as well (get_market_context), for wake-ups that do not come from chartgen.

For the warning:

  1. An event of its own, setup.near_stop, from a log entry the scorer writes (chosen).
  2. A field on setup.status. The status does not change at that bar, and a reader that keys on change.to would misread it.
  3. An alert the server makes on every take. A second object for the agent to keep in step, deleted or not, and one that fires on a price where the setup has its own measure, R.

An object, context, for one symbol:

FieldWhat it isSource
symbol, atthe symbol, and when this was read (ms)
pricethe last price: the close of the forming 4h barfetch_ohlcv, 150 bars of 4h
trend{"4h": …, "1d": …}, each up, down or mixedthe same bars, and 150 of 1d
funding_pctthe latest funding rate, in percentthe funding_rate indicator, on 150 bars of 1h
long_short{long, ratio}: the share of accounts that are long, and long/shortthe long_short indicator (Binance’s global account ratio, 4h period)
fear_greed{value, label}, today’s index and its bandthe fear_greed indicator (alternative.me)
lineall of it in one line

Trend. Of the bars as read, the forming one included, so its close is the current price: up when the price is above the 20-bar EMA and that EMA is above the 50-bar EMA, down when both are the other way round, mixed otherwise. The EMA is chartgen’s (Pine ta.ema, SMA-seeded); 150 bars let the 50-bar EMA settle. It is a description of the chart, not a signal.

Bands of the index: below 25 extreme fear, below 45 fear, below 55 neutral, below 75 greed, else extreme greed, the bands the Fear & Greed panel colours by.

What is left out. Funding, long/short and the index are for Binance symbols only; a stock has its price and trends. Any part that could not be read is absent, never guessed. A macro blackout is not in it: chartgen has no macro calendar. That stays with the agent’s own tools.

Funding is read on 1h bars, not the 4h ones: the indicator lines each rate up with the nearest bar open, and a symbol funded every hour has a newer rate than the open of the forming 4h bar.

Cost. A context is held per symbol for one minute, the index for ten. A cold read is three kline requests plus the funding and long/short series, which the indicators already cache and pace through the Binance gate. Callers that come while a symbol is being read wait for that read rather than start another, also after a wake-up stopped waiting for it.

In a wake-up. The body is made when it is sent (ADR-0061), and context is made with it. It comes after change or alert and note, before setup, so the thesis stays last. A wake-up waits at most 5 seconds for it; a read that takes longer goes on in the background and is held for the next attempt, and this attempt goes without context. It is absent too when the symbol’s bars cannot be read. The symbol is the fired alert’s, else the setup’s. The event in the log does not hold it: an event says what happened, and the context, like the setup, is shown as it is when sent.

get_market_context answers the same object for 1 to 10 symbols, in the order asked, with {symbol, error} for one that cannot be read. It is an ordinary MCP tool, on the same authenticated path as every other, so a script that already holds a token can call it. A replay server refuses it: the fixtures hold no funding or index.

A filled take is near its stop at the first 1m bar that takes it 0.8 R or more against the position and leaves it open. R and the excursion are the ones mae_r already has: the distance from the fill price to the invalidation, before costs, read from the bar’s low (long) or high (short) as the scorer reads them. Concretely, the first bar whose excursion sets a new mae_r at or below −0.8 while the setup stays triggered.

  • Once per setup. The record keeps near_stop_at, that bar’s open time, and is never told again, however often the price comes back and goes out.
  • A bar that stops it out says so with its status, not also with a warning. The bar that fills it can be the one, and then logs the fill and the warning together.
  • Takes only. A skip is not warned about, as no change of a skip wakes anyone.
  • The scorer writes it as an entry of the setup’s log, kind near_stop, by the server, with mae_r from/to and near_stop_at. That entry becomes an event of the owner’s log, setup.near_stop, with what: {mae_r, gap} and the time of the bar, in the same transaction, as a status is.
  • It wakes the agent if the take is still triggered when it is written. After the server was away, several bars are written at once; if one of them stopped the take, the status says more than the warning before it, which is then logged and not sent.
  • Its wake-up has "type": "setup.near_stop" and change: {mae_r, gap}, then context, then the setup, which now shows near_stop_at.

0.8 R is chartgen’s threshold: near enough that it is rarely noise, with room left to act before the stop.

  • Good: a woken agent starts with the context it otherwise reads first, in the body it already gets. A receiver that renders a template can show it with one placeholder, {context.line}.
  • Good: a take that goes wrong is seen before the stop, without an alert to keep in step with it.
  • Good: nothing is renamed or moved; a receiver that ignores new fields and kinds is unaffected.
  • Bad: a wake-up can be up to 5 seconds later than before when the context is not held. Wake-ups are minutes apart; their schedule and slots are unchanged.
  • Bad: the context of a wake-up sent on a retry can differ from the first attempt’s, as its setup can. The setup in a body is read after the context, so a bar that closed it meanwhile shows.
  • Neutral: setup.near_stop events count toward the owner’s 1000 kept events and the wake-up limits like any other.
  • Unit tests: the trend’s three cases and its refusal with too few bars, the index bands, the line, the tool’s argument checks; the near-stop rule (once, not on the stop bar, on the fill bar, not on a skip, stored and read back).
  • A service test: a take filled and taken to −0.83 R logs setup.near_stop once at the bar’s time and wakes the agent with context between change and setup; the same bars with a stop in one write log the warning and send only the statuses.
  • On a server with the sample source, get_market_context answers every part but the index.