Write the prompt. Tune the inputs. Both in one place.

First agent in the pipeline. Reads the prompt's KEY OBSERVATIONS block, computes the indicators on the chart's TF, and writes the deterministic technical state the rest of the pipeline reasons over. Cheap value-tier model is fine — this agent is a transcriber, not a forecaster.

Argues FOR the long (or, in symmetric mode, FOR the short). Reads the prompt's BULL ANGLES block and constructs the strongest available case from the current technical state. Default is Claude Sonnet 4.6 — its long-form reasoning makes for better debate transcripts the Judge can grade.

Counter-arguer. Reads the prompt's BEAR ANGLES block and the Bull's case, then tries to invalidate. Default is GPT-4o — its terser, more confrontational style breaks the Bull's argument cleanly, which is what the Judge needs.

Reads the full Bull↔Bear transcript across all debate rounds plus the prompt's CONFIDENCE GUIDANCE block. Emits a single HIGH / MEDIUM / HOLD verdict with rationale. Default is Gemini 3 Pro for its long context — multi-round debates can run thousands of tokens and the Judge must read every line.

The only agent whose output reaches the broker. Takes the Judge's verdict, the prompt's ENTRY MECHANICS / STOPS / SIZING blocks, and the live broker spec (spread, stop level, lot step), then emits the final EXECUTE / HOLD directive with SL / TP / lot. Default is GPT-5 — this is the most expensive seat in the pipeline by design.

How much of the 5-agent pipeline runs each cycle. MINIMAL = Portfolio Manager only (1 LLM call). STANDARD = Technical Analyst + Research Manager + Portfolio Manager (3 calls). FULL = all five agents with Bull/Bear debate.

How many Bull↔Bear exchange rounds the Judge sees before voting. Valid range 1–5. Only used in FULL pipeline mode. The EA refuses to start if this is outside range.

Required if any agent in Group 1 / 1C is on the GPT-5 family (including GPT-5 mini) or the GPT-4o family (including GPT-4o mini). Get a key at platform.openai.com → API keys. Defaults: 1.1 + 1.3 + 1.5 + 1C.1 all use this key, so it's the single most-required slot.

Required if any agent is on the Claude Sonnet 4.6 / Opus 4.6 family. Default Bull (1.2) uses it. Get a key at console.anthropic.com → API Keys.

Required for any agent on Gemini 3 Pro or Gemini 3 Flash. Default Research Manager (1.4) uses it — bound to long context. Get a key at aistudio.google.com → Get API key. The free tier rate limit is enough for one chart at DebateRounds = 1.

Only needed if you point any agent at DeepSeek-Chat or DeepSeek-Reasoner. Strong value-tier option for the Technical Analyst slot. Get a key at platform.deepseek.com.

Only needed if you point any agent at the Grok family. Get a key at console.x.ai. Leave empty if no agent selects a Grok model.

Only used when an agent is set to MODEL_CUSTOM. Pair with the internal CustomModelName + CustomEndpointUrl constants (in source — recompile to change). Any OpenAI-compatible Chat Completions endpoint works: OpenRouter, a local vLLM server, an Azure-hosted Mistral, etc.

The model used when a primary agent's call fails transiently. Engaged only for the failing seat, only for that cycle — the next cycle re-tries the primary. Pick something with a different rate-limit pool from your primaries so a provider-wide 429 doesn't take the backup down with it.

The built-in preset the EA loads at startup. v1.0 ships with one verified, long-history-validated XAUUSD H4 strategy (Stoch MR-XAUUSD-H4) as the built-in default. The other seven community-curated archetypes (Donchian Turtle, Connors RSI2, Holy Grail Pullback, IBS Equity MR, Turtle Soup, NR7 Compression, ACD Gold NY) ship in the public Prompt Library and load via the eatrading_prompts.txt file in MQL5\Files\ — see input 2.6 below.

The master switch. When false the EA runs the full pipeline and writes decisions to the panel and the chart comment, but does not place a single trade. When true, every EXECUTE verdict from the Portfolio Manager is sent to the broker. Also toggle-able at runtime via the panel checkbox.

Hard cap on simultaneous open positions for this magic number. Valid range 1–50 — the EA refuses to attach if outside range. Counts toward Group 4 layers as well when DCA is enabled.

How often the EA fires the full pipeline. Set to 0 for manual mode — the pipeline runs only when you press Analyze on the panel. Defaults to 60 (1 hr) which is correct for H4: oversampling burns tokens without new bar information. Valid range 0 or 15–1440 — the EA refuses to attach if outside that band.

