Adds three new synthetic-Candidate emitter engines parallel to the
existing leader_follower scheduler hook, plus look-ahead defenses
(LookaheadViolationError + per-engine assertions). Each engine is
covered by a standalone PoC config (no PEAD/parking/idle alpha) for
isolation backtests against the midlarge or broad snapshot.
Engines:
EarningsRunup (libs/backtest/earnings_runup.py)
- Trigger: days_to_earnings ∈ [3,7] AND attention_zscore_20d ≥ 1.5
AND dollar_volume_20d_zscore ≥ 1.0 (all evaluated at T-1 close)
- Entry: T+1 next_open. Exit: -4% / +8% / max_holding_days =
days_to_earnings - buffer (forced flat by close before announcement)
- PIT calendar: PointInTimeEarningsCalendar adapter for backtest;
oracle_surprise_prefetch fallback when parquet calendar absent
- PoC verdict (configs/experiments/earnings_runup_poc_v1.json):
119 trades over 1051 days, +37.27% total return, 44.46% MDD,
SQS 45.2 (profitability=55.5, risk=23.5, robustness=50.1).
VIABLE BUT NEEDS WORK — signal exists; standalone risk profile
too aggressive for v7.356 baseline (8.8% MDD on v7.364). Path
forward: per_trade_risk reduction, VIX gate, position cap, or
integrate as PEAD sleeve adjunct (not as standalone replacement).
PeerSympathy (libs/backtest/peer_sympathy.py)
- Trigger: leader passes PEAD filter (earnings_release / guidance_update
/ material_contract) AND leader reaction_close ≥ +5% AND peer 60d
correlation ≥ 0.55 over [T-65, T-5]. Top-2 peers by correlation
from leader_follower_extra_peer_symbols_by_sector + sector ETF
holdings.
- Entry: T+1 next_open on peer. Exit: -3.5% / +6% / max_holding=3 /
peer-earnings blackout
- PoC verdict (configs/experiments/peer_sympathy_poc_v1.json):
256 trades over 1051 days, -52.92% total return, 54.47% MDD,
SQS 19.6 (profitability=0.0, risk=5.4, robustness=100.0).
DEAD. The leader's catalyst is already absorbed by T+1 next_open
— peers gap up overnight before entry. robustness=100 confirms
the negative result is not noise. Salvage paths (not implemented):
reaction_close entry, raised-guidance-only restriction.
- Note: initial run_id was 0 trades due to a select_candidates
filter mismatch (engine.event_types=['peer_sympathy'] dropping
real event_type='earnings_release' rows). The runner adapter
was patched to bypass strategy_engine filtering for leader
selection; the manual peer_sympathy_leader_event_types filter
does the gating.
VolBreakout52w (libs/backtest/vol_breakout_52w.py)
- Trigger: close_T-1 > max(high[T-252:T-2]) AND volume_T-1 ≥
2 × median_volume_20d_T-2 AND ATR_14_T-1/close ∈ [0.015, 0.06].
Entry T next_open, exit -3% / +5% / max_holding=2 / MOC.
- Honest, look-ahead-safe descendant of the retired topgainer v1-v54
family. Five layers of strict-before assertions guard the bar
provider, candidate construction, trigger evaluation, and feature
timestamps. A leaky-provider proof-by-contradiction test
demonstrates the categorical catch.
- PoC verdict (configs/experiments/vol_breakout_52w_poc_v1.json,
broad-liquid universe): 1,332 trades, -87.28% total return,
88.74% MDD, SQS 24.4 (profitability=0.0, robustness=100.0).
DEAD AND HONEST. This is the most important finding of the three
PoCs: the topgainer v1-v54 lineage's headline returns (+267%
Sharpe 13.73 in best variants) were 100% lookahead bug. With
the bug removed, the 52w-high + volume + ATR signal has no real
alpha — the lookahead-corrected -4.3% from prior memory is
confirmed and amplified to -87% on a fuller universe and longer
horizon. Future "revive topgainer" proposals can cite this run
(bt_return_max_long_v1_broad-liquid_20260509042903892342_3bb473d9)
as definitive falsification.
