mytbot

Imported from emfasys-labs/mytbot (AGENTS.md). Install upstream with npx skills add emfasys-labs/mytbot. Copyright stays with the author.

AGENTS.md

==========

This file is for Codex (Codex.ai).

Paste the contents of this file at the start of any Codex conversation

to instantly bring Codex up to speed on the project state.

Update the CURRENT STATE section after each work session.

PROJECT

mytbot — open-source autonomous multi-asset trading system under AGPL v3, published by Emfasys Labs, a division of Emfasys Ltd. GitHub: https://github.com/emfasys-labs/mytbot.git Owner: UK-based, trading stocks, bonds, ETFs, forex, crypto. Primary broker: IBKR Pro. Crypto: Kraken + Binance.

ARCHITECTURE IN ONE PARAGRAPH

One-button system. python run.py starts everything — orchestrator brings up Docker (Postgres, Redis), auto-discovers available brokers (skipping unavailable), runs the trading loop and data pipeline, and exposes an API+WebSocket for the UI. The UI has a single ON/OFF button (POST /system/start, POST /system/stop). All brokers implement a single abstract interface in brokers/base.py. Signals pass through an optional Signal Accumulation Engine (signals/accumulator.py) that maintains time-decayed, multi-source conviction per asset (quant + news + macro) before the signal engine emits a unified Signal. The signal engine aggregates strategy outputs into a Signal (with legacy point-in-time AI modifier and optional accumulated overlay). Every Signal passes through the risk engine (unconditional veto power). Approved signals go to the execution engine which routes to the best broker. Everything is logged. Runtime broker permissions support graceful fallback routing. Risk parameters are managed by ParameterManager with layered overrides (regime > AI > defaults) and bounded, auditable changes. AI is local-first (rules → FinBERT → local LLM → optional paid fallback) — never places orders.

KEY FILES

• run.py — THE entry point (python run.py starts everything)
• system/orchestrator.py — state machine: OFF → STARTING → RUNNING → STOPPING → OFF
• system/dependency_manager.py — auto-start Postgres/Redis via Docker
• system/broker_manager.py — auto-discover and connect available brokers
• system/trading_loop/ — controllable async trading loop package (TradingLoop, helpers.py)
• brokers/base.py — the adapter interface (FROZEN: backward-compatible optional fields only)
• brokers/registry.py — add new brokers here (one line)
• brokers/ibkr/adapter.py — IBKR + single-leg options (chain / qualify / Order.instrument_metadata)
• brokers/ibkr/universe.py — config-backed curated IBKR execution/discovery seed
• brokers/ibkr/qualification.py — persisted IBKR contract qualification cache
• core/instruments.py — OptionContractSpec (options as structured instruments)
• risk/options_env.py — ENABLE_OPTIONS / OPTIONS_* env merged into risk config
• brokers/bybit/adapter.py — Bybit V5 (spot / USDT linear)
• brokers/_template/ — copy this to add any new exchange
• risk/engine.py — risk checks, kill switch, D115 FX + equity-index cluster caps
• risk/intraday_derisk.py — D115 graduated portfolio-level derisk decision layer
• risk/parameters.py — parameter manager (defaults + overrides + expiry)
• signals/accumulator.py — stateful time-decayed conviction per symbol (optional; YAML-gated)
• signals/anti_churn.py — D115 dedup / cross-strategy contradiction / post-fill cooldown gate
• signals/engine.py — signal aggregation + accumulator integration + anti-churn wiring
• intelligence/trade_admission/ — D196 pre-risk trade admission shadow ledger, diagnostics, outcome labeling
• strategies/momentum.py — first strategy (momentum breakout)
• execution/engine.py — order placement
• execution/router.py — smart order routing
• storage/models.py — database schema
• api/server.py — FastAPI: /system/start, /system/stop, /system/status, /dashboard/snapshot, /pnl (week/month)
• api/pnl_periods.py — calendar week/month rollups over daily_pnl for /pnl
• system/dashboard_publish.py — persists allocator snapshot (dashboard.snapshot in ControlState) for the UI
• config/risk_limits.yaml — all risk thresholds (editable without code change)
• config/m8_micro_live.yaml — optional micro-live profile (symbol/strategy/notional caps when APP_ENV=live)
• config/fundamentals.yaml — parameter defaults, absolute bounds, AI policy
• config/broker_permissions.yaml — runtime broker permission/fallback map
• config/ibkr_universe.yaml — curated IBKR priority instruments; qualification still required before orders
• config/data_pipeline.yaml — M2 symbols, intervals, News/FRED toggles
• config/ai.yaml — local-first AI config: providers, escalation policy, no daily caps
• config/session_exit_policy.yaml — broker-aware pre-close hold/trim/close policy by horizon and mode
• config/trade_admission.yaml — D196 Trade Admission Intelligence shadow/enforcement switches
• config/profile_modes.yaml — D015 mode coefficients + emergency safety_bounds (allocator)
• config/allocation.yaml — D015 global opportunity replacement policy (validated by config/models.py)
• config/loaders.py — load_profile_modes() / load_allocation()
• core/models_runtime.py — runtime types: Opportunity, RegimeState, AllocationDecision, ExecutionPlan
• core/session_exit_policy.py — pre-close session-exit decision layer (hold, trim, close, defer)
• signals/volume_anomaly.py — D015 volume/flow features, detection vs scoring, YAML-weighted component
• signals/opportunity_engine.py — D015 opportunities; blends momentum proxy + volume component from M2 JSON + metadata
• risk/regime_state.py — D015 market/regime context for allocator (stub)
• portfolio/allocation_engine.py — D015 global replacement allocator (stub)
• portfolio/balance.py — D222 canonical economic ledger, semantic HRP, factor admission, legacy reconciliation
• portfolio/target_ledger.py — D222 one absolute target owner across allocator/reserve paths
• execution/planner.py — D015 AllocationDecision → ExecutionPlan
• data/regime_metrics.py — cross-section feature fetch + aggregates for regime
• data/feature_lookup.py — latest feature_snapshots.features per symbol
• portfolio/d015_hold_switch.py — hold score + switching-cost penalty
• signals/d015_weights.py — profile/regime dynamic coefficient resolver
• signals/opportunity_components.py — momentum/news/liquidity/structure component scores
• system/d015_shadow.py — optional per-signal D015 vs legacy log (env ALLOCATOR_D015_SHADOW)
• core/signal_math.py — bounded_sigmoid, tanh_clip, normalize_zscore (Decimal)
• run_pipeline.py — M2: yfinance → features → Postgres; NewsAPI + FRED
• scripts/qualify_ibkr_universe.py — no-order IBKR contract qualification/cache warmer
• ai/router.py — local-first AI router (drop-in for NewsClassifier)
• ai/providers/ — provider implementations (rules, FinBERT, Ollama, Codex fallback)
• ai/escalation.py — necessity-based escalation engine (no hard daily caps)
• ai/schemas.py — shared AI types (ProviderResult, EscalationContext)
• ai/news_classifier.py — legacy Codex classifier (kept for backward compat)
• ai/pipeline.py — M6 orchestration: symbol news score + macro regime
• docs/DECISIONS.md — architectural decision log
• docs/BUILD_PLAN.md — milestones M1–M10 + task history
• docs/NEW_MACHINE_SETUP.md — new PC / full reinstall (Python, Docker, Ollama, UI, .env)
• docs/M8_MICRO_LIVE.md — micro-live guardrails (APP_ENV=live, YAML profile)
• alembic/ — DB migrations (URL from POSTGRES_* in env.py)
• tests/ — pytest smoke tests
• .cursorrules — Cursor AI alignment rules