The integer tag stamped on every order from this chart. Used to differentiate multiple EA instances on the same account and to namespace the per-instance memory file + CSV audit log. Must be ≥ 1000 — the EA refuses to attach below that floor (low magics collide with manual orders).

File name (inside MQL5\Files\) of your custom prompt library. On first run the EA auto-deploys a working template here so you don't have to hunt for the format. File presets append after the built-in — cycle order: builtin → file presets in order written.

Safety guard. Every preset declares its optimal timeframe (the verified default declares H4; library presets declare via the TIMEFRAME= header). When this is false (default), the EA forces the chart to the preset's TF on init, on reload, and on every preset cycle — and prints e.g. Quant AI Agents: restoring chart TF H4. When true, the EA respects whatever TF you put it on and runs the preset there anyway.

How position size is computed. LOT_FIXED uses 3.2 as a literal lot value. LOT_LOW_RISK sizes for 0.5% risk per trade. LOT_MEDIUM_RISK sizes for 1.0%. LOT_CUSTOM_RISK reads the percent from 3.3.

Literal lot value used when 3.1 = LOT_FIXED. The default is the minimum tradable size on most brokers — enough to grade EA behavior without meaningful exposure.

Percent of current equity risked on each entry. Used only when 3.1 = LOT_CUSTOM_RISK. The EA back-solves the lot size from this value and the AI-defined SL distance.

The master switch. When false, every other input in this group is ignored. When true, the EA may stack additional entries as price moves against the basket — bounded by 4.2 / 4.4 and the hardcoded group-kill DD safety.

Maximum number of additional entries the EA will add when price moves against the basket. Valid range 1–5. 3 means the original entry plus up to two adds.

How far (in vol-anchored multiples) price must travel against the basket before the next layer fills.

Hard cap on total open volume across all layers in the basket. Even if max layers (4.2) hasn't been reached, no further layer fills once aggregate lots exceed this number.

Percent of day-start equity at which the EA locks itself out for the remainder of the calendar day (broker server time). On rollover the cap resets and the EA resumes.

Percent of week-start equity at which the EA locks itself out for the rest of the trading week. Resets on Monday open.

The HARD HALT. When account equity drops below this percent of the EA's recorded initial deposit, every position closes at market and the auto-trade flag flips off — sticky until you manually re-enable it. Designed as the last-resort backstop after daily, weekly, monthly caps have already failed.

When true, the EA queries the MT5 Calendar API on every cycle and refuses new entries inside the configured window around high-impact events (default: ±2 h on the USD/EUR/GBP/JPY block at importance level 3). Existing positions still get managed by the exit engine — the filter only blocks new opens.

Fires once at attach if the Google Apps Script license check rejects this MT5 account number. The Experts tab is printed first with a multi-line diagnostic (header LICENSE VERIFICATION FAILED) — read it before clicking through the dialog.

The EA caches the MT5 login at init and re-checks every cycle. If the live login differs (you logged into a different account on the same terminal while the EA was attached), this alert fires and auto-trade is disabled until you re-attach. Common cause: switching demo ↔ live on the same MT5 instance.

The EA re-checks its license once per 24h (cheap — 1 HTTPS call). If the check flips from OK to FAIL mid-session (subscription lapsed, refund, revocation), this alert fires, new orders are blocked, and the Experts tab logs === LICENSE INVALIDATED === New orders blocked.

The Group 5.3 equity floor tripped. Every open position from this magic was closed at market, auto-trade is forced OFF, and the kill is sticky — you must manually re-enable from the panel and confirm the diagnosis before the EA resumes. The Experts tab logs === HARD HALT === <reason> — closed=N.

The init scan found at least one agent in Group 1 / 1C whose matching provider key in Group 1B is empty. The next line tells you the default mapping (Technical=GPT-5-mini, Bull=Claude-Sonnet-4.6, Bear=GPT-4o, Research=Gemini-3-Pro, Portfolio=GPT-5). Per-agent rows logged as Agent N (Name): API Key is empty. follow.

The init validator rejects out-of-range inputs and refuses to attach. Possible rows: CandleCount must be between 10 and 200, MaxOrders must be between 1 and 50, MagicNumber must be >= 1000, UpdateIntervalMin must be 0 or between 15 and 1440, DebateRounds must be between 1 and 5, MaxDailyLossPct must be between 0.1 and 50.0, MaxWeeklyLossPct must be between 0.1 and 80.0, EquityFloorPct must be between 10.0 and 99.0, FixedLotSize must be > 0, CustomRiskPercent must be > 0, ExitTriggerATR must be > 0, ExitBufferPoints must be >= 0, ExitTrailDistanceATR must be > 0.

