Skip to content

CHANGELOG.md

Latest commit

 

History

History
235 lines (187 loc) · 27.2 KB

CHANGELOG.md

File metadata and controls

235 lines (187 loc) · 27.2 KB

Changelog

Feature — 2026-04-30 (v3: strategy framework — config-driven council)

Implements specs/v3.md phases 1–6. Users can now define a complete trading strategy in a single YAML file without writing Python. Test suite: 305 → 344 passing.

Added

  • src/strategies/config.py — typed StrategyConfig Pydantic model: AgentConfig (name, model, prompt, weight, is_veto), DeliberationConfig (optional), ScoringConfig (confidence_floor, trade_threshold), RiskRules (stop_loss/take_profit SLTPRule + max_position_pct), ValidationConfig (max_drawdown_pct, atr_spike_multiplier, min_news_count); model_validator enforces weight-sum=1.0 and at-most-one veto agent
  • src/strategies/loader.py — load_strategy(path): loads YAML, validates schema, and checks all prompt files exist on disk; prompt paths are resolved relative to project root
  • src/strategies/context.py — render_context_block(ctx): renders the full MarketContext as a structured text block (PRICE / INDICATORS / NEWS / SENTIMENT / ON-CHAIN / MACRO / PORTFOLIO sections); single shared input for all agents in a generic run
  • src/strategies/runner.py — run_strategy_council(ctx, config, client): config-driven generic runner; runs all agents in parallel; short-circuits to HOLD on veto; calls compute_generic_score; optionally runs deliberation; computes SL/TP via compute_sl_tp; returns GenericCouncilResult
  • src/strategies/scoring.py — compute_generic_score(outputs, config): config-driven confidence-weighted signed-score; weights normalised at call time; confidence_floor and trade_threshold sourced from ScoringConfig; score_to_position_size_pct(score, config) scales size from |score|
  • src/strategies/risk_rules.py — compute_sl_tp(rules, ctx, entry_price): computes SL/TP prices from RiskRules; supports atr_multiple (price ± N×ATR-14), fixed_pct (price × (1 ± pct)), and none (0.0, 0.0); SL clamped > 0
  • strategies/default.yaml — current v2 council expressed as a strategy file (technical 0.40 / sentiment 0.25 / fundamental 0.35 / risk veto; deliberation enabled; scoring thresholds and validation thresholds matching current hardcoded constants)
  • strategies/example.yaml — fully-commented example strategy (momentum + macro + risk_guard veto); shows every config option
  • prompts/deliberation_generic_v1.txt — built-in deliberation prompt for generic runs; requests a 2–4 sentence narrative synthesis
  • prompts/agent_template.txt — starting-point prompt for directional agents with annotated output schema and confidence calibration guide
  • prompts/veto_agent_template.txt — starting-point prompt for veto agents with VetoAgentOutput schema guidance
  • prompts/risk_manager_veto_v1.txt — purpose-built veto prompt for default.yaml's risk agent (cleaner than repurposing risk_manager_v2.txt)
  • tests/test_strategy_loader.py (12 tests) — valid configs, weight validation, veto count, missing fields, missing prompt files, default.yaml and example.yaml smoke loads
  • tests/test_strategy_scoring.py (16 tests) — all directions, confidence floor exclusion, threshold gating, weighted score formula, SL/TP for all three rule types
  • tests/test_strategy_runner.py (8 tests) — BUY/HOLD/SELL signals, veto short-circuit, veto=false passthrough, SL/TP on result, conviction derivation, deliberation failure non-fatal, no-veto-agent skips veto call
  • tests/test_strategy_parity.py (3 tests) — BUY/HOLD/veto scenarios confirm generic and legacy scorers agree on identical inputs

