Implements compute_7d_return, compute_metrics, load_baseline_metrics,
and make_decision functions with full TDD coverage (11 tests passing).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements compute_7d_return, compute_metrics, load_baseline_metrics,
and make_decision functions with full TDD coverage (11 tests passing).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pending hypotheses queue by priority and promote when a slot opens,
rather than pausing a running experiment mid-streak.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Covers 5 tasks: knowledge base structure, /iterate command,
/research-strategy command, and two GitHub Actions workflows with
rolling PR logic.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
At most one open PR per skill at any time. Daily runs push onto the
existing branch and update the PR description. Merging resets the cycle.
Prevents PR accumulation from unreviewed automated runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Scanners and strategies are 1:1 in the current codebase — separate folder
was artificial. Each scanner file now captures both implementation and thesis.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds daily /iterate and weekly /research-strategy cron workflows to the
spec — full autonomous loop with PR-gated merges, no auto-merge to main.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Defines /iterate and /research-strategy skills + docs/iterations/ folder
structure for a generic learn-improve-repeat cycle, demonstrated with the
trading agent discovery pipeline.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The previous implementation redirected sys.stderr to /dev/null using a
context manager. This is not thread-safe: 8 concurrent scanner threads each
mutate sys.stderr, and when one thread's context manager closes the devnull
file, another thread that captured devnull as its saved stderr attempts to
write to the closed fd and raises "I/O operation on closed file".
This corrupted sys.stderr state caused _fetch_batch_prices to fail and
all per-ticker get_stock_price fallback calls to return None, resulting in
every candidate being dropped with "no data available".
Fix by suppressing at the Python logging level instead of redirecting
sys.stderr. Logger.setLevel() is protected by internal locks and is safe
to call from concurrent threads.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs causing zero recommendations:
1. risk_metrics.py was untracked — importing it raised ModuleNotFoundError which
was caught by the outer try/except in filter.py, silently dropping all 32
candidates that reached the fundamental risk check stage.
2. Minervini scanner at max_tickers=200 took >5 min to download 200 tickers x 1y
of OHLCV data. ThreadPoolExecutor.cancel() cannot kill a running thread, so the
download kept running as a zombie thread for 20 more minutes after the pipeline
completed, holding the Python process alive until the 30-min workflow timeout
killed the entire job.
Reducing to 50 tickers brings the download to ~75s, well under the 300s global
scanner timeout.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The same-day mover filter used abs() to check intraday movement, which
filtered both gap-ups AND gap-downs. On volatile/crash days (e.g. 2026-04-07)
all stocks dropped >10% from open, causing every candidate to be filtered and
leaving zero recommendations.
The filter's purpose is to avoid chasing stocks that already ran up. A stock
down 20% intraday is not "stale" — it should be evaluated on its merits.
Changed threshold check from abs(pct) >= threshold to pct >= threshold so only
upside movers are filtered.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
yf.download(592 tickers, period=1y) takes 20+ minutes in CI, causing
the 30-minute job timeout to trigger. Add max_tickers=200 (configurable)
to limit the batch download to the first N tickers from the file. The
concurrent scanner pool already has a 5-min global timeout, but the hung
download thread monopolises network connections and starves the filter stage.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
minervini.py existed but was never committed. Without the file on the
remote, the __init__.py import added in the previous fix causes an
ImportError in CI.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add permissions: contents: write so git push works (was failing with 403)
- Add continue-on-error: true on discovery step so partial output still commits
- Change all commit/tracking/position steps to if: always() so they run regardless of discovery outcome
- Use commit-then-pull-rebase-then-push pattern to handle branch divergence
- Fix minervini scanner missing from scanners/__init__.py (enabled in config but never loaded)
- Fix .gitignore: results/* + !results/discovery/ so CI run logs can be committed
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Same issue as options_flow: early exit on candidate count discards strong
signals that happen to be later in iteration order.
insider_buying: Dict iteration order matched OpenInsider HTML scrape order,
not signal quality. Now scores by cluster buys + C-suite + dollar value,
then takes top N.
technical_breakout: Stopped at limit*2 in file order despite data already
being batch-downloaded (zero API cost to check all). Removed early exit,
scan full universe, sort by volume_multiple.
sector_rotation: Checked laggards in arbitrary dict order, spending API
calls on random tickers. Now sorts by most-negative 5d return first so
the strongest laggard candidates are checked before hitting the budget.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the scanner stopped as soon as self.limit candidates were found
from as_completed() futures. Since futures complete in non-deterministic
network-latency order, this was equivalent to random sampling — fast-to-
respond tickers won regardless of how strong their options signal was.
Fix: collect all candidates from the full universe, then sort by options_score
(unusual strike count weighted 1.5x for calls to favor bullish flow) before
applying the limit. The top-N strongest signals are now always returned.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
tqdm writes to stderr immediately on __enter__, before any loop iteration.
In Streamlit's thread/subprocess context stderr can be a closed pipe, causing
'I/O operation on closed file' which _run_call catches and returns {} — so
the entire news enrichment step was silently skipped every run.
Replaced tqdm progress bars with logger.info() calls in:
- get_batch_stock_news_google() in openai.py
- get_batch_stock_news_openai() in openai.py
- Reddit DD parallel evaluation in reddit_api.py
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Youssef Aitousarrah
Aitous
github-actions[bot]