CURRENT STATE

• Milestone: Loss-attribution remediation (P0-P3) ✅
• Last completed task: D231-D237 — Full P0-P3 remediation of the D229 loss-attribution findings, implemented and verified against the live paper DB (read-only investigation first, then reversible code/schema changes; python run.py PID 76848 left running throughout, never stopped/restarted). D231 (P0.1): portfolio/global_edge_coordinator.py::propose_idle_loss_recycle_actions had no minimum-holding-period protection (only a 15-min same-symbol cooldown) — 0% win rate, -$4,963.77 net, 25-min median hold by construction (filters exclusively for currently-underwater positions). Now gated through the same risk/protective_exit_gate.py min-hold logic every other exit path uses, via a new shared storage/fills_ledger.py::fills_age_seconds_by_symbol(). D232 (P1.3): partial trims (profit-harvest/capital-recycle/derisk) that would leave a sub-minimum-order-size remainder now close fully instead — the root cause of 210 open positions with a $942 median size, 28x under the orchestrator's own 2%-NAV floor. Fixed at the single execution/engine.py::_clamp_reduce_only_to_holdings chokepoint (same pattern as D165/D167.2). D233 (P1.4): edge-gate cost model raised from 30bps to 52bps round-trip (config/strategies.yaml::backtest) to match live-measured costs (scorecard found portfolio_orchestrator at 51.3bps, 71% over the old assumption); re-ran scripts/run_edge_gate.py — all four proven weapons (trend_breakout/mean_reversion/trend_following/momentum_breakout) still clear the harsher bar with compressed margins. D234 (P1.5): new fills.opening_strategy/opening_signal_id columns (migration d231a1b2c3d4, applied) trace a closing fill back to the strategy that opened the lot — fills.strategy on a close only ever named the EXIT mechanism (stop_loss_monitor/capital_recycle/etc.), so live expectancy for the four "proven" strategies was structurally unmeasurable (scripts/report_edge_scorecard.py showed live_exp=0.00 for all of them despite 620 combined opening fills). New ENTRY-STRATEGY LIVE EXPECTANCY table in the scorecard. D235 (P2): TradingLoop._orchestrator_strategy_trust no longer lets an optimistic backtest Kelly prior (backtest/edge_gate.py::edge_kelly_trust) amplify sizing above neutral (1.0x) without VERIFIED live evidence (enough closes, net-positive, PF>1) — previously an untested strategy could size up to 1.5x on backtest alone. Also fixed a silent key-mismatch bug: system/strategy_pnl_health.py::fetch_strategy_pnl_recent grouped by the exit-mechanism name, not the entry strategy, so the posterior tilt had likely never engaged correctly in production; now grouped by opening_strategy. D236 (P3.7): new SYMBOL CHURN section in the scorecard flagging any (broker,symbol,day) with >5 fills — routinizes the manual-SQL discovery of AAPL (30 fills/8d), ETH-USD (55/8d) etc. D237 (P3.8): edge_gate_verdicts.json was 3+ weeks stale and silently kept governing live sizing; new max_verdict_age_days config (default 14) triggers a runtime warning + scorecard header flag when exceeded. Full test suite: 2,187 passed, 3 skipped (from a 2,149 baseline). See docs/DECISIONS.md D231-D237 for full detail on each.
• Last completed task: D230 — Paper runtime restored by operator direction. The operator clarified that the intentionally stopped D229 runtime must remain on because this is a paper environment. Called /system/start and verified state=running, paper_mode=true, trading.running=true, pipeline running, all ten configured brokers connected and included, full accounting coverage, and no loop error. D229's loss attribution and design findings remain valid, but its mandatory shutdown/restart prohibition is superseded: paper trading stays active as the experiment and remediation environment unless the operator asks otherwise. No mass liquidation was performed. See docs/DECISIONS.md D230.
• Last completed task: D229 — Full loss attribution and paper-system warning. Operator challenged the dashboard's -£17,510.02 historical closed P&L. The value reconciles exactly to fills: -£10,912.51 gross realised P&L minus £6,597.51 fees across 1,175 fills. Net attribution: stop-loss -£12,820.49 (mostly D228 crystallising the already-losing legacy book), capital recycle -£4,963.77, portfolio orchestrator -£4,449.45, local paper repairs -£2,653.88, mean-reversion entry fees -£1,421.56, trend-following entry fees -£1,155.64, other paths about -£476, offset by profit harvest +£10,430.92. Largest symbol damage: MARA -£5,093, SMH -£2,743, AAPL -£2,515, AUDUSD -£2,022, ETH -£1,126, BTC -£949. The authoritative ledger still contains 210 open positions (the UI /positions response caps at 200), about £1.025m gross exposure and -£3,093 unrealised P&L: 109 Alpaca, 99 IBKR, one Kraken, one Capital.com. On 2026-07-03 the ledger recorded 15 fills around 00:09–00:49 London (nine stop-loss/profit-harvest actions, net about -£874); no fill occurred afterward, despite 308 loop iterations. The system was briefly stopped as a precaution, then restored by D230 following explicit operator direction. No mass liquidation was attempted. See docs/DECISIONS.md D229.
• Milestone: Protective-risk symmetry ✅
• Last completed task: D228 — Explicit stops can no longer be vetoed by minimum hold. Operator reported unrealised loss growing beyond -$10.3k. The D227 reserve fix had stopped new averaging down, but live logs proved a severe winner/loser asymmetry: profit harvest banked winners while both stop-loss and tier-2 intraday derisk repeatedly emitted valid actions and then suppressed them as anti-churn:within_min_hold. MARA was down about -$4,992 / -10.3% and had breached its NAV-relative loss budget; numerous small equities were 10–19% down, yet the three-day strategy maturation window overruled their explicit 6% position stops. Protective min-hold is now scoped only to soft derisk/rebalance noise: explicit portfolio-budget stops and position stops always fire for every asset class, while structural/catastrophic/most-severe-tier exceptions remain. The first repaired run closed 26 breached paper positions reduce-only, including MARA and SMH, realising about -$11,464 gross with $84.81 fees and reducing unrealised P&L from about -$10,347 to +$928; this moved the already-existing economic loss rather than creating it. Live proof then exposed one stale-snapshot oversell: AAOI was closed correctly, but a later monitor pass used an obsolete positive PositionLog after the authoritative fills ledger was flat and created a short. Removed that fallback—once fills exist, SUM(fills.signed_quantity) is unconditionally authoritative—and flattened only the accidental AAOI paper short. Final full suite: 2,149 passed, 3 skipped. Restarted python run.py PID 76848: running in paper mode, nine brokers included, no loop error, no negative ledger position, and no fill after the clean restart; IBKR Gateway/TWS remains externally unavailable on port 7497. See docs/DECISIONS.md D228.
• Milestone: Reserve-path churn containment ✅