- Pre-open gap guard inactive (no premarket data in broad snapshot).
skip_if_no_gap_data=true; the +4% gap-fade guard would not move
the result given the magnitude.
Shared infrastructure additions:
- libs/backtest/domain.py: LookaheadViolationError class +
StrategyEngineConfig fields (11 EarningsRunup + 11 PeerSympathy
+ 13 VolBreakout52w = 35 new fields)
- apps/backtester/run.py: _BacktestAttentionZscoreAdapter,
_RunnerPeerResolver, _schedule_earnings_runup_candidates,
_schedule_peer_sympathy_candidates,
_schedule_vol_breakout_52w_candidates wired into the daily
scheduler block. PeerSympathy adapter bypasses strategy_engine
filtering on leader selection (manual filter handles gating).
Tests: 21 (EarningsRunup) + 27 (PeerSympathy) + 38 (VolBreakout52w)
= 86 new unit tests, all passing. Broader unit suite: 1392 passed,
2 pre-existing failures unrelated.
Net engine state: EarningsRunup is the only viable new engine class.
PeerSympathy and VolBreakout52w are kept in-tree as falsification
evidence, not as production engines.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The _classify_event_type mapping (2.02 → earnings_release, 7.01 →
guidance_update, 1.01 → material_contract, 1.03 → other_material_event,
8.01 → other_material_event, 5.02 → management_change) was already in
place but used naive 'in items' string matching. Upstream extractors
sometimes deliver items as 'Item 2.02' or '2.02 - Results of Operations'
(full-description form), which silently slipped through to event_type
'unknown' and were rejected by all 12 v7.356 PEAD engines.
A reparse using the patched classifier touched 9,779 historical 'unknown'
rows; only 16 actually flipped (the rest are genuinely off-vocab 8-Ks
like 9.01-only, 3.01, 5.07). The fix is therefore small in retroactive
impact, but defends against future ingestion drift.
Changes:
- libs/parser/rule_parser.py: rewrote _classify_event_type with
_normalize_item_codes (regex \\b(\\d+\\.\\d+)\\b token extractor) and
tuple-of-pairs _ITEM_TO_EVENT_TYPE mapping. Earnings_release wins
priority over management_change when 2.02 + 5.02 co-occur, consistent
with the strategy's vocabulary intent.
- tests/unit/test_rule_parser.py: 7 new regression tests covering
AMD/MNST-style 2.02+9.01, dirty 'Item 2.02' / '2.02 - Results...'
forms, and negative cases (9.01-only, 2.03, 3.01 remain unknown).
Note: a follow-up vocabulary normalizer is still needed for the
Oracle-fallback path in apps/pipeline/event_parser/main.py:140, which
writes raw oracle_event.event_type values like 'earnings_result',
'shareholder_vote', 'regulation_fd' that don't match the strategy
vocabulary. Flagged for separate ticket.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pre-market label_generator runs request future-dated price windows from
Stock Oracle, which correctly returns 404 because the data does not yet
exist. The labeler was swallowing this as label_status='unavailable' with
entry_date=None. Snapshot export then filtered these rows out, so live
PEAD trading silently lost candidates whose entry_dates fell on
later trading days (e.g., post-market 8-K filings late Friday → Monday
open entry). This explains today's missed RKLB/SNDK/AKAM/MNST/AMD/MRNA
even though their 8-Ks parsed correctly.
Changes:
- libs/labeler/label_generator.py: in 404/empty-bars path, when
entry_date >= today, preserve entry_date and mark label_status='pending'.
New log event label_price_pending_future_window distinguishes from real
data-unavailable failures (past dates still log label_price_unavailable).
