Sydnee nightly — config_drift audit — 2026-06-07 P0 findings: 0 P1 findings: 5 Risks: 9 - Area: Config Drift (Sunday theme) - Bugs found (P0 / P1): 0 / 5 (1 new, 4 carried) - Risks noted: 10 (1 new, 9 carried) Cross-check: `git log --oneline -30` and `docs/strategy_decisions.md` reviewed before filing. `docs/audit_2026-05-31_config_drift.md` (last config-drift) and `docs/audit_2026-06-07_weekly.md` reviewed. All 4 carried P1s confirmed still present in current code; 1 new P1 (DISABLE_DAILY_LOSS_LIMIT CLAUDE.md gap — carried from weekly). Recent commits: feat(tv-bxt) scale-in redesign 3b0d6ac, pre-market window relaxation 82c80de, risk flag 44aec76 — none resolved carried items. --- Full report (dev branch): https://github.com/kanex1/sydnee.signals/blob/dev/docs/audit_2026-06-07_config_drift.md Reply FROM [email protected] to [email protected] to request fixes, e.g.: "code_task on sydnee-signals-dev: apply fix for the P0 about RVOL threshold in bot.py" Sydnee Agent will propose + you APPROVE (or plain 'approve') + auto-push to dev. --- Full audit below (first 12 KB) --- # Nightly Audit 2026-06-07 — CONFIG DRIFT ## Summary - Area: Config Drift (Sunday theme) - Bugs found (P0 / P1): 0 / 5 (1 new, 4 carried) - Risks noted: 10 (1 new, 9 carried) Cross-check: `git log --oneline -30` and `docs/strategy_decisions.md` reviewed before filing. `docs/audit_2026-05-31_config_drift.md` (last config-drift) and `docs/audit_2026-06-07_weekly.md` reviewed. All 4 carried P1s confirmed still present in current code; 1 new P1 (DISABLE_DAILY_LOSS_LIMIT CLAUDE.md gap — carried from weekly). Recent commits: feat(tv-bxt) scale-in redesign 3b0d6ac, pre-market window relaxation 82c80de, risk flag 44aec76 — none resolved carried items. --- ## Findings ### BUG [P1] (new since 2026-05-31): `DISABLE_DAILY_LOSS_LIMIT` absent from CLAUDE.md — production safety warning buried in strategy_decisions.md only **Files:** `core/risk.py:106–116`; `docs/strategy_decisions.md:1059`; `CLAUDE.md` **Evidence:** ```python # core/risk.py:112 if os.environ.get("DISABLE_DAILY_LOSS_LIMIT", "").lower() not in ("1", "true", "yes"): if self.state.realized_pnl <= -self.cfg.daily_loss_limit: return False, ... ``` ``` # docs/strategy_decisions.md:1059 "Hard re-enable the gate before any real-money promotion — this flag is NOT for prod." ``` `grep "DISABLE_DAILY_LOSS_LIMIT" CLAUDE.md` → zero results. The flag was added on 2026-06-02 (commit 44aec76) and is currently active on dev (`TV_BXT_SCALE_IN=true` re-enabled on dev docker-compose per strategy_decisions.md:1009). The hard warning lives only in strategy_decisions.md, which is not the first file an operator or AI session reads when deploying. **Impact:** An operator or automated session that copies the dev docker-compose to prod (or cherry-picks env vars) silently removes the $5K daily-loss circuit breaker on real money. Config.json already sets `daily_loss_limit: 50000` (see carried P1 below), so the DLL is already effectively disabled by the value — but if the value were fixed, this flag would bypass the fix. **Fix:** Add to CLAUDE.md env-var reference (under Dev-Only Features or a new Env Vars section): ``` | DISABLE_DAILY_LOSS_LIMIT | "" (gate ON) | ⚠️ DEV ONLY — removes 5% NLV daily-loss circuit breaker. MUST be unset/absent before any real-money promotion. | ``` --- ### BUG [P1] (carried since 2026-05-24): `TV_BXT_MAX_AGE_SEC` documented but never implemented — stale Friday signals persist through the weekend **File:** `bot.py:3705` **Evidence:** ```python # bot.py:3705 — docstring for _tv_bxt_signal(): """... 0 (none). Optionally ignore stale signals via TV_BXT_MAX_AGE_SEC.""" # grep "TV_BXT_MAX_AGE_SEC" bot.py → 1 hit, docstring only. No os.environ.get call. ``` `tv_bxt_signal:{sym}` Redis key has no TTL. A Friday-close TV webhook sets the signal; the Redis key persists through the weekend. On Monday's first loop tick the bot reads the signal and may enter before any live webhook arrives, acting on a direction that is up to 65+ hours stale. **Impact:** TV-BXt model (SNDK, MU — real money since 2026-05-01) can open a position at Monday open based on a Friday signal in the opposite direction from the weekend gap. **Fix:** In `_tv_bxt_signal()` webhook writer (`bot.py:~11832–11843`), add `EX=TTL` to the Redis SET. A 14-hour TTL (50,400s) survives