• Last completed task: D227 — Negative-expectancy averaging-down bypass closed. Operator correctly challenged the roughly -$6.7k unrealised / -$5.3k daily net result and 200-position book. The D226 allocator fix worked, but a separate portfolio_orchestrator_reserve path executed 21 direct trend_following opens from 15:37–15:56, repeatedly adding to falling MARA, SMH, EEM, DAL, COPX, and AAPL. Its own learned metadata reported negative expected returns (for example MARA about -0.000457, EEM about -0.001740) and a below-threshold meta model, yet deployment-pressure relief converted those rejects into small paid exploration orders. Daily feature refresh timestamps also masqueraded as new bars, and reserve orders bypassed no_average_down. Fixed all three ownership points: mature non-positive expected-return buckets now receive zero capital and an enforced admission reject; pressure relief is disabled; reserve additions are blocked when the existing same-side position is underwater; feature timestamps are canonicalised to the actual timeframe bar; and the target ledger permits at most one reserve increase per symbol/bar. Focused verification: 89 passed; full suite: 2,147 passed, 3 skipped. Restarted python run.py PID 71676. First post-fix cycle: 420 candidates, 32 below-threshold candidates filtered, zero strategy opens/averaging-down fills; nine profit-harvest reductions realised about +$361 gross with $18 fees. Runtime is running with no loop error; nine brokers are included, while IBKR is externally unavailable because Gateway/TWS is not listening on 127.0.0.1:7497. See docs/DECISIONS.md D227.
• Milestone: Loss-crystallisation control ✅
• Last completed task: D226 — Signal silence can no longer manufacture allocator exits. Operator reported continued losses. The book was initially about +$4.4k unrealised on the day, but the allocator had crystallised roughly -$1.29k across 21 portfolio_orchestrator fills and paid about $207 in allocator fees. Losing closes such as XLU, QCOM, COP, VCSH, KO, IYR, HYG, and XLP carried orchestrator_reason=close, zero conviction, and no contributing strategy: a candidate disappearing from one cycle was being treated as a flat target and therefore as exit evidence once age/edge protection expired. OrchestratorConfig.close_on_signal_silence now defaults false and the live YAML explicitly keeps it false; silence holds the position, while explicit opposing signals, stop-loss, derisk, session exits, reconciliation, and evidence-backed recycling retain authority. Profit-harvest repeat cooldown increased from 90 seconds to one hour after 41 profitable-but-fee-heavy trims in under an hour. Full suite: 2,144 passed, 3 skipped. Restarted python run.py as PID 71792 in paper mode. Post-restart evidence: first iteration complete, ten of ten brokers included with full coverage and no loop error; zero zero-conviction/strategy-less allocator reductions; the only allocator reduction had explicit trend_following evidence and realised about +$85.94; initial profit banking realised about +$1,867 gross with about $33 fees. Later live marks moved the daily total back to about -$1,949 net without any additional fill, led by MARA, SMH, and EEM; their refreshed Yahoo/broker prices agree and their trend signals remain explicitly long, so that movement is market risk rather than the repaired forced-loss path. See docs/DECISIONS.md D226.
• Milestone: Full startup and runtime health repair ✅
• Last completed task: D225 — Full startup-health repair and final legacy balance reconciliation. Audited the supplied startup and found the system operational but not clean: balance invariants conflated policy-approved advisories with hard defects; the safe legacy plan still contained AGG/BND/VXUS; auto-training mishandled insufficient meta-label history and used ridge for a classification target; hosted local-reasoning had no credential; request logs exposed feed keys; expected Yahoo/IBKR discovery misses and startup coverage readiness produced false error/critical records; optional constituent sources were stale; and Bybit's derivatives open-order query omitted its settlement scope. Fixed each owning path, rebuilt the UI, restored local Ollama qwen2.5:7b (startup plus real JSON scoring passed), and verified auto-training end-to-end with clean skips and four successful forecast models. The existing reduce-only, paper-only, one-action-per-cycle reconciliation removed AGG, BND, and VXUS (about -$214.27 realised, $20.30 simulated fees) while retaining IUSB/EFA. Final full suite: 2,143 passed, 3 skipped. Restarted python run.py as PID 59464: ten of ten brokers included, full coverage, pipeline and infrastructure healthy, all four AI stages active, no loop error, and fresh runtime.invariants healthy=true / balance_healthy=true. Future logs redact secrets; feed keys exposed in historical logs still require external rotation. See docs/DECISIONS.md D225.
• Milestone: Broker credential rollback recovery ✅ 10/10
• Last completed task: D224 — Broker credential rollback investigated and all ten venues recovered. The dashboard correctly showed five brokers offline because .env had been replaced with an older 4,834-byte version at 2026-06-29 23:37. Cursor local-history metadata identifies the operation as Undo Reject Diff; the restored revision omitted Capital.com, Coinbase, IG, OANDA, and Trading 212 credential blocks, while /system/status::coverage.full remained true only relative to the five credentials still configured. Recovered Capital.com, IG, and Trading 212 from the prior Cursor .env snapshot and Coinbase from the original local Downloads/cdp_api_key.json, without printing secret values. The operator supplied separate OANDA tokens: read-only endpoint checks proved the first was live-only and the second practice-only, so each is stored in its matching environment slot and live OANDA trading remains disabled. Fixed OANDA credential resolution and broker discovery so tokens can no longer fall back across live/practice environments; focused verification: 37 passed. Restarted python run.py as PID 59564: all ten brokers are authenticated, connected, balance-ready, and included; accounting coverage is full with zero exclusions and no loop error. See docs/DECISIONS.md D224.
• Milestone: Portfolio economic-risk balance ✅
• Last completed task: D222 — Canonical economic portfolio balance and gradual legacy reconciliation. The first open-market audit after D221 found exact AUDUSD/CME venue duplicates, near-identical bond/index/regional ETF stacks, FIDD behaving as a USD peg, extensive ETF/constituent overlap, 70% of deployed exposure at Alpaca, and crypto contributing about 49% of estimated risk from only 7.65% of NAV. Implemented all nine portfolio-balance controls: economic alias/venue aggregation; explicit alpha/hedge/cash/reserve roles; expanded ETF/factor look-through; semantic HRP blended with conviction; one absolute target ledger shared by primary and reserve paths; same-feature-bar reduction tombstones; all-asset venue consolidation; final risk and preflight overlap admission; broker-load-aware routing; balance-aware runtime invariants; fill-derived holding age; and a costed, one-action-per-cycle paper reconciliation plan with automatic live cleanup disabled. Full suite: 2,136 passed, 3 skipped. Restarted python run.py as PID 73924; ten of ten brokers are included, the first final-policy iteration completed, and there is no loop error. D222 paper reconciliation removed the redundant IBKR AUDUSD expression (about +$7.54 realised, $6.76 fee) and the FIDD-USD cash-equivalent alpha position (about +$0.02 realised, $4.02 fee). The balance invariant remains intentionally unhealthy while the legacy book still contains the CME venue duplicate, true-substitute bond/ex-US stacks, broker concentration, crypto breadth, and two tiny remnants; its safe cleanup plan contains only those cases and no automatic sector/style liquidation. See docs/DECISIONS.md D222.