- libs/export/snapshot_export.py: include 'pending' in the
label_status filter so today's not-yet-labeled events flow into the
live snapshot.
- apps/pipeline/label_generator/main.py: regeneration logic also
retries existing 'unavailable' rows whose entry_date is null or future
to recover events already mis-labeled in the DB.
- tests/unit/test_labeler.py: regression test reproducing the
RKLB/SNDK/AKAM failure mode and asserting label_status='pending' with
entry_date preserved.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
event_volume (and potentially other columns) can arrive as int64 from the
pipeline while the stored snapshot uses double, causing pa.concat_tables to
fail with "incompatible types" every run and silently fall back to a full
rebuild. _coerce_schema() casts new rows to the existing snapshot's types
before concatenation so incremental works without a full rebuild.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On the last simulation day, parking was entered at EOD close price
(when _had_event_activity_today=True) and immediately liquidated at
close by end-of-backtest cleanup → entry == exit → PnL = 0.
Fix: force all six parking entry code paths to use "open" price when
date == last_simulation_date, so entry and cleanup-close are always
different prices.
Also adds _parking_cap logic in _extend_store_to_requested_window to
cap _requested_end_date at the last date where QQQM/TQQQ/SGOV all
have Oracle close-price data, preventing the simulation from including
days where macro is incomplete and the exit fallback would fire.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
V31 research findings (2026-04-22):
- Hard gate (max_gap_zscore_20d=1.0): 45.2% vs V24 95.3% — catastrophically bad.
All three terciles are profitable; hard rejection removes positive-EV trades.
- Negative weight (weight_gap_zscore=-0.05): 90.6% DD-12.33% Sh=2.649.
Signal too weak (G2 failed at 0.181R < 0.30R threshold). G2 ≥ 0.30R
validated as reliable promotion gate: OBV-slope (G2=0.394R) passed; all
signals below 0.30R failed in backtest.
All 7 signal axes exhausted — V24 is the peak for current feature library.
domain.py: add max_gap_zscore_20d param (no-op at None default)
orb_simulator.py: add gainers_leader hard-gate (no-op at None default)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- orb_simulator.py: min_abs_gap_pct filter, premarket dollar vol filter,
rolling_loss circuit breaker, drawdown_governor, streak_sizing,
trailing_tighten_at_r, allow_doji/red_to_green breakout, abs_gap scoring
for gainers_leader, ORBSimulationState, run_orb_simulation_with_state API
- metrics.py: loss_containment_score and related metrics
- features.py: enrich_daily_bars with gap_zscore, ATR ratio, range compression
- domain.py: extended ORBStrategyParams with new fields
- cache.py: DailyBarCache with merged parquet storage and coverage metadata
- simulator.py: base simulator updates for new entry/exit mechanics
- configs/intraday: updated orb_gainers_v23.yaml with canonical params
- Added BLD to midlarge symbol snapshot
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two fixes:
1. _rebuild_daily_from_intraday_cache now filters to regular market hours
(9:30–16:00 ET) before computing OHLCV. Previously used bars[-1] which
included after-hours data, distorting prev_close for gap calculations.
Root cause of V23 regression: HIMS Aug-4 after-hours drop to $54.81
made it appear as a +0.89% gap on Aug 5 instead of the correct -12.85%
gap (from $63.45 market close), causing it to fail min_abs_gap_pct filter.
V23 with fix: +109.32%, WR 58.1%, Sharpe 3.01, DD -12.91%
2. Remove temporary debug instrumentation (HIMS/2025-08-05 trace blocks)
that was left in orb_simulator.py during regression investigation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
New mode (risk.daily_budget_reset=True) where cash_available and sizing
equity reset to initial_equity at the start of each day, regardless of
how many open positions or realized P&L exist. Unlike fixed_capital_sizing
(단리, sizing only), this also treats buying power as if no positions are
held — useful for evaluating signal quality independent of capital constraints.
