[HaoweiChan]

Why

The actual-portfolio propose pass ran at 2026-10-08 13:24 ET, mid-NYSE-session, and fired EXIT RVMD: "現價 185.43（2026-10-08 交易日）跌破 187.00". 185.43 was yfinance's still-forming intraday bar; RVMD closed 188.12 that day and never closed at or below 187 through 10-13. The book's stop veto (core/book.py stopped_out) is close-based, so the two rules disagreed about the same stop.

Task (verbatim): TASK-26 — propose compares stops against completed sessions only

Description

Seen 2026-10-09 01:24 Taipei (2026-10-08 13:24 ET, mid-NYSE-session): the actual-portfolio
propose pass fired EXIT RVMD because the "2026-10-08 session" price 185.43 was below the
187.00 stop. That number was yfinance's still-forming intraday bar; RVMD closed 188.12 that day
and never closed at or below 187 through 10-13. core/book.py stopped_out is close-based, so
the two rules disagreed about the same stop.

Change: each market profile (core/screening/__init__.py) carries its session close as
(timezone, HH:MM) — us America/New_York 16:00, tw Asia/Taipei 13:30. _run_propose drops
the panel's final row when that row is the market's today and the market's clock is before the
close, so every price propose reads (stop/EXIT, MA50, trail, scores) is a finished session.

Probe: KUROSHIO_LEDGER_DIR=$(mktemp -d) PYTHONPATH=. .venv/bin/python -m kuroshio.cli propose --ips ~/.kuroshio/ips-us.md --holdings ~/.kuroshio/book/latest/actual/holdings_us.yaml --market us --no-ledger · budget $0 · reps 1 — during the NYSE session the cards' session label is the previous trading day; after 16:00 ET it is today.

Acceptance Criteria

• groundwork: adopt pr-loop + portfolio-management diagnosis and plan #1 with the clock stubbed at 13:24 ET on a session whose partial bar is below a holding's invalidation price, propose emits no EXIT for it and reads the previous session; at 16:30 ET the same panel fires the EXIT — a test each.
• CLAUDE.md: drop hardcoded local path from the dev-env note #2 a TW market read at 21:00 Taipei keeps today's bar; a panel whose last row is not the market's today is never trimmed — tests with a stubbed clock, never the local one.

What changed

• Each MarketProfile carries session_close = (IANA tz, HH:MM): us/us-leadership America/New_York 16:00, tw Asia/Taipei 13:30.
• _run_propose passes the fetched panel through _completed_sessions: the final row is dropped when it is dated the market's today and the market's own clock is before its close. Every price propose reads (stop/EXIT, MA50, ATR trail, auto-scores, the asof label and the stop-ledger date) is therefore a finished session. Covers the CLI, kuroshio book's propose passes and the MCP propose tool, which all route through _run_propose.
  Not changed: screen, book's price columns and the other fetch_panel callers; early-close half days keep the bar out until 16:00 (delays the check, never fires early); card wording.

Verification

Gate: ruff clean · pytest 509/509 · 4fe0285 (after merging main)
Red-first: test_task26_mid_session_bar_does_not_fire_an_exit watched red on 5ab0fd5 with _completed_sessions as identity (EXIT RVMD at 185.43, asof 2026-10-08), green at 9498456; test_task26_after_the_close_the_same_bar_fires_the_exit and test_task26_completed_sessions_uses_the_market_clock red on the same stub, green at 9498456
Live: run local-9498456-1420Z ×1 on build 9498456 (2026-10-14 11:20 ET) · $0 — every card reads the 2026-10-13 session, 0 EXIT, exit 0; control on main (b10fe45) at the same minute reads the open 2026-10-14 bar
Not verified: a run after 16:00 ET reading today's bar live (covered by the 16:30 stubbed-clock test); TW live

Problems found

none

Follow-ups

none

Reviewer notes

Start here: _completed_sessions in kuroshio/cli.py
Reproduce: pytest tests/test_cli.py -q -k task26

Decision: met — verifier APPROVED, 1 model call

🤖 Generated with Claude Code

[HaoweiChan]

pr-loop/verifier APPROVED · no blocking findings
Gate: ruff 0 · pytest 482/482 · 9498456
Probe: local-9498456-1420Z