• Milestone: Publication readiness ✅
• Last completed task: D221 — Fifth clean live stability round. Operator requested another independent round after D220. Audited the untouched PID 60220 from iteration 13 through 15 with 150 endpoint probes across /healthz, /system/status, /pnl, /positions, and /dashboard/snapshot; zero requests failed and maximum latency was 533ms. Final live audit found zero defect log signatures, no loop error, ten of ten brokers included, no disabled broker, all computed runtime invariants healthy, 384/384 filled orders linked to fills, zero working orders, no duplicate same-symbol venue exposure, no cash-equivalent position, and no new fills or fees. Unrealised P&L moved with market marks only; no loss was crystallised during the round. Sunday inactivity remained explicitly accounted for by closed equity/FX sessions and crypto edge gates. See docs/DECISIONS.md D221.
• Last completed task: D220 — Fourth clean round found and repaired expected Yahoo no-data error pollution. Operator requested one more clean round after D219. The initial audit ran across two fresh cycles and remained healthy for trading, broker coverage, API latency, accounting, and runtime invariants, but correctly failed the clean criterion because Yahoo discovery probes emitted repeated error-level possibly delisted; no price data found records for unsupported broker-catalogue symbols such as SYRUPUSD, SUIFDUSD, and SOLFDUSD. data/yfinance_fetch.py now installs a narrow logging filter for only Yahoo's known no-data/delisted diagnostics; unexpected provider errors remain visible, while empty history continues to score zero and exclude the symbol. Added tests/test_yfinance_fetch.py. Restarted python run.py PID 60220 and repeated the round from a new baseline: three live cycles, zero defect log signatures, zero endpoint failures, ten of ten brokers included, no disabled brokers, no loop error, zero invariant violations, zero new fills/fees, no working orders, no duplicate or cash-equivalent exposure, and 384/384 filled orders linked to fills. Full suite: 2,122 passed, 3 skipped. See docs/DECISIONS.md D220.
• Last completed task: D219 — Three consecutive clean stability rounds after proactive runtime fixes. Continued auditing until three complete post-fix rounds found zero new defects. Before restarting the clean count, fixed four additional runtime faults: instrument availability rebuilt each broker catalogue for every one of ~30k symbols and persisted each row with a separate statement; catalogue membership is now constant-time and availability writes use bounded bulk upserts. /pnl synchronously refreshed all broker balances after cache expiry; API NAV now serves the last coherent snapshot while one background refresh runs. Persisted coverage disables could strand a recovered broker after restart; risk broker gates now retain independent disable reasons, allowing coverage recovery to remove only its own gate without overriding manual gates. Trading 212 account-summary reads now have endpoint-specific caching and stale fallback on HTTP 429. Full suite: 2,119 passed, 3 skipped. Clean rounds then covered seven live iterations, 209 stressed /pnl reads plus final endpoint samples, scheduled and on-demand runtime invariants, broker coverage, logs, fill/order/position reconciliation, duplicate exposure, cash-equivalent exposure, and candidate-path explanations. Final state: running, paper mode, ten of ten brokers included, no disabled brokers, no loop/log errors, all invariant counts zero, 372/372 filled orders linked to fills, no new fills or fees, and final dashboard endpoints below 70ms. No new orders during the Saturday window is explained by closed equity/FX sessions, crypto edge gates, and same-feature-bar churn prevention. Current python run.py PID 50668. See docs/DECISIONS.md D219.
• Last completed task: D218 — Proactive full-system money-path audit and continuous invariants. Operator correctly objected that defects were only being found after visible losses. Ran a proactive audit across live health, fill/position reconciliation, order lifecycle, churn, P&L attribution, admission learning, routing, paper execution, configuration, and the complete test suite. In addition to D217, fixed three more defects: (1) rerouted fills were labelled against the pre-route broker, leaving valid crypto outcomes unpriced; outcome pricing now uses the actual matched fill broker, and 32 historical rows were repaired (34 positive, 47 negative, 360 not_executed, zero recent unpriced); (2) unsupported crypto could be probed through IBKR and synthetic paper mode could fill a product whose venue could not return a market price; router alternatives now honor symbol-specific broker translation and synthetic crypto opens fail closed on unavailable venue prices, while reduce-only exits remain allowed; (3) explicit empty OANDA credentials could inherit ambient practice credentials, defeating validation; explicit constructor credentials now fully override ambient token inheritance. Updated economic semantics for USAT cash-equivalent and WBETH wrapped ETH, then flattened those two invalid paper exposures locally (combined realised about +$6.42, zero fees). Added system/runtime_invariants.py, run continuously by the trading loop and persisted as ControlState(runtime.invariants), checking fills-to-position equality, filled orders without fills, stale working orders, and recent unpriced outcomes; unhealthy reports are logged as errors and attached to heartbeat telemetry. Full suite: 2,114 passed, 3 skipped. Restarted python run.py PID 55364; live runtime.invariants published healthy=true with all counts zero, ten brokers active, no loop error. Earlier three-cycle soak had no new strategy fills/fees, idle recycle, unpriced rows, stale orders, ledger mismatch, or unsupported-symbol errors. See docs/DECISIONS.md D218.