Changed

  • src/models.py — added GenericAgentOutput, VetoAgentOutput, GenericCouncilOutputs, GenericDeliberationOutput (additive; all existing models unchanged)
  • src/validation.py — check_drawdown_halt and check_volatility_halt now accept max_drawdown_pct and atr_spike_multiplier params (backward-compatible defaults match previous constants); validate_context accepts both params and passes them through
  • src/data/assembler.py — assemble_context accepts max_drawdown_pct and atr_spike_multiplier params and passes them to validate_context
  • src/main.py — added --strategy PATH CLI argument; when provided, calls run_strategy_council instead of legacy run_council; validation params derived from strategy config
  • src/backtest/signals.py — generate_signals accepts strategy_config param; routes to generic runner when provided; agent logs written per-agent-name for generic runs
  • src/backtest/engine.py — run_backtest accepts strategy_config param; _parse_args adds --strategy PATH
  • pyproject.toml — added pyyaml>=6.0 as explicit dependency (was transitively available; now declared)

Bugfix — 2026-04-27 (risk manager stop-loss above entry price)

Fixed

  • prompts/risk_manager_v2.txt — rewritten to prevent the model from anchoring recommended_stop_loss to historical BTC ATH levels (~$70–74k) instead of computing downward from the current price. Key changes: explicitly states the system is long-only (removed the "above entry for SELL" clause that confused the model); requires recommended_stop_loss MUST be LESS THAN the current price; provides the placement formula as a concrete equation (stop = current_price − multiplier × ATR-14); adds a worked example with actual numbers ($23k BTC / $1,500 ATR-14); adds "DO NOT use historical ATH prices as reference points"
  • src/agents/risk.py — PROMPT_FILE flipped from risk_manager_v1.txt to risk_manager_v2.txt
  • src/backtest/signals.py — new _sanitize_stop_levels(close_price, sl_pct, tp_pct, atr_14): code-level guard that corrects invalid stop/target ratios regardless of LLM output. If sl_pct ≥ 1.0 (stop above entry) or ≤ 0, replaces with 2×ATR below entry floored at 80% of entry. If tp_pct ≤ 1.0 (target below entry), replaces with 3×ATR above entry (preserves ≥1.5 R:R). Applied immediately after the raw ratio computation in the signal loop.

Tests

  • tests/test_backtest.py — new TestSanitizeStopLevels (14 cases): ATH-anchoring bug reproduction, ATR fallback formula verification, 80% floor enforcement, valid values pass through unchanged, tp-below-entry correction, R:R ≥ 1.5 check after sanitization, zero-ATR and zero-close edge cases. Test suite: 291 → 305 passing.

Feature — 2026-04-22 (v2: restore agent sight + replace rigid consensus)

Implements specs/v2.md phases 1–4. Phase 5 (tuning + holdout backtest) is pending real API runs. Test suite: 235 → 283 passing.

