Release v4.2.0 — Portable skills, real portfolio risk, Elo+Dixon-Coles football · artherahq/aria-code

Portable skill catalog — Aria now actually loads external SKILL.md workflows

ARIA_SKILLS_PATH (or a sibling aria-skills/skills checkout) is discovered at
startup; each skill registers as plugin:skill and its instructions are injected
only when a task matches
Every skill tree is verified against .claude-plugin/skills.lock.json (SHA-256)
before anything runs. A skill with no lock entry can never activate automatically —
only by explicit $name invocation — and one whose contents no longer match its
lock is refused outright
/skills doctor — catalog integrity and declared permissions
/skills trace — why each skill was selected or blocked, with match score
/skills now lists external catalog skills alongside built-in commands
A skill can declare which specialist agents it orchestrates (agents in
skill-policy.json); the active skill's agent list is surfaced in the prompt

New commands

/export-pdf <report.md> — renders a structured report to a designed PDF
(--theme=institutional|bloomberg, --sections=, --exclude=)
/artifacts — manage generated files: open, reveal, path, copy-path,
stats, prune
Structured Excel export and bilingual Markdown→PDF deliverables

/portfolio analyze now uses your real positions, not an imagined equal-weight basket

Previously every risk number — portfolio volatility, diversification ratio, and the
HEALTHY/NEEDS_ATTENTION/HIGH_RISK verdict — was computed as if every holding were
the same size, regardless of what portfolio_ledger actually held. On a
90%-concentrated book the difference is 45.4% annualized vol versus 27.4%: the gap
between "high risk" and "medium risk" on the tool's own thresholds, and it always
erred toward understating a concentrated position
With no arguments it now reads your ledger (cost-basis weighted) instead of falling
through to the static watchlist; the output states which basis it used, so an
equal-weight run can't be mistaken for real exposure

Football prediction switched to the Elo + Dixon-Coles engine

FootballAgent was calling a separate, simpler model: independent Poisson over a
fixed hand-maintained attack/defense table with a flat 1.25× home-advantage
constant. It now uses the Elo + Dixon-Coles predictor (negative-binomial tail for
lopsided fixtures, dynamic Elo/DC mixing, recency form, head-to-head,
self-calibration) that already shipped in the quant engine but was never wired up
Expect different numbers: on Germany vs Curaçao the two models' home-win
probabilities differ by 13.1 percentage points
New neutral_venue argument (defaults to False, preserving league semantics)

Agent team orchestration is no longer hardcoded to finance vocabulary

pip install aria-code on Python 3.14 failed with an error naming neither
aria-code nor its cause. requires-python was >=3.10 with no upper bound,
so pip would start the install, resolve the core dependency pandas_ta →
numba, and then fail on a numba source build (Cannot install on Python version 3.14.6; only versions >=3.10,<3.14 are supported). The bound is now
>=3.10,<3.14, so pip refuses up front with a clear "requires a different
Python" message. Python 3.10–3.13 are unaffected — verified 3.13 installs and
aria-code --version runs. Will be raised once numba ships 3.14 wheels
Dark-theme input text rendered gray instead of white
Core file tools (read_file, etc.) stayed unavailable when a restrictive skill
policy was active
weasyprint mangled emoji and CJK glyphs in exported PDFs
npm postinstall treated a non-zero exit code as uv install failure even when the
binary had installed correctly — now checks the filesystem
npm postinstall's critical notices (PATH reminder, terminal reopen) were being
swallowed instead of reaching the terminal

send_message's ~580-line inline agent loop deleted; CLI, SDK and self-host backend
now all run the same runtime.run_turn path (validated with real turns —
plain chat, forced-inline parity, native tool-calling with multi-round recovery —
before removal)

Navigation Menu