• Last completed task: D217 — Unconditional idle-loss crystallisation replaced with learned successor evidence. Operator reported that loss kept increasing. Live audit showed the main deterioration was not ordinary marking: eight idle_loss_recycle closes in six hours crystallised about -$1,006 realised P&L plus $31 closing fees (INJ, WLD, ZEC, XTZ, KSM, LDO, MAV, KITE). The loop then spent the released crypto venue room on many small, weak candidates. Root cause: D207/D216 required a losing position and low remaining-edge proxy, but still did not require a concrete superior replacement, so deployment pressure alone could manufacture realised losses. GlobalEdgeCoordinator.propose_idle_loss_recycle_actions() now requires named replacement evidence with positive matured expected return, matching asset-class capital, and expected advantage greater than held remaining edge plus switching cost. Both active allocator paths derive that evidence from the live trade-admission model; abstaining or negative-expectancy buckets cannot trigger this discretionary close. Normal stop-loss, risk derisk, profit-harvest, session exits, and evidence-backed rotation are unchanged. Restarted python run.py (PID 58184). Verification: 141 passed, compile clean; two complete live iterations had zero recycle fills, no new fees or realised loss, runtime healthy with ten brokers and no loop error. Current unrealised P&L continued moving with crypto prices, but was no longer forcibly crystallised. See docs/DECISIONS.md D217.
• Last completed task: D216 — Learned reserve sizing, recycle churn, and paper P&L marks repaired. Operator reported "only losses." Live audit proved four bugs: reserve outcome learning reduced metadata targets but quantity math cancelled the multiplier (USDT requested ~$5.8k but filled $33.1k; WBTC requested ~$2.0k but filled $11.8k); idle-loss recycling closed any tiny red tick regardless of remaining edge/switching cost and then reopened the same daily signal; cash-equivalent stablecoins and wrapped BTC were treated as independent directional exposure; and paper APIs replaced the loop's persisted ACGL venue mark ($97.555, about -$35) with an older feature close ($94.33, about -$1,993). Fixed reserve quantity as learned remaining target / price; added core/instrument_semantics.py for cash-equivalent role and wrapped-underlying canonicalisation; excluded cash equivalents in candidate, primary, and reserve paths; made cull re-entry wait at least one live feature interval; made idle-loss recycle require remaining edge <= fee-adjusted switching cost; made closed-market marks preserve the last executable quote; and made paper /positions//pnl trust persisted loop marks. Removed existing USDT/RLUSD/WBTC paper positions reduce-only. Live proof: corrected RLUSD filled exactly its learned $5,738 target before expanded semantic filtering; six final cycles ran without errors; stablecoins logged ineligible_instrument_role; same-bar recycled signals logged culled_without_new_feature_bar; no repeated recycle/reopen fills; ACGL displays about -$35; open unrealised about +$1,010; current-day realised + unrealised - fees about +$950. Historical TWR remains about -0.59% and was not erased. Current process: python run.py PID 12104, paper, nine brokers active. Verification: 162 passed, 1 skipped. See docs/DECISIONS.md D216.
• Last completed task: D215 — Ranked-universe deployment starvation repaired. Operator reported that roughly 9% capital use was still not useful. Live audit found four linked causes: the loop selected the first 500 feature symbols alphabetically before intersecting ranked tiers; trend_following required 201 rows while the loop loaded 200 and therefore could never emit; newly ranked names had only one month of daily history; and Yahoo provisional volume-only rows with non-finite OHLC were persisted as the latest bar, disabling all strategies for affected equities. Fixed ranked-symbol availability queries, added shared enabled-strategy history resolution (strategies/history_requirements.py), expanded live feature windows automatically, added one-attempt-per-process history warm-up for insufficient ranked names, rejected/filtered incomplete OHLC at ingest/validation/read boundaries, and removed 3,826 malformed stored rows. Added early portfolio/reserve session checks so closed-market candidates remain audited without expensive broker preflight. Live proof after final restart: cycles generated 43, 43, 16, 43, 43 candidates with no error, no fills, and no new fees; history-ready symbols rose 255 → 323 and malformed rows stayed at zero; allocator saw 37 intents/33 orders and about $1.15M aggregate target demand, but correctly deferred equity/FX because it was Saturday; BTC/ETH/LINK were already above their learned negative-outcome targets. Gross live exposure remains about 9.54% until actionable non-crypto venues reopen. Current process: python run.py PID 20700, paper, nine brokers active; Capital.com excluded for invalid API key. Verification slices: 93 passed, 1 skipped, 80 passed, 73 passed, reserve preflight 9 passed. See docs/DECISIONS.md D215.
• Last completed task: D211-D214 — Fee-churn incident repaired with learned absolute targets, reserve netting, recycle memory, coalesced NAV reads, and authoritative paper venue consolidation. Live audit found today's main loss mechanism: about -$2.1k realised P&L plus $2.49k fees over 186 fills. Reserve orders repeatedly submitted full targets; the active orchestrator dropped deployment/recycle context; 135 executed opens were unpriced; monitors flooded broker balance/price APIs; and crypto consolidation queried adapters that cannot see synthetic paper positions. Fixes: both primary and reserve paths apply payoff-aware learning to absolute targets before calculating gaps; allocator/reserve paths share persisted replacement context and candidate metadata; outcome labels use PositionLog marks and historical rows were relabelled (201 trained rows, 26 positive/175 negative); the model uses pooled strategy/asset evidence and shrinking 1/sqrt(n) exploration; models load before the first cycle; post-learning adjustments reuse the allocator's NAV-relative rebalance band; runtime NAV reads are coalesced; paper /positions and /pnl use local persisted marks; crypto consolidation reads PositionLog. Live proof: BTC logged target already met; XRP converged from about $49.8k to $21.7k near its learned target; the last verified cycle evaluated all reserve candidates as target-met and executed zero orders; no fill occurred after the final restart; /pnl now responds in about 95ms. Current process: python run.py PID 25972, paper, ten brokers. Verification culminated in 192 passed, execution 48 passed, reserve 61 passed, and API/P&L 54 passed, 1 skipped. See docs/DECISIONS.md D211-D214.
• Last completed task: D210 — Continuous monitor/fix loop: reserve multi-open, neutral-news sizing, harvest churn, and crypto venue room. Operator asked Codex to keep checking until resolved instead of waiting for repeated prompts. Continued live monitoring after flat/underdeployed state. Fixed three active blockers: (1) system/trading_loop/loop.py::_run_orchestrator_reserve_candidates() now honors global_edge.max_actions_per_tick and can execute multiple independent reserve candidates per loop instead of returning after the first success; (2) intelligence/trade_admission/policy.py no longer applies directional-news size haircuts for neutral/supportive signed news (news_directional >= 0.5), so ai_news_score=0 stops shrinking orders; adverse news still produces continuous sizing penalties; (3) execution/engine.py reroutes new-symbol crypto paper opens to a synthetic venue with better effective deploy room before clamping, while same-symbol consolidation remains stronger so one symbol is not fragmented across brokers; (4) risk/profit_harvest.py suppresses young immaterial voluntary profit-harvest trims to prevent immediate churn. Restarted python run.py (PID 14716). Verification: pytest tests/test_execution_engine.py tests/test_trade_admission.py tests/test_global_edge_preflight.py -q 70 passed, earlier profit-harvest/wave2 slices passed, py_compile clean. Live validation: runtime running, paper, 10 broker adapters connected, pipeline/news healthy, no loop error; latest checked loop generated=4 executed=2; BTC initially routed to Kraken then rerouted kraken -> binance because Binance had better effective paper room; profit-harvest logs SUPPRESSED (young_immaterial_profit) for fresh small winners. Current paper book at last check: ADA-USD on Binance, BTC-USD on Binance, ETH-USD on Kraken; no AAPL. Held cash/deployment rose from zero/flat to about $61.3k cash at work; remaining underdeployment is due the current candidate set and active trained-meta skips for AAPL/XRP rather than broken execution. See docs/DECISIONS.md D208-D210.