Added

  • src/data/news_historical.py — fetch_news_for_date(as_of): GDELT DOC 2.0 artlist query filtered to a source allowlist (Reuters, Bloomberg, CoinDesk, WSJ, FT, CoinTelegraph, Forbes, CNBC, The Block, Decrypt); strict as_of - 24h publish-lag cutoff; disk cache at tmp/cache/gdelt/YYYY-MM-DD.json; returns [] on HTTP error so the caller can fall back to the stub
  • src/data/onchain_historical.py — fetch_onchain_for_date(as_of): CoinMetrics Community API (community-api.coinmetrics.io/v4/timeseries/asset-metrics); preserves existing OnchainData field names but fills them with documented proxies — MVRV (CapMrktCurUSD / CapRealUSD) as sopr, 7-day transfer-volume momentum as exchange_net_flow_btc, active-address count / 1000 as whale_transactions_24h; slices strictly date < as_of.date(); disk cache at tmp/cache/coinmetrics/
  • src/data/macro.py — fetch_macro_for_date(as_of): yfinance bundle for DX-Y.NYB, ^VIX, ^GSPC, ^TNX (divides by 10 to correct the yfinance ×10 quote); computes macro_bias from VIX band, DXY 20-session trend, and SPX 20-session trend; falls back to macro_bias=neutral on any empty frame
  • src/models.py — new MacroData model (vix, dxy, spx_20d_change_pct, tnx_yield_pct, macro_bias); new optional macro: MacroData | None = None field on MarketContext; new score: float = 0.0 field on DeliberationOutput
  • src/agents/scoring.py — compute_signed_score(outputs): confidence-weighted signed-score over the three directional agents (weights 0.40 tech / 0.25 sentiment / 0.35 fundamental); confidence floor 30 (agents below abstain rather than dilute); threshold 0.25 to trade; risk-veto overrides unconditionally. score_to_position_size_pct(score) derives size from |score|, capped at 20%
  • prompts/technical_analyst_v2.txt, sentiment_analyst_v2.txt, fundamental_analyst_v2.txt, deliberation_v2.txt — v1 prompts retained in-tree; v2 prompts instruct directional agents to use the full 0–100 confidence range and be more assertive in clear regimes; deliberation rewritten to produce narrative only (scoring happens in Python); fundamental prompt documents the MVRV-as-SOPR proxy so the agent reasons about it correctly
  • tests/test_news_historical.py, tests/test_onchain_historical.py, tests/test_macro.py — full coverage with mocked HTTP and yfinance; tests lookahead boundaries, cache round-trips, fallback paths
  • tests/test_agents.py — TestComputeSignedScore (8 cases: all-BUY / all-SELL / veto override / all-HOLD / confidence floor / split directions / sub-threshold / all-abstain) and TestScoreToPositionSize (4 cases); new test_scored_signal_overrides_llm_deliberation verifying the runner discards the LLM's final_signal and substitutes the deterministic score

Changed

  • src/validation.py — validate_news_count and validate_context take min_news_items: int = 10; backtest callers pass 5 (GDELT allowlist can be thin on some dates)
  • src/data/assembler.py — new macro_override: MacroData | None parameter; new min_news_items: int = 10 parameter plumbed through to validate_context
  • src/agents/deliberation.py — PROMPT_FILE flipped to deliberation_v2.txt
  • src/agents/technical.py, sentiment.py, fundamental.py — PROMPT_FILE flipped to _v2.txt variants
  • src/agents/fundamental.py — user-message builder now includes a Macro backdrop: block when ctx.macro is populated; on-chain field labels re-worded to match the v2 proxy semantics
  • src/agents/runner.py — imports compute_signed_score; after deliberation, calls llm_deliberation.model_copy(update={"final_signal": scored_signal, "score": score}) so the deterministic scorer owns the final signal while the LLM still supplies narrative; risk-veto short-circuit preserved (skips the deliberation LLM call entirely)
  • src/backtest/signals.py — replaces NEUTRAL_NEWS_STUB with fetch_news_for_date (stub fallback when < 5 items); replaces NEUTRAL_ONCHAIN_STUB with fetch_onchain_for_date; threads fetch_macro_for_date into macro_override; sizes positions via min(score_to_position_size_pct(score), risk.approved_position_size_pct); SIGNAL_COLUMNS gains a score column (surfaced in the CSV and the DataFrame)
  • README.md — new v2 phase table; lists each module created

Update — 2026-04-20 (per-cycle agent response logs)

Added

  • src/backtest/signals.py — _append_agent_log(path, date, data): appends a JSON Lines entry {"date": "...", "output": {...}} to an agent file and closes immediately; generate_signals accepts a new agent_log_dir: Path | None parameter and calls this after each successful cycle, writing technical_analyst.json, sentiment_analyst.json, fundamental_analyst.json, risk_manager.json, and deliberation.json under tmp/YYYYMMDD_HHMMSS/agents/
  • src/backtest/engine.py — derives agent_log_dir = run_dir / "agents" and passes it to generate_signals; end-of-run summary now prints the agents/ folder

Fixed

  • src/models.py — RiskManagerOutput: added field_validator on recommended_stop_loss, recommended_take_profit, and risk_reward_ratio to strip commas before float parsing, fixing AgentError when the model returns formatted numbers like '71,001.50'

Update — 2026-04-20 (backtest run output to tmp/, incremental CSV)