intraday reconnects but expires before Monday open. Alternatively, gate signal read on age using a companion `tv_bxt_signal_ts:{sym}` key. --- ### BUG [P1] (carried since 2026-05-17): `REVERSE_TIME_DECAY` uses `== "1"` — YAML-style `"true"` silently has no effect **File:** `bot.py:6019` **Evidence:** ```python # bot.py:6019 — sole check for the flag if os.environ.get("REVERSE_TIME_DECAY", "0") == "1" and _in_rth: ``` Every other boolean env flag in bot.py uses `.lower() in ("1", "true", "yes")`. `REVERSE_TIME_DECAY: "true"` in a docker-compose YAML silently disables the morning size ramp with no log or diagnostic. The only working form is the bare `"1"`. **Impact:** If the dev compose is edited to YAML boolean form, the opening-hour size ramp is silently off — affecting all non-TV-BXt entries that rely on `MORNING_SIZE_SCALE`. **Fix:** Change to `os.environ.get("REVERSE_TIME_DECAY", "").lower() in ("1", "true", "yes")`. One-line change; update default guard accordingly. --- ### BUG [P1] (carried since 2026-05-17): `config.json:daily_loss_limit: 50000` equals `capital` — DLL fires only at 100% intraday drawdown **Files:** `config.json:3,6`; `core/risk.py:112` **Evidence:** ```json { "capital": 50000, "daily_loss_limit": 50000 } ``` ```python # core/risk.py:112 if self.state.realized_pnl <= -self.cfg.daily_loss_limit: return False, ... ``` Gate fires only after a full $50,000 intraday realized loss — a 100% drawdown. The bot is live on real money. This is also the backstop the TV-BXt SHIELD exemption relies on. The RiskConfig dataclass default is `daily_loss_limit=1_500.0` (bot.py:769) — correct — but config.json overrides it to 50,000 on every load. **Impact:** No intraday loss circuit breaker is operational on real money. A runaway session (cascading fills, bad data) can lose up to the full account value before the gate fires. **Fix:** `"daily_loss_limit": 1500` (3% of $50K) in config.json. The RiskConfig default is already correct; config.json is the only thing overriding it to an unsafe value. --- ### BUG [P1] (carried since 2026-05-17): `TV_BXT_TRAIL_AGAINST` default mismatch — live engine `"1.0"` vs history-replay endpoint `"1.5"` **Files:** `bot.py:3774`; `bot.py:14344` **Evidence:** ```python # bot.py:3774 — _obv_trail_stop_hit() LIVE TRADING path: agnst = float(os.environ.get("TV_BXT_TRAIL_AGAINST", "1.0")) # bot.py:14344 — api_trade_obv_trail() HISTORY REPLAY endpoint: agnst = float(os.environ.get("TV_BXT_TRAIL_AGAINST", "1.5")) ``` When `TV_BXT_TRAIL_AGAINST` is unset (the default path on both dev and prod), the live engine uses a 1.0×ATR against-stop, but the dashboard's OBV-trail history panel reconstructs using 1.5×ATR. Every closed-trade stop timeline is wrong by 50% of the against-trail width. **Impact:** Post-trade stop-width analysis on the trade card is systematically misleading. "The stop was too tight" conclusions may be reversed — the replay shows a wider stop than what actually triggered the exit. **Fix:** Change `bot.py:14344` default from `"1.5"` to `"1.0"`. One-line fix. --- ### RISK (new): `TV_BXT_RTH_ONLY` flag name misnomer after 2026-06-03 window relaxation **File:** `bot.py:4047–4065`; `docs/strategy_decisions.md:938` **Evidence:** ```python # bot.py:4047-4048 — comment in code: # TV_BXT_RTH_ONLY=true (dev opt-in, kept name for compat): block NEW # entries outside the active session — pre-market + RTH = 04:00-16:00 ET ``` ``` # strategy_decisions.md:938: "semantics now 'block after-hours + overnight' rather than 'RTH-only'." "Reason for keeping the flag name: backward compat with the dev docker-compose env" ``` The flag was introduced 2026-06-01 (9603296) as a true RTH gate (09:30–16:00 ET). On 2026-06-03 (82c80de), the window was relaxed to 04:00–16:00 ET to allow pre-market entries. The flag name `TV_BXT_RTH_ONLY` was kept for compose backward-compat — but "RTH" is an industry-standard term meaning 09:30–16:00 ET. CLAUDE.md does not document this flag at all. **Impact:** An operator reading the flag name expects it to gate entries to RTH (09:30–16:00 ET). Enabling it on prod would actually allow pre-market entries into thin 04:00–07:00 ET liquidity (wider spreads, shallow book). The mismatch between name and behavior is documented only in strategy_decisions.md; CLAUDE.md has no mention. **Recommendation:** Either rename to `TV_BXT_SESSION_GATE` (document the old name as alias) or add a note to CLAUDE.md: `TV_BXT_RTH_ONLY=true gates NEW entries to 04:00–16:00 ET weekdays (pre-market + RTH; after-hours/overnight blocked).