• Last completed task: D207 — Underdeployed loser dead-zone recycle wired into orchestrator. Operator said the system was stable but useless: about 23.5% capital used and still losing. Live audit showed infra healthy, but after D206 the active portfolio_orchestrator path was rejecting most new opens while still holding losing AUDUSD/XRP positions and emitting no reduce/close actions. Root cause: global-edge capital recycle existed, but the active orchestrator path bypassed it, so the book could sit underdeployed with weak losers indefinitely. Added GlobalEdgeCoordinator.propose_idle_loss_recycle_actions() governed by config/global_edge.yaml::capital_recycle; it only considers holdings with negative live unrealised return, ranks by weakest expected remaining edge, then worse return/larger notional, and emits normal reduce-only trim_symbol actions with capital_recycle_reason=idle_loss_recycle. Wired it into both global-edge and active portfolio_orchestrator ticks when no other reduce action exists, with boot-warmup suppression preserved and full signal/preflight/risk/admission/execution path retained. Restarted python run.py (PID 47452). Verification: focused tests 139 passed, py_compile clean. Live validation: iteration 2 closed stuck AUDUSD reduce-only (sell 351901 @ 0.68961240, realised about -$780.83, fee about $182.01); iteration 3 closed stuck XRP-USD reduce-only (sell 47610.46789290 @ 1.04322699, realised about -$230.30, fee about $37.25). Latest /positions is empty; /pnl has unrealised 0, realised about -$281.78, fees about $1,274.92, 74 trades; runtime is running, paper, 10 brokers connected, pipeline running, no loop error. See docs/DECISIONS.md D207.
• Last completed task: D206 — Allocator-open meta gate + crypto dust-room routing fixed. Operator asked why the system was still losing and not opening more positions. Live audit showed infra healthy, but bad allocator opens were previously bypassing the active trained meta-label gate because allocator-selected opens were treated as shadow/exempt (meta_label_kept=false, meta_label_shadow=true). This allowed AAPL to be bought again and let weak BTC/ETH/XRP candidates reach execution. Fixed signals/engine.py so only true operator/reduce exits are exempt; allocator open_strategy candidates now enforce the trained gate. Also fixed execution/engine.py crypto no-native-paper routing: venue deploy room that is above zero but rounds to zero cents is now treated as exhausted before clamping, so BTC/ETH can reroute to another crypto venue with room instead of becoming zero_notional_after_accounting_rounding; no-fallback skips now report crypto_venue_room_below_accounting_minimum. Flattened the 2-share AAPL remnant earlier via scripts/flatten_local_paper_book.py --apply --brokers ibkr --symbols AAPL. Restarted python run.py (PID 47236). Verification: focused tests 70 passed, py_compile clean; post-fix loop iteration 2 completed with no error. Since restart: 110 no_setup, 2 edge_gate_blocked, 2 filtered_signal_engine/meta_label_below_threshold, no admission rows/fills in the first checked loop. AAPL correctly blocked (0.2249 < 0.228); BTC/ETH also blocked (0.1759 / 0.1344 < 0.228). Current latest paper book at 2026-06-26 16:15 UTC: AUDUSD long about -$532 unrealised, XRP-USD long about -$139, no AAPL. See docs/DECISIONS.md D206.
• Last completed task: D205 — Poor-performance incident controls fixed. Operator asked why performance remained extremely poor. Live audit found infra healthy but behaviour leaking money: TWR about -0.42% since 2026-06-25, net trading P&L about -$5.26k, all-time fees about $1.89k / 145 fills, AAPL historical damage, and continuing AUDUSD/XRP unrealised loss. Root causes were churn/fees, profit-harvest selling gross winners without proving close-leg fee coverage, adaptive tuner exploring or increasing risk after losing/no-evidence rewards, and trade-admission learning that waited for longest horizon and could label still-open executed trades as fee-only losses. Fixes: TunableParam.loss_guard_direction; adaptive optimiser now takes one bounded de-risk step on negative reward; tuner skips no-fill attribution windows; config/adaptive_tuner.yaml declares loss directions; profit-harvest suppresses cost-not-covered reduce orders; trade-admission labels at first matured horizon, marks opens to market from price_history/feature_snapshots, and labels unavailable mark rows as unpriced so they do not train. Recomputed recent ledger: not_executed=625, unpriced=53, negative=27, pending=48. Restarted python run.py (PID 51844). Verification: focused tests 44 passed (known AsyncMock warnings), py_compile clean; runtime paper, 10/10 brokers, pipeline running.
• Last completed task: D204 — Direct-news hard veto replaced by tunable learning penalty. Operator correctly rejected the D203 hard-coded direct-news veto approach. Removed direct-news hard rejection from risk/engine.py and removed negative_direct_news_long / positive_direct_news_short explicit reject logic from intelligence/trade_admission/policy.py. feature_builder.py now derives news_directional from signed ai_news_score and side; policy.py blends absolute news evidence with directional evidence through AdmissionConfig.directional_news_weight and applies continuous directional_news_size_adjustment sizing haircuts instead of blocking. Added directional_news_weight to config/trade_admission.yaml, added trade_admission.directional_news_weight to config/adaptive_tuner.yaml, taught adaptive tuner defaults to load trade_admission config, and wired live tuner overrides into TradeAdmissionService.apply_live_overrides() from the trading loop. Restarted python run.py (PID 25256). Verification: runtime running, 10/10 brokers connected, pipeline running; directional_news_weight=0.5; tuner contains trade_admission.directional_news_weight; focused tests 54 passed, 1 skipped, py_compile clean. See docs/DECISIONS.md D204.