Changed

  • src/backtest/engine.py — _setup_run_dir default base changed from runs/ to tmp/; timestamped run folders now live under tmp/YYYYMMDD_HHMMSS/
  • src/backtest/engine.py — run_backtest now resolves csv_path before calling generate_signals and passes it in; the post-hoc signals_df.to_csv() call is removed
  • src/backtest/signals.py — generate_signals accepts a new csv_path: Path | None parameter; when set, opens the file before the signal loop, writes the header row, then flushes one CSV row per completed cycle so the file is preserved if the run is killed mid-way

Feature — 2026-04-20 (run output folder)

Added

  • src/backtest/engine.py — _setup_run_dir(base): creates runs/YYYYMMDD_HHMMSS/ on each invocation; adds a logging.StreamHandler pointing at run.log inside that directory (StreamHandler flushes after every record); tees sys.stdout and sys.stderr through _TeeStream so all print() output is also captured; uses buffering=1 (line-buffered) so every newline triggers an OS write and output is preserved if the process is killed
  • _TeeStream: lightweight wrapper that mirrors writes to two streams simultaneously
  • run_backtest now accepts an optional run_dir: Path parameter; when provided, the signals CSV is written into that directory instead of the working directory
  • CSV filename changed from backtest_signals_<full-datetime-with-tz>.csv (contained colons, broke on some OS) to backtest_signals_YYYYMMDD_YYYYMMDD.csv (start/end dates)
  • End-of-run summary prints the run folder path, CSV filename, and log filename

Fixed

  • run_backtest previously called datetime.now() twice for the CSV filename, producing a different timestamp in the returned csv_path than the one used when writing the file

Bugfix — 2026-04-20 (structured JSON output via tool use)

Fixed

  • src/agents/base.py — replaced free-form text generation + markdown-fence stripping with Anthropic tool use (tools + tool_choice={"type": "tool"}); the API now returns a tool_use block whose .input is a guaranteed parsed dict, eliminating the non-JSON output errors caused by models wrapping responses in ```json ``` fences with trailing prose
  • Removed _extract_json helper and import json (no longer needed)
  • AgentError now raised with "tool_use block" message when the response contains no tool-use block (e.g. unexpected end_turn)

Tests

  • tests/test_agents.py — updated _mock_response/_mock_client to return a mock with type="tool_use" and .input=payload instead of a text block; replaced test_strips_markdown_fences and test_raises_on_non_json (scenarios no longer possible) with test_raises_on_missing_tool_use_block; removed unused import json; 235 tests passing

Bugfix — 2026-04-06 (fractional trading + dotenv)

Fixed

  • src/backtest/engine.py — switched from Backtest to FractionalBacktest (backtesting.lib); BTC trades in fractional units so any --capital value now works without orders being silently canceled
  • src/backtest/signals.py — replaced sl_price/tp_price (raw USD) with sl_pct/tp_pct (ratio relative to close, e.g. 0.95 = 5% stop below close); FractionalBacktest scales OHLCV prices internally (to satoshi units) so absolute USD prices caused constraint failures (SL < execution_price < TP always failed against the scaled price)
  • src/backtest/engine.py strategy — next() now computes sl = Close * sl_pct and tp = Close * tp_pct, which scales correctly under any price transformation
  • tests/test_backtest.py — updated all test fixtures and _make_signals_df to use sl_pct/tp_pct; _run_strategy cash reverted to $10,000 (fractional support makes large cash workaround unnecessary)

Bugfix — 2026-04-06

Fixed

  • src/main.py, src/backtest/engine.py — added load_dotenv() call at startup; .env file was never being read, so ANTHROPIC_API_KEY and all other env vars were silently missing at runtime

v1 Phase 6 (Historical Data Source) — 2026-04-06

Changed

  • src/backtest/data.py — fetch_full_ohlcv rewritten to use yfinance.download("BTC-USD") instead of CCXT; removes exchange parameter and pagination loop; handles MultiIndex columns (yfinance 0.2+); UTC-localizes index
  • src/backtest/engine.py — run_backtest no longer creates a Kraken exchange for data fetching; get_exchange import removed
  • pyproject.toml — added yfinance>=0.2.0

Fixed

  • Kraken's public OHLC API only serves the most recent ~720 daily candles regardless of the since parameter; requesting data older than that (e.g. --start 2024-01-01 from April 2026) returned 0 candles. yfinance provides BTC-USD history back to 2014 with no API key.