Group 4 (Position Averaging) range validators. Possible rows: DCAMaxLevels must be between 1 and 5, DCALotMultiplier must be > 0 and <= 2.0, DCAMaxTotalLots must be > 0, DCADistanceATR must be between 0 and 10, DCAMaxDrawdownPct must be between 0.1 and 50.0, DCAGroupKillDDPct must be > DCAMaxDrawdownPct and <= 95.0. Only logged when 4.1 DCAEnabled = true.

The fail-fast broker spec audit. On success: one line dumping tick size / tick value / min-max-step lot / stop level / fill modes. On failure: any of trade mode DISABLED — symbol not tradable, is CLOSE-ONLY — new positions blocked by broker, tickSize <= 0, tickValue <= 0, minLot <= 0, maxLot < minLot, stepLot <= 0.

Confirmation the prompt selected via input 2.1 was loaded. Default reads Initial prompt: Stoch MR-XAUUSD-H4. Followed by the full init banner — Quant AI Agents initialized successfully! — with symbol, TF, pipeline, agent models, lot info, DCA summary, V2 features, auto-trade state, max orders, update cadence, exit strategy.

The prompt library file (input 2.6) loaded. N = total presets after merging built-ins with file presets. If the file is malformed: Prompt library: file '<file>' load failed (<err>) — built-ins only. The panel still functions on built-ins — but your custom presets won't appear in the cycle.

ML hook status. Examples: Regime classifier auto-picked: <path>, Regime classifier disabled (no model at <path>), Meta-label gate active: <file>, Meta-label gate disabled (model file not loaded). Informational only — both gates are no-ops when the model files aren't shipped for your symbol.

The cycle ran a gate and the gate refused to run the pipeline. Possible reasons: RiskGovernor block (daily / weekly / monthly cap, loss-streak damper, profit-lock, margin guard, cooldown, aggregate risk cap), news blackout window, TF mismatch when input 2.7 = false. Expected behaviour — not an error.

Two charts share the same symbol + magic and both tried to fire the pipeline. The first chart held the GlobalVariable lock; this chart yielded. Expected when running deliberate multi-instance setups — the lock prevents duplicate API spend on the same symbol.

Hard pipeline error (an agent returned malformed JSON, a WebRequest timed out beyond backup retry, or the panel rejected the verdict). Or, on init, the ATR handle for the exit engine couldn't be created — exit engine inactive, positions managed only by agent-supplied SL/TP.

Randomised attach-time delay so re-attaching every chart in a multi-chart bundle doesn't fire N parallel pipeline calls. On re-attach within the same terminal session: Startup delay skipped (already completed in this terminal session). On first attach completion: Startup delay complete. Running first analysis...

Input 2.7 (AllowTFMismatch) = false. The preset declares a different optimal TF from the current chart TF, so the EA pushes the chart back. The failure variant means MT5 rejected the TF switch — usually no history on that TF for that symbol.

You hit the panel's preset cycle or reload button while a state forbids reconfigure (open position on this magic, pipeline in flight, RiskGovernor hard-halt active). The Experts tab line gives the precise reason. Expected behaviour — protects against config swap mid-trade.

Confirmation that the panel's cycle / reload took effect. Followed (when the chart TF needs switching) by Preset cycle: switching chart <old TF> → <new TF> and (on full pipeline rebuild) Pipeline sync'd to library preset: <name>.

The panel's Reload button pushed a re-parse of the prompt file. Always followed by either -> N builtin+file presets (success) or -> file load failed (<err>) — built-ins only (parser rejected the file). Same file-format rules as P6.

A close-by-stop transaction landed but the entry-side price couldn't be resolved from the deal history (broker delayed the open deal). Audit log row is written with empty entry price; everything else proceeds normally. Cosmetic — no remediation needed unless it repeats every trade.

The per-magic CSV trade log (eatrading_trades_<symbol>_<magic>.csv) couldn't be opened. Usually means the file is locked open in Excel — close it. Less commonly: disk full or MQL5 sandbox permission issue.

Catch-all for low-frequency lines: License rejected (HTTP <code>): <response> (raw upstream rejection — see A1 fix), Auto-deploy preset file: FileOpen for write failed (err N) (sandbox blocked the template seed; create the file by hand instead), Auto-deployed preset template to MQL5\Files\<file> (first-attach success — your starter template is now editable on disk), [Telemetry] license_key not available (len=…) (telemetry needs the license-key UUID returned by Apps Script; legacy "OK"-only responses can't authenticate — request a re-issue from support if you want telemetry credit), and WriteExampleFileIfMissing-related notices.