- domain.py: daily_budget_reset field on RiskConfig
- run.py: _daily_budget_reset flag; _sizing_equity / _sleeve_equity_est /
_build_portfolio_state all honor the new flag
- backtest_sim.py: daily_budget_reset param threaded through
- direct_runner.py: --daily-budget-reset CLI flag
- routers/backtest.py: BacktestRequest field + cmd arg
- client.ts: BacktestParams / BacktestTask types updated
- Backtest.tsx: checkbox in form + DBR badge in task list
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- screener: switch from non-existent single-ticker endpoint to multi-ticker
/alpaca/intraday batch calls (grouped by date, chunk ≤ 75); fixes 0-trades
- cache: bump version 2→3 to invalidate stale IEX Parquet files
- oracle_client: add get_multi_intraday_bars_today() for IEX real-time feed
- paper_trader: use /alpaca/intraday/today for live sessions, /alpaca/intraday
for historical (SIP)
- intraday.py: define _BUILTIN_STRATEGIES={} to fix /api/orb/strategies import
- delete orb_p1–p10_winner + variant configs; add strategies/orb_default.yaml
(Phase 10 params) as the single registered web strategy
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add run_pre_screen() at 9:20 ET: fetch daily bars + enrichment + quality filter
before market open, narrowing universe for faster orb_detect intraday fetch
- run_orb_detection() uses cached pre-screen data when available; falls back to
full pipeline if pre_screen missed (late start, failure)
- Add _last_trading_day() helper to skip weekends/holidays for bars_end,
preventing Alpaca 502 on Mondays (today-1 = Sunday was causing failures)
- Fix Oracle client chunk_size 300→75: Alpaca rejects 100+ ticker URL requests
- Add pre_screen event to build_schedule() at 9:20 ET and dispatch in _run_trading()
- run_session_now() runs pre_screen before orb_detect for efficiency
- Add ORB daemon, engine, models, state, screener, and intraday strategy configs
- Add intraday library (libs/intraday/) and web routes for ORB/intraday trading
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- libs/oracle_client/alpaca.py: Added get_multi_daily_bars() and
get_multi_intraday_bars() helpers that call Oracle's /api/v1/price/data
and /api/v1/alpaca/intraday endpoints respectively. Oracle handles
symbol normalization (e.g. BF-B → BF.B) internally, so symbols like
BF-B no longer crash the screening chunk.
- apps/paper_trader/alpaca_broker.py: get_bars() and get_intraday_bars()
now use the new Oracle client helpers instead of the Alpaca SDK
StockBarsRequest, eliminating direct Alpaca bar API calls from broker.
- apps/orb_trader/engine.py: Removed per-symbol BF-B workaround (now
unnecessary since Oracle normalizes the symbol server-side); kept outer
try/except for chunk-level resilience.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Bug #2 (paper trader): lookback entries sized using historical entry_price_est
but filled at current market price, causing cash overdraft. Fix: override
entry_price_est with get_latest_bars() close before entering _process_entries.
Bug #3 (paper trader + backtester): paper trader was missing the per-candidate
MHD expiration check that the backtester already had. Also adds
lookback_min_remaining_days (default 3) to reject candidates with too little
holding time remaining — prevents entering a position the day before forced exit.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Delete v7.360-v7.363 experiment configs (rotation/momentum tests)
- Remove _schedule_momentum_breakout_candidates() from backtester run.py
- Remove MomentumBreakoutConfig from domain.py
- Delete momentum_calendar.py, momentum_screener.py, build_momentum_calendar.py
- Delete data/momentum_calendar/ parquet data
Valid period performance was -31.36% vs +152.4% baseline — sleeve is not viable
without walk-forward validation. Abandoning for now.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- enrich_tier2: prefetch price bars (parallel ThreadPool) and short ratio
(single batch DB query) instead of per-row HTTP/DB calls (~20min → ~2min)
- canonical_snapshots: add PYTHONUNBUFFERED=1 to enrichment subprocesses
so progress output is visible in real time
- backtest_sim: use incremental_update_canonical_snapshot when existing
snapshot is present, falling back to full rebuild only when needed
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a backtest starts mid-stream (via --start), events that fired
before the start date but are still within their max_holding_days
window can now be entered on the first simulation day.