Tests

  • tests/test_backtest.py — replaced CCXT-based fetch_full_ohlcv tests with yfinance mocks; removed get_exchange mock from run_backtest tests

v1 Phase 5 (Exchange Migration) — 2026-04-06 (revised)

Fixed

  • src/data/price.py — corrected CCXT symbol from "XBT/USD" to "BTC/USD"; CCXT normalizes Kraken's internal XBT ticker to BTC in its unified API, so XBT/USD raises BadSymbol at runtime

v1 Phase 5 (Exchange Migration) — 2026-04-06

Changed

  • src/data/price.py — replaced ccxt.binance with ccxt.kraken; SYMBOL changed from "BTC/USDT" to "XBT/USD"; removed Binance testnet URL logic; credentials now read from KRAKEN_API_KEY / KRAKEN_SECRET; sandbox param retained for interface compat but ignored (Kraken has no spot sandbox)
  • src/execution/router.py — SYMBOL = "XBT/USD"
  • src/execution/state.py — default symbol updated to "XBT/USD"
  • src/execution/orders.py — docstring updated (Binance → Kraken)
  • src/db/store.py — default symbol in get_open_trade updated to "XBT/USD"
  • src/main.py — SYMBOL = "XBT/USD"; env var docs reflect Kraken credentials; paper trading note updated (DRY_RUN=1 replaces testnet)
  • src/data/assembler.py — docstring updated
  • .env.example — replaced BINANCE_TESTNET_API_KEY/SECRET and BINANCE_API_KEY/SECRET with KRAKEN_API_KEY / KRAKEN_SECRET
  • tests/test_execution.py, tests/test_db.py, tests/test_reporting.py — all "BTC/USDT" references updated to "XBT/USD"

Why: Binance is geo-restricted in the United States. Kraken is fully US-accessible, free for market data, and provides BTC/USD history back to 2013 via CCXT public endpoints.

Paper trading: Kraken has no spot sandbox. Paper trading uses DRY_RUN=1 (signals logged, no orders placed). Live trading requires KRAKEN_API_KEY + KRAKEN_SECRET.

v1 Phase 1–4 (Backtesting) — 2026-04-06

Added

  • src/backtest/__init__.py — new backtest package
  • src/backtest/data.py — fetch_full_ohlcv (paginated CCXT download), slice_ohlcv (strict no-lookahead enforcement), fetch_historical_sentiment (Alternative.me Fear & Greed history), get_sentiment_for_date (nearest-earlier fallback), NEUTRAL_NEWS_STUB (15 generic items), NEUTRAL_ONCHAIN_STUB
  • src/backtest/signals.py — generate_signals: async LLM council loop over each historical date; _PortfolioTracker dataclass (cash/position/drawdown tracking); SIGNAL_COLUMNS constant; graceful HOLD fallback on AgentError
  • src/backtest/engine.py — LLMCouncilStrategy (backtesting.py Strategy replaying pre-computed signals); run_backtest async runner (fetch → signals → simulate → formatted stats dict); CLI via python -m src.backtest.engine --start --end --capital
  • tests/test_backtest.py — 26 tests: TestBacktestData (slice no-lookahead, pagination, sentiment fallbacks, news stub), TestGenerateSignals (column schema, date count, lookahead guard, error fallback, portfolio tracker), TestRunBacktest (strategy execution, stop-loss trigger, stats keys)
  • backtesting>=0.3.3 added to pyproject.toml

