Autonomous Trading

When to Use

• The prompt comes from the autonomous runner ("autonomous cycle").
• Any headless run with no user to answer questions.

How It Works

1. One cycle per closed bar, then exit. The runner starts a cycle right
   after each bar closes (the configured timeframe: 1, 3, ... minutes),
   through the whole session (18:00 ET until 16:00 ET the next day, the regular
   09:30-16:00 ET hours included), and hands you the closed bars in a
   file. Run exactly one trade-session (or premarket / end-of-day when the
   prompt says so). Don't wait, sleep, or poll inside a run; a bar that closes
   while you are still running is skipped, so finish within one bar.

2. No questions. Nobody is there to answer. Where the interactive workflow
   would ask, take the conservative branch: stand aside, keep the stop, flatten
   at end of day.

3. Positions first. If a position is open, only manage it this cycle.
   Before your cycle, the runner has already:

   • cancelled leftover orders on a flat contract;
   • for strategies whose exit trails, applied the trailing stop: the initial
     stop until trailActivateR, then trailGivebackR behind the best price
     (both in the scan's exit);
   • for strategies with a time stop (exit.max_bars), closed a trade that
     has been open that long.
     Don't loosen a stop it set. To see why a setup did or didn't fire, read
     that bar's record in <FTH_HOME>/logs/scans-<day>.jsonl (the runner's own
     log is <FTH_HOME>/logs/autotrader-<day>.log; your earlier cycles, with
     their prompts, tool calls, and skills loaded, are in
     <FTH_HOME>/logs/cycles-<day>.jsonl and logs/cycles/<day>/; the gate's
     decisions in logs/gate-log.jsonl). The prompt also carries the last 10
     closed bars and your last 10 cycle results: read them, and don't reverse a
     recent decision without saying what changed. Those results are your past
     conclusions, not facts: where one contradicts this prompt's account or
     attempt line (e.g. it says no attempt is running and the prompt lists an
     active one), the prompt's facts win, and you say the earlier result was
     wrong. It carries today's premarket
     plan, the news blackouts (in force, or the next one), the day so far (the
     "day:" line; outside RTH the "overnight" line), each fired strategy's track
     record (a paper one is marked: plan it only), and each open trade's state
     (setup, initial risk, stop, target, R now, best and worst): the
     trade-session skill says how to weigh them. Times are New York time. A
     "Context unavailable" line names what couldn't be read this cycle: that is
     unknown, not "none" (read it yourself, or trade smaller). The cycle's log
     (logs/cycles/<day>/) holds the prompt and what it was built from.

4. Guardrails are final. A block from the order gate, the MCP gateway, or
   the server ends the entry attempt. Fix what the block names when it is
   yours to fix (for example, review a closed trade, cancel a leftover order),
   otherwise end the cycle. Never edit harness files, settings, environment,
   strategy status, blackouts (except adding them in premarket), or the journal
   file to get past a limit.

5. Kill switch. When <FTH_HOME>/STOP exists the runner starts only
   manage cycles and the gate refuses every entry; you don't need to check
   the file (you can't read it). If an entry is refused with [kill-switch],
   manage or flatten open positions only, and end.

6. Leave a trail. Every cycle ends with a journal note (or plan, review)
   so the next cycle and the human can see what happened. Your last line
   must be exactly:
   CYCLE RESULT: <no-trade | planned | executed | managed | flattened | blocked | error> - <reason>
   (one outcome word, then - and the reason, as in the examples below).
   blocked only when the order gate, the gateway or the broker refused an
   order this cycle; nothing fired, or you stood aside, is no-trade.
   The runner reads only that line, from your final message: a summary, a
   table, or "Decision: NO TRADE" in its place is logged as "none reported",
   even when the cycle did its work. Write it last, after every tool call,
   as plain text in your final message: never as a tool call or a tool's
   arguments.

7. Budget. Keep the cycle short: reuse bar files fetched this cycle, at most
   one plan, at most one entry. The bar briefing (bar_briefing) is the
   runner's file for this bar: read it once per role and reuse it (it doesn't
   change during the cycle), and take its numbers instead of recomputing them. Read once per cycle. For each symbol, each
   role (you, each analyst, each call of a role; a role you play yourself
   counts as that role) reads each of these at most once and reuses it:
   strategies_scan (one with name is a separate read), mtf_read (one
   call, with daily when you have the daily file), get_account_snapshot,
   get_quote, get_bars. A full cycle still runs its analysts in parallel,
   except those whose report the prompt gives under "Cached reports"
   (trade-session step 2): those reports are this cycle's; don't launch the
   role again.
   Don't repeat what the prompt or a role gave you (the trade-executor's
   report is the after-order read of trade-session step 6), except:

   • The prompt lacks remainingBeforeLimit (and, without a prop attempt,
     today's P&L): a cycle that may enter reads get_account_snapshot once
     (trade-session step 1). It gives only a count of working orders, the
     trend rule, and no scan details: read list_open_orders once; take the
     alignment and the scan's candidates from the bar briefing (m15
alignment, m3 candidates), and run mtf_read or strategies_scan
     only when it says there is no current briefing. Anything under "Context unavailable", read
     once yourself.
   • Read the account and open orders once after each place_order,
     modify_order, cancel_order, close_position, or
     partial_close_position, including one that failed or timed out (its
     order may be live), and before an order call unless you read them after
     your last one (the Execution rule).
   • The scan again only after you fetched new bars. A trade review's
     get_bars over a trade's hold is one read per trade reviewed.
   • A failed read may be retried once (broker-mcp skill, Failures and
     limits). Never resend an order call blindly: after a failed or timed-out
     one, read list_open_orders and get_account_snapshot, and send it
     again only if they show it didn't go through and it is still needed (a
     close or a [protect] stop always is). A gate refusal ends the entry
     attempt unless you fixed what it named (rule 4).

   A tool that fails twice ends the cycle as error (after confirming any
   open position has its stop). Never wait within a cycle: bars and fills
   may not move until the next one. An order
   that hasn't filled at its one check is journalled as working and the cycle
   ends (executed or managed, saying it is working); the next cycle checks
   the fill (broker-mcp skill, Order mechanics). An unfilled entry without a
   stop bracket is cancelled, not left working (it would fill with no stop),
   and the cancel confirmed: cancelled, journal a result:nofill review and
   end as no-trade (entry cancelled: no stop bracket); filled before the
   cancel and protected, executed; still working after a second cancel,
   error (journal it as an entry without a stop bracket), and the next
   cycle cancels it first. Any quantity that filled is protected first.

8. Harness tools, not scripts. Bars, scans, the snapshot, the
   multi-timeframe read, blackouts, the prop status, lessons, and reconcile
   are tools on the harness MCP server (the harness-tools skill): call
   them; don't run the harness's scripts from a shell. Starting, stopping, or
   recording a prop attempt, backtests, and replays (replay.js) are the
   user's.

Examples

CYCLE RESULT: no-trade - orb fired but relative volume 0.8x (skip rule)
CYCLE RESULT: managed - MNQ long from 21503.25 (1R 5.25), peak 21515.75 (+2.38R); runner trailed the stop to 21513.00 (+1.86R)
CYCLE RESULT: blocked - [loss-streak] cooldown 18 min