- Add `lookback_entry_enabled: bool = False` to ExecutionConfig
- On first sim day, _collect_lookback_candidates() gathers pre-start
events, runs them through the same select_candidates() pipeline,
and injects them before normal candidates
- Entry fills at the first day's open price; gap-cap check is skipped
since the event is multi-days old
- days_held is initialized to the elapsed trading days so TIME exits
fire at the correct time relative to the original event date
- Store slice is extended backward by max_mhd calendar buffer so
pre-start rows survive slice_by_date_range when feature is enabled
- Enabled in return_max_long_v7.119 for testing
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Backtester (run.py):
- cash_available = (self._cash + parking_value) * multiplier caused trades to be
approved even when self._cash ≈ 0 (all money in SGOV/QQQ). Trades executed
by deducting from self._cash → negative cash (phantom money).
- Fix: after simulate_entry, if self._cash < actual trade cost and parking exists,
call _liquidate_parking_for_cash(shortfall) before deducting from cash.
- Verified: 2022-2026 backtest with qqqm_low_dd shows 0 cash_negative events.
Live engine (engine.py):
- Add _parking_liquidate_for_event(): frees parking cash to fund event entries.
SGOV (virtual) reduces entry_value in DB; QQQM/QQQ sells real shares via broker.
- Both entry loops (engines mode + flat/reaction_close mode) now attempt parking
liquidation when plan.skip_reason == "insufficient_cash" before giving up.
Also includes prior session work (accumulated since last commit):
- 6 novel parking gate signals: VRP, Market Temperature, Hurst exponent, Rolling
Kurtosis, Return Autocorrelation, SPY-QQQ Correlation (composite risk score v2)
- QQQM parking symbol support (lower expense ratio vs QQQ)
- Snapshot auto-refresh + bar extension cache (pickle) to avoid 10-min re-fetches
- Bar extension clamps to last market-closed date (ET 4PM check)
- fithia2 refresh command; --no-refresh flag for paper backtest
- Paper backtest macro extension beyond last event date (parking-only periods)
- parking_state DB schema: 7 new columns (peak_price, gate_in_sgov,
committed_target, pending_target, pending_days, sgov_entry_value, sold_today)
- Live engine: target confirmation (2-day), top-up drawdown gate, trailing stop,
SGOV interest accrual, full 6-signal gate evaluation
- New PARKING_PRESETS: qqqm_low_dd, composite_v2, vv_24_vrp8, vt_24_t13, etc.
- Web GUI / CLI result parity fix (Oracle URL via get_settings().stock_oracle_url)
- Force-close uses last_exec_date (has bar data); parking liquidates at last_date
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Builds a full synthetic market data pipeline to test strategies against
12 diverse market regimes (bull/bear/crash/chop/rotation/liquidity drought)
that may not exist in historical data. Computes Regime Robustness Score (RRS)
to detect overfitting and environment-specific fragility.
- libs/backtest/scenarios/: price_gen, macro_gen, event_gen, coupling,
store_builder, scenarios (12 pre-built), robustness (RRS)
- apps/scenario/cli.py: `fithia2 scenario-test` with Rich output
- apps/tracker/cli.py: scenario-test command routing
- tests/: 83 unit tests across 3 new test files
- docs/scenario_test.md: usage guide and result interpretation
- docs/research_workflow_and_handoff.md: Step 5.5 scenario test added
Fix: no_signal scenario uses drift=0% (was +10%) for fair signal integrity scoring.