Changed

  • src/data/price.py — added since: int | None = None to fetch_ohlcv for historical backfill
  • src/data/assembler.py — added timestamp, news_override, sentiment_override, onchain_override, skip_freshness optional params to assemble_context; backward compatible (all default to None/False)
  • src/validation.py — added skip_freshness: bool = False to validate_context; when True, skips price freshness check (used by backtest engine)
  • tests/test_assembler.py — updated to cover new override parameters
  • tests/test_validation.py — updated to cover skip_freshness flag

Fixed

  • src/backtest/data.py — fetch_full_ohlcv now correctly handles timezone-aware end_date arguments using tz_localize/tz_convert instead of pd.Timestamp(dt, tz=...) which raises on already-aware datetimes
  • tests/test_backtest.py — _run_strategy default cash raised from $10k to $1M so 20% position sizing can purchase at least one whole BTC unit at realistic price levels ($40k–$50k) without backtesting.py silently canceling orders

v0 Phase 4 — 2026-04-05

Added

  • src/db/schema.py — added weekly_summaries table (window_start, window_end, summary, metrics_json)
  • src/db/store.py — added get_closed_trades_in_window, get_reflections_for_trades, get_cycles_in_window, save_weekly_summary, get_weekly_summaries
  • src/reporting/__init__.py — new reporting package
  • src/reporting/metrics.py — compute_sharpe_ratio (annualised, sqrt(252)), compute_max_drawdown (equity-curve walk), compute_expectancy (avg_win × win_rate − avg_loss × loss_rate), compute_council_consistency (unanimous rate, confidence spread, veto rate), compute_all_metrics (all four, with target-met flags)
  • src/reporting/weekly.py — run_weekly_summary: loads week's trades + reflections + cycles, computes metrics, calls claude-sonnet-4-6 with weekly_summary_v1.txt, persists summary to DB; runnable as python -m src.reporting.weekly
  • prompts/weekly_summary_v1.txt — system prompt for the weekly summary agent (5 structured sections: performance verdict, pattern analysis, agent accuracy ranking, prompt refinement recommendations, next-period watchpoints)
  • src/main.py — TRADING_MODE wiring (paper = testnet, live = mainnet with COUNCIL_LIVE_CONFIRMED=1 safety gate); --weekly CLI flag; weekly APScheduler job (Sunday 08:00 UTC)
  • .env.example — documented TRADING_MODE, COUNCIL_LIVE_CONFIRMED, BINANCE_API_KEY/SECRET, COUNCIL_DB_PATH, DRY_RUN
  • tests/test_reporting.py — 49 tests: Sharpe, max drawdown, expectancy, consistency, all-metrics, weekly message builder, run_weekly_summary, store windowed queries

v0 Phase 3 — 2026-04-05

Added

  • src/db/schema.py — SQLAlchemy table definitions (cycles, trades, portfolio_state, reflections)
  • src/db/store.py — CRUD operations: log_cycle, open_trade, close_trade, get_open_trade, get_portfolio_state, update_portfolio_state, save_reflection, get_reflection
  • src/execution/orders.py — place_market_buy, place_market_sell, place_stop_loss, cancel_order with retry logic (max 3 attempts, exponential back-off)
  • src/execution/state.py — build_portfolio_dict: loads live portfolio state from DB for context assembly
  • src/execution/router.py — route_signal: translates CouncilResult → exchange orders + DB updates; reconcile_open_position: detects stop-outs at cycle start
  • src/agents/reflection.py — post-trade reflection agent (claude-sonnet-4-6)
  • prompts/reflection_v1.txt — system prompt for the reflection agent
  • src/main.py — full cycle entry point; supports one-shot and --schedule (APScheduler daily at 00:05 UTC); DRY_RUN=1 skips order execution
  • tests/test_db.py — 16 tests for DB schema and CRUD
  • tests/test_execution.py — 22 tests for orders, state, and router
  • tests/test_reflection.py — 8 tests for the reflection agent
  • sqlalchemy>=2.0.0 and apscheduler>=3.10.0 added to dependencies