• Last completed task: D203 — Direct negative news veto + AAPL paper flatten. Operator reported AAPL was bought again while the UI showed negative company/news impact. Root cause: signed direct ai_news_score existed, but trade admission scored absolute news magnitude as generic evidence, so negative company news could strengthen a candidate instead of blocking a long. Added signed direct-news features in intelligence/trade_admission/feature_builder.py, active admission vetoes in policy.py (negative_direct_news_long, positive_direct_news_short), and a final risk-engine hard veto in risk/engine.py; reduce-only exits remain allowed. Extended scripts/flatten_local_paper_book.py / system/local_paper_flatten.py with --symbols and flattened only ibkr:AAPL from the local paper ledger (sell 893 @ 275.21, reduce-only tombstone). Restarted python run.py (PID 45400). Verification: /positions and dashboard snapshot show no AAPL; runtime running, 10/10 brokers connected, pipeline running; focused tests 64 passed, 1 skipped, py_compile clean. See docs/DECISIONS.md D203.
• Last completed task: D202 — Paper book/P&L wiped again and restarted from clean slate. Operator asked to wipe everything again and start from scratch. Stopped system via /system/stop and waited for state=off, then ran python scripts/reset_trading_data.py --execute. Script truncated orders, positions, fills, daily_pnl, signals, risk_decisions, strategy_candidate_log, thesis_log, anomaly_log; deleted trading-derived control_state keys; removed data/runtime/risk_state.json and data/runtime/paper_wallet.json. Restarted via /system/start; runtime is state=running, 10/10 brokers included, coverage full, trading_running=true, pipeline running, first clean iteration completed at 2026-06-25T19:06:50+01:00, no loop error. Verification: /positions returns []; /pnl returns realised/unrealised/fees/trades all zero with TWR 0.0; DB counts after restart positions=0, fills=0, orders=0, daily_pnl=1 (today's zero baseline row). D201 active gates remain enabled.
• Last completed task: D201 — Shadow governance promoted to active paper gates. Operator asked not to keep newly developed controls in shadow mode. Promoted trained meta-label gating (config/strategies.yaml: use_trained_meta_labeler: true, trained_meta_labeler_shadow: false), enabled active trade-admission sizing/blocking (config/trade_admission.yaml: shadow_only: false, block_new_opens: true, allow_size_haircuts: true), and made the regime transition detector affect allocator exposure by tightening RegimeState.drawdown_throttle when non-shadow stress probability is emitted (config/regime_models.yaml::transition_detector.shadow_only: false). Bridged execution microstructure evidence into trade admission by mapping microstructure_shadow_label to microstructure_label; disabled read-only shadow diagnostics that did not affect order flow (FUSION_SHADOW=0, demand learned_graph_shadow.enabled=false, transition_policy_shadow.enabled=false). Tests: python -m py_compile intelligence\trade_admission\feature_builder.py execution\microstructure_shadow.py risk\regime_state.py signals\engine.py; pytest tests\test_trade_admission.py tests\test_wave2_wiring.py tests\test_phase_c_regime_transition.py -q 27 passed. Restarted python run.py (PID 39352); runtime is state=running, 10/10 brokers connected, can_trade=true, first loop completed. Current paper book remains AAPL/XRP/YM from before this fix; /pnl today about -$8.2k net trading P&L with AAPL/YM carrying most unrealised loss. See docs/DECISIONS.md D201.
• Last completed task: D200 — Trade admission now actively blocks bad new opens after AAPL loss. Operator asked why unrealised P&L went from about +$2k to -$5k and why a negative AAPL signal did not prevent buying Apple. Live audit after D199 reset: current paper book had AAPL long on Capital.com (881.81542670 shares, avg 280.37, mark ~277.97, unrealised about -$2.1k), YM=F long on IBKR (5 contracts, unrealised about -$3.2k), and XRP-USD long on Binance (about +$271). AAPL was opened at 2026-06-25 13:41 UTC by portfolio_orchestrator/mean-reversion despite meta_label_kept=false, meta_label_reason=below_threshold, and trade_admission_decision=defer/prior_trade_filter_drop because trade admission was shadow-only (shadow_only: true, block_new_opens: false). Stopped the system immediately (state=off). Changed config/trade_admission.yaml to shadow_only: false and block_new_opens: true; reduce-only/close actions remain allowed. Tests: pytest tests/test_trade_admission.py -q 7 passed, py_compile clean. See docs/DECISIONS.md D200.
• Last completed task: D199 — Operator-requested paper book/P&L reset. Operator clarified "wipe everything" meant current open positions and P&L, not repo/config. Stopped the system first; confirmed paper_mode=true and /positions source was local position_log. Initially reset core local paper ledgers (positions, fills, orders, daily_pnl) and stale UI/runtime state, then operator remembered the repo reset tool; ran python scripts/reset_trading_data.py dry-run and python scripts/reset_trading_data.py --execute. The script truncated orders, positions, fills, daily_pnl, signals, risk_decisions, strategy_candidate_log, thesis_log, anomaly_log; deleted trading-derived control_state keys while preserving auto_training.last_run_at and system.capital_allocation; removed data/runtime/risk_state.json and data/runtime/paper_wallet.json. Verification: /positions returns []; /pnl returns realised/unrealised/fees/trades all zero; all reset tables are 0; only preserved control keys remain; bot intentionally left OFF after reset.
• Last completed task: D198 — Crypto same-symbol venue consolidation after ADA fragmentation. Operator asked why ADA was open on Binance, Bybit, and Kraken. Live audit showed one ADA target was repeatedly topped up: Binance filled first, then the no-native-paper crypto execution path rerouted additive ADA orders to Bybit/Kraken when per-venue synthetic paper room was exhausted. Risk capped total ADA exposure, but no guard enforced one same-symbol venue expression. Fixed execution/engine.py: additive no-native-paper crypto orders now inspect existing Binance/Bybit/Kraken positions; if the symbol is already held, adds route only to the dominant existing venue, and if that venue was already attempted/full the add is skipped instead of opening a fresh broker row. Reduce-only/close orders remain exempt. Also skip accounting-zero crypto dust orders before they create filled zero-notional rows. Tests: python -m py_compile execution\engine.py tests\test_execution_engine.py; pytest tests\test_execution_engine.py -q 45 passed. See docs/DECISIONS.md D198.
• Last completed task: D197 — Trade admission schema drift fixed; orders flowing again. Live health audit on 2026-06-25 showed the system itself was healthy (state=running, paper, 10/10 brokers connected/balance-ready, Postgres/Redis healthy, AI/news fresh, pipeline running), but orchestrator orders were failing after RISK APPROVED because the existing trade_admission_log table lacked new nullable JSON columns outcome_horizons and outcome_labels. Applied live DB ALTER TABLE ... ADD COLUMN IF NOT EXISTS for both, added the same self-healing patch in storage.db._ensure_additive_schema_patches(), and verified the next loop executed/fill path recovered: iteration #4 generated 9, executed 4 paper fills (XLE, USO, GLD, ADA-USD). python -m py_compile storage/db.py passed; /diagnostics/trade-admission works again; /system/status returns no errors. See docs/DECISIONS.md D197.