Fix: synthetic candidates now carry macro_vix/macro_hy_spread from macro_by_date
to pass selector engine filters.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Additional tracker/leaderboard updates, overlay leaderboard, and
documentation improvements.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces per-experiment rglob with single-pass manifest/metrics indexing
and adds lru_cache. Removes rarely-used commands from help display.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extends selector with new scoring model support, adds execution
enhancements, and improves snapshot store loading and split handling.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds earnings surprise extraction to parser/features/labeler pipeline,
improves filing fetcher robustness, and extends snapshot export with
new field support.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bounce engine (buy negative reaction, bet on mean reversion) could not
execute: system architecture ties scoring to single model per backtest,
and selector/store indexes are optimized for positive-reaction PEAD.
Negative-reaction candidates get score=0 from PEAD scoring, blocking
engine selection regardless of engine-level threshold overrides.
Implementing bounce trades requires: dual scoring model support,
selector changes for negative-reaction candidate routing, and
store indexing changes. Deferred to future refactor.
Current best CW return: 293.2% (v6new.255)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Data analysis revealed OBV Q1 (distribution) has 56.4% WR vs Q5 51.2% —
contrarian signal confirmed. Previous OBV bonus was applied in wrong
direction. Corrected with v15 scoring models.
Best result: v6new.185 (entropy + risk 0.058) CW 274.4% but SQS 72.2,
still below v6new.122 (72.4). WFV/robustness offsets CW gains.
v6new.122 confirmed as optimal under current SQS v4 formula.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously only used rows[0].keys() — columns present in later rows
(like earnings_surprise_pct from sparse features) were silently dropped.
Now collects all unique keys across all rows.
YoY earnings surprise tested: WR spread only 2.5pp (55.2% vs 52.7%).
Not actionable — YoY growth != analyst consensus surprise.
v6new.30 remains the framework optimum.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New data source integration:
- EarningsSurpriseService: GET /api/v1/earnings/surprise/{symbol}
Returns actual vs estimated EPS with surprise_percentage
- Feature builder: creates earnings_surprise_v1 snapshots for earnings events
- Backfill script runs for existing 1,273 tickers (Alpha Vantage rate limited)
New scoring (v11):
- Small beat (0-3% surprise): +10% bonus (82.4% WR in sample)
- Medium beat (3-8%): +5% bonus
- Big beat (>8%): no bonus (already priced in)
- Miss (<=0%): -5% penalty
Signal validation (n=66 sample):
Small beat: 82.4% WR, +1.79% mean 5d return
Big beat: 54.8% WR, +0.47%
Miss: 55.6% WR, -0.10%
Backfill running (~4 hours). Experiment pending data completion.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
FilledTrade now carries event_type and score from the Candidate.
These fields are written to trade_blotter.parquet and displayed in
paper backtest trade logs.
Previously score showed as 0.00 for all trades because the field
wasn't propagated from Candidate → FilledTrade → Parquet.
Score=0.00 is valid for trades from engines with score_threshold_override=0.0
(e.g. guidance_unknown_orderly) where engine gates, not score, determine entry.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase 1-4 of engine unification to eliminate research/live divergence.
Phase 1 — Scoring (event_detector.py):
EventDetector now uses config's scoring_model (v5/v9 etc.) when
event_v1 features are present (parse_confidence_overall not null).
Falls back to compute_entry_score only for incomplete events.
Phase 2 — Execution config (execution.py):
Extracted build_effective_execution_config() as shared function.
BacktestRunner delegates to it. PaperTradingEngine can now use
identical per-engine overrides, adaptive exit, tiered targets.
Phase 3 — Attention filtering (attention.py):
New AttentionFilterService class extracted from BacktestRunner.
Provides: engine_requires_attention, apply_filters, rescoring.
BacktestRunner now delegates to this service.
PaperTradingEngine can import and use the same service.
Phase 4 — Gap cap (execution.py):
check_next_open_gap_cap() shared function for next-open gap rejection.
All 450 unit tests pass. Paper backtest verified working.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
I Luk Kim