v0 Phase 2 — Agent Framework

Added

  • src/models.py — extended with all agent output models: TechnicalAnalystOutput, SentimentAnalystOutput, FundamentalAnalystOutput, RiskManagerOutput, CouncilOutputs, DeliberationOutput, AgentWeights
  • src/agents/base.py — call_agent() helper: loads prompt file, calls Anthropic API at temperature=0, strips markdown fences from response, validates JSON against Pydantic schema; raises AgentError on any failure
  • src/agents/technical.py — Technical Analyst agent (claude-haiku-4-5); analyses RSI momentum, MACD crosses, Bollinger Band position, EMA alignment, ATR volatility
  • src/agents/sentiment.py — Sentiment Analyst agent (claude-haiku-4-5); analyses news tone, fear/greed index, social volume, dominant narrative
  • src/agents/fundamental.py — Fundamental Analyst agent (claude-haiku-4-5); analyses exchange net flows, whale transactions, SOPR, macro bias
  • src/agents/risk.py — Risk Manager agent (claude-sonnet-4-6); evaluates portfolio risk, outputs position size, stop-loss, take-profit, R:R; has veto authority
  • src/agents/deliberation.py — Deliberation Chair agent (claude-sonnet-4-6); synthesises all four agent outputs into final BUY/SELL/HOLD signal with conviction level
  • src/agents/runner.py — run_council(): runs all four agents in parallel via asyncio.gather(); short-circuits to HOLD immediately if risk manager vetoes (skips deliberation); returns CouncilResult
  • prompts/technical_analyst_v1.txt — system prompt for Technical Analyst
  • prompts/sentiment_analyst_v1.txt — system prompt for Sentiment Analyst
  • prompts/fundamental_analyst_v1.txt — system prompt for Fundamental Analyst
  • prompts/risk_manager_v1.txt — system prompt for Risk Manager
  • prompts/deliberation_v1.txt — system prompt for Deliberation Chair (includes consensus rules: all-3-agree ≥70% → full size; 2-of-3 ≥60% → 50% size; else HOLD)
  • tests/test_agents.py — 38 tests: call_agent base utility, individual agent schema validation, deliberation consensus logic, runner veto short-circuit, prompt file existence checks

v0 Phase 1 — Data Pipeline + Tier-0 Validation

Added

  • src/models.py — Pydantic v2 models for the full context schema: PriceData, IndicatorData, NewsItem, SentimentData, OnchainData, PortfolioData, MarketContext; literal types for Direction, MACDSignal, BBPosition, Regime, etc.
  • src/data/price.py — CCXT Binance testnet OHLCV fetcher (fetch_ohlcv, build_price_data); raises ValueError if fewer than 200 candles returned
  • src/data/indicators.py — compute_indicators(): RSI-14, MACD cross detection, Bollinger Band position, EMA 20/50/200, ATR-14, ATR-30 rolling average, regime classifier (trending/ranging/high_volatility via ADX + ATR)
  • src/data/news.py — async CryptoPanic fetcher; returns empty list if API key absent
  • src/data/sentiment.py — async Alternative.me Fear/Greed Index fetcher; LunarCrush stub returning neutral defaults
  • src/data/onchain.py — Glassnode/CryptoQuant stubs returning zero defaults
  • src/data/assembler.py — assemble_context(): fetches OHLCV synchronously (ccxt is sync), fetches news/sentiment/on-chain in parallel with asyncio.gather(), assembles MarketContext, runs Tier-0 validation
  • src/validation.py — validate_context(): price freshness check (< 5 min), minimum news count (≥ 10), numeric range checks; HardRuleViolation raised for drawdown halt (> 15% from peak) and ATR volatility halt (ATR-14 > 2× ATR-30 avg)
  • tests/test_indicators.py — 35 tests for all indicator computations
  • tests/test_assembler.py — context assembly tests with mocked data feeds and default portfolio
  • tests/test_validation.py — 35 tests for price freshness, news count, numeric ranges, drawdown halt, volatility halt