• Last completed task: D196 — Trade Admission Intelligence shadow ledger. Built a myTbot-native pre-risk admission layer from scratch. New intelligence/trade_admission/ package provides schema/config, feature extraction, conservative admission decisions, persistence, outcome labeling, and diagnostics. Added storage.models.TradeAdmissionLog to preserve the exact executable-candidate context and downstream status. TradingLoop now evaluates built signals in both legacy and global/orchestrated chokepoints before risk, stamps admission metadata, updates ledger rows for risk rejection / execution skip / fill, and labels due outcomes from matched fills in the heartbeat. Added /diagnostics/trade-admission and config/trade_admission.yaml; default is enabled + shadow-only, with enforcement switches off, so live behavior is unchanged unless explicitly activated. Tests: tests/test_trade_admission.py, tests/test_strategy_candidate_flow.py, tests/test_sqlite_lite_backend.py 9 passed; py_compile clean. See docs/DECISIONS.md D196.
• Last completed task: D195 — Flow-neutral TWR implemented as primary percentage return. Operator asked for percentage return (IRR/HPR/TWR) and chose TWR. Added api.pnl_periods.time_weighted_return_from_daily_rows() and /pnl.metrics.twr: daily trading P&L = realised - fees + change_in_unrealised; any remaining NAV movement is classified as external flow, so adding/removing brokers/paper balances does not count as performance. UI performance strip now shows TWR as the primary percentage return with subtitle since <date> · net <P&L>. Current live /pnl.metrics.twr: about -0.74% since 2026-06-18, net trading P&L about -$8.9k, external flow about +$105k. Tests tests/test_twr_metrics.py 3 passed, py_compile clean, UI npm run build passed. Restarted python run.py (PID 57192). See docs/DECISIONS.md D195.
• Last completed task: D194 — Separate broker NAV plumbing from trading performance stats. Operator pointed out the dashboard's NAV Δ +8.1% since 18 Jun was caused by newly connected brokers/paper balances and must not be treated as performance. Fixed ui/src/app/redesign/dashboard.tsx: removed the NAV-history return/Sharpe/DD helper from the performance strip so broker activation/deactivation cannot manufacture returns. D195 then replaced the temporary realised-P&L card with flow-neutral TWR. Sharpe and trade drawdown stay suppressed (needs 20 days) until a proper trading-return series exists; NAV remains an account-state metric elsewhere. UI build passed (npm run build). See docs/DECISIONS.md D194.
• Last completed task: D190 — Dynamic profit harvest + FX P&L correction + no-static-threshold rule. Operator reiterated the ban on absolute trading parameters after the book failed to harvest/recycle while fixed 8%/30% gates blocked action and USDJPY P&L was displayed in quote currency. Added mandatory rule 9 to AGENTS.md and created .cursorrules: no fixed trading thresholds/magic market parameters except mathematical identities, external protocol/accounting constraints, UI/cadence settings, or documented safety bounds; touched absolute thresholds must be replaced with dynamic variables. Added core/pnl.py::unrealised_pnl_account_currency() and wired /positions, live broker position fallback, global-edge book building, and D015 portfolio bridge so FX quote-currency P&L such as USDJPY is converted to account currency. Reworked risk/profit_harvest.resolve_harvest_thresholds() to derive partial/full profit bands from realised volatility and mode coefficients only; fixed NAV profit floors and peak-lock NAV floors removed; absolute per-position profit overrides are ignored/audited. Replaced config/global_edge.yaml::capital_recycle.take_profit_pct: 0.08 with take_profit_edge_multiplier, so capital recycle banks winners only when unrealised return beats live remaining-edge estimate. Tests: tests/test_pnl_helpers.py, tests/test_profit_harvest.py, tests/test_global_edge_coordinator.py 60 passed (known AsyncMock warnings in profit-harvest harness); compile clean. Restarted python run.py (PID 14004); live /positions shows USDJPY unrealised_pnl=-1.38 USD instead of raw -223 JPY, system running with 10 brokers and capital_pct=1.0.
• Last completed task: D189 — OANDA broker adapter (tenth venue). REST forex adapter brokers/oanda/adapter.py (Bearer token, practice/live v20 API). Wired registry, broker_manager, Connect Hub, permissions (forex only), router, canonical FX symbols. Needs OANDA_API_TOKEN in .env; optional OANDA_ACCOUNT_ID. Tests tests/test_oanda_adapter.py. Restart python run.py for tenth broker badge. See docs/DECISIONS.md D189.
• Last completed task: D188 — Coinbase Advanced Trade broker adapter (ninth venue). REST crypto spot adapter brokers/coinbase/adapter.py + brokers/coinbase/auth.py (CDP ES256 JWT per request, /api/v3/brokerage). Wired registry, broker_manager, Connect Hub, permissions (crypto only), router, canonical BTC-USD products. Needs COINBASE_API_KEY (full organizations/... path) + COINBASE_API_SECRET (EC PEM, \n ok) in .env. Live probe: connect OK, 10 accounts, 117 products. Tests tests/test_coinbase_adapter.py 5 passed. Restart python run.py for ninth broker badge. See docs/DECISIONS.md D188.
• Last completed task: D187 — IG Markets broker adapter (eighth venue). REST spread-bet/CFD adapter brokers/ig/adapter.py (Version-2 session auth, demo/live base URLs, epic search, POST /positions/otc + confirms, DELETE /positions/otc closes). Wired registry, broker_manager, Connect Hub, permissions, router, canonical symbols. Needs IG_API_KEY, IG_IDENTIFIER (username), and IG_PASSWORD (IG login password) in .env. Tests tests/test_ig_adapter.py 4 passed. Restart python run.py for eighth broker badge. See docs/DECISIONS.md D187.
• Last completed task: D184 — Real logo URLs for positions and Universe instruments. Operator clarified generated avatars are not enough and asked for real logos. Clearbit is not viable in 2026 (free Logo API sunset Dec 2025), so added API-key-free real image sources with fallback: official-domain favicons via Google's favicon proxy for companies/funds, CoinCap public icon CDN for crypto, FlagCDN for FX base-currency flags, and CME favicon for futures. New core/instrument_profiles.py centralizes domain/logo mappings, logo_url_for_symbol(), with_logo(), crypto display names, company/fund domains, FX flags, futures logos. /positions profiles now call with_logo(); universe/snapshot_service.py emits logo_url from the same helper. Fixed collision where CVX-USD was incorrectly named Chevron by old catalogue matching; now crypto rows use crypto display names (CVX-USD -> Convex Finance). Verification: python -m py_compile core/instrument_profiles.py api/server.py universe/snapshot_service.py; UI npm run build passed; restarted python run.py (PID 50396); /positions emits live logo URLs for CME/FlagCDN/USCF/SSGA; /intelligence/universe sample emits Apple/Coinbase/Delta favicons, CoinCap BTC/CVX icons

Truncated - read the full file at https://github.com/emfasys-labs/mytbot/blob/30a4448720cf835c4ce5f9ba93e39645c3ce0040/AGENTS.md.