- Add opt_max_agents config (default 75% of CPUs, capped at 20)
- Set LocalAgents in tester .ini and pin MT5 to CPUs 0..max_agents via taskset
- Resolve .set filename consistently between tester + ini config
- Move .mt5mcp_jobs to temp dir for persistence across CWD changes
- Fix optimization XML row parsing to descend into Cell/Data nodes
- Add run_rolling_backtest tool for N consecutive weekly backtests
- Add max_passes support for optimization to cap genetic passes
- Kill stale MT5/agent processes before optimization launch
- Auto-parse optimization results on status poll when process completes
- Fix XML namespace for SpreadsheetML parsing
- Change optimization model from 4 to 1 for better accuracy
- Persist completion results to job metadata
- CHANGELOG: inactivity_kill_secs default is disabled (None), not 120
- TROUBLESHOOTING: clarify inactivity_kill_secs is opt-in (must be
passed explicitly); correct "default 120" claim
- MCP_TOOLS: inactivity_kill_secs is disabled by default; add note
that Wine/macOS may not exit naturally even with ShutdownTerminal=1
- handlers/backtest.rs: remove stale comment claiming inactivity
watchdog is "intentionally skipped when shutdown=true" — replaced
with accurate description of the new 30s HTML-wait + kill behavior
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- fix: inactivity watchdog now waits 30s for HTML report then kills
terminal64.exe unconditionally — ShutdownTerminal=1 does not cause
MT5 to exit on Wine/macOS; relying on natural exit caused all runs
to fall back to journal extraction
- fix: tester agent log deduplication — each deal is written twice in
the log; use HashSet to skip already-seen deal numbers
- fix: position tracker for entry direction inference — infer "in"/"out"
from per-symbol signed lot accumulation instead of guessing
- fix: is_closed_trade() no longer gates on profit!=0.0 — journal deals
have profit=0.0 legitimately; the old gate hid all 140 deals from
list_deals
- fix: log file selection priority — prefer Agent-127.0.0.1 over
Agent-0.0.0.0 (startup-only log), tiebreak by file size
- feat: launch_backtest default shutdown=true; inactivity_kill_secs
default 120s exposed as tool parameter
- feat: report DB stores deals in SQLite; analytics resolve by
report_id / report_dir / latest
Verified: 1-month DPS21/XAUUSD.cent/M5 backtest produces full HTML
report with all metrics (win_rate=70%, profit_factor=0.77, sharpe=-3.61,
max_dd=6.93%) and all 17 analytics tools returning real data.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add cargo-publish job that runs in parallel with binary builds.
The release job now depends on [build-macos, build-linux, cargo-publish]
so the crate is always published to crates.io before the GitHub release
and binaries are attached.
Requires CARGO_REGISTRY_TOKEN secret in repo settings.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add ini_safe() helper that strips CR/LF from user-supplied string params
before they are written into terminal.ini and backtest_config.ini.
Applies to: expert path, symbol, timeframe, set_file — any value that
could carry an embedded newline and inject extra INI directives.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Devid HW