` --- ### RISK (carried from 2026-05-31): `TV_BXT_NLV_PCT = 50.0` sizing bypass absent from `strategy_decisions.md` **File:** `bot.py:4216`; `bot.py:7171`; `docs/strategy_decisions.md` **Evidence:** ```python # bot.py:4216 — TV-BXt single-entry sizing (also bot.py:7171): _tv_pct = float(os.environ.get("TV_BXT_NLV_PCT", "50.0")) _tv_shares = int((nlv * _tv_pct / 100.0) / price) ``` This explicitly bypasses `MAX_POSITION_PCT` (10% cap), ATR-risk sizing, and all conviction multipliers. At `$100K NLV × 50%` = $50K per trade. `grep "TV_BXT_NLV_PCT" docs/strategy_decisions.md` → zero results. Six weeks unfixed. **Recommendation:** Add a strategy_decisions.md entry for the 50% NLV sizing with hypothesis, approval date, validation criteria, and the `TV_BXT_NLV_PCT` env-var name. --- ### RISK (carried from 2026-05-25): `config.json:max_concurrent_positions: 50` — concurrent-position gate never fires **File:** `config.json:11`; `core/risk.py:114` **Evidence:** `"max_concurrent_positions": 50`. With 8 symbols in the watchlist, max realistic concurrent positions ≈ 8–12. Gate fires only at 50 — effectively disabled on real money. **Recommendation:** `"max_concurrent_positions": 5` (or 8 at most for an 8-symbol book). --- ### RISK (carried from 2026-05-25): `config.json:risk_per_trade_pct: 2.0` bypassed by env-var live sizing path **Files:** `config.json:7`; `bot.py:6780`; `bot.py:7126` **Evidence:** ```python # bot.py:6780 — live sizing (reads env vars, not config.json): sym_risk_pct = max(1.0, min(3.0, float(os.environ.get("RISK_ATR_BUDGET", "12.0")) / daily_atr_pct)) # bot.py:7126 — per-trade cap: _max_risk_pct = float(os.environ.get("MAX_RISK_PCT", "0.015")) ``` Config.json `risk_per_trade_pct: 2.0` is read only by the backtester. Live sizing reads `RISK_ATR_BUDGET` and `MAX_RISK_PCT` env vars; the config.json value is dead for live trading. **Recommendation:** Comment `risk_per_trade_pct` in config.json as backtester-only; document `MAX_RISK_PCT` (default 1.5%) as the live cap in CLAUDE.md. --- ### RISK (carried from 2026-05-24): Oracle `HEALTH_PORT` and bot `DASH_PORT` both default to `8083` **Files:** `oracle/config.py:56`; `bot.py:184`; `Dockerfile.oracle:12`; `Dockerfile:20` **Evidence:** ```python HEALTH_PORT = int(os.environ.get("ORACLE_HEALTH_PORT", "8083")) # oracle/config.py:56 DASH_PORT = int(os.environ.get("DASH_PORT", "8083")) # bot.py:184 ``` Both `Dockerfile` and `Dockerfile.oracle` declare `EXPOSE 8083`. A compose stack that maps both containers' `8083` to the host gets a silent bind conflict; whichever service starts second loses. **Recommendation:** Change `ORACLE_HEALTH_PORT` default to `8084` in `oracle/config.py:56` and `Dockerfile.oracle:12`. --- ### RISK (carried from 2026-05-24): `config.json:atr_stop_multiplier: 1.5` diverges from code fallback `2.0` **Files:** `config.json:8`; `bot.py:769` **Evidence:** ```python # bot.py:769 self.atr_stop_mult = self.cfg.get("atr_stop_multiplier", 2.0) ``` Removing or misspelling the config key silently widens stops by 33% — no log or alert. The calibrated value (1.5×) is incompatible with the code default (2.0×). **Recommendation:** Comment `config.json` to note the code fallback is 2.0; or align both to 1.5. --- ### RISK (carried from 2026-05-31): `pages.py:1448` hardcodes `"TradingView webhook (SNDK only)"` label — stale when `TV_BXT_SYMBOLS` changes **File:** `pages.py:1448` **Evidence:** ```js var nm = { ..., tv: 'TradingView webhook (SNDK only)' }[m]; ``` `TV_BXT_SYMBOLS` env var controls the actual symbol list; dashboard label is static. MU was added to TV_BXT_SYMBOLS but the label still reads "SNDK only." **Recommendation:** Pass `TV_BXT_SYMBOLS` through Jinja2 into the JS map, or update label to `"TradingView webhook"` (no symbol list). --- ### RISK (carried from 2026-05-17): `sentiment.cashtags` mismatches trading `symbols` **File:** `config.json:2,37` **Evidence:** ```json "symbols": ["WDC", "STX", "MU", "SNDK", "APP", "META", "TSLA", "PLTR"], "cashtags": ["APP", "KORU", "MU", "SNDK", "STX", "WDC"] ``` `META`, `TSLA`, `PLTR` are traded but absent from cashtag seed — sentiment bypas