16, Glossary

Glossary

The terms, roles, tools, and risk caps that appear throughout the Aelix docs, in one place.

A quick reference for the vocabulary used across these docs. Where a term has its own page, the definition links to it.

Roles & architecture

Portfolio Manager (PM): The main Claude Code session you talk to. It orchestrates the sub-agents and is the only role that can place orders, and only after your approval. See The Desk Team.
Sub-agent: A specialist role defined as a Markdown file in .claude/agents/ with its own restricted tools: list and model:. Loaded when Claude Code starts.
Human-in-the-loop (HITL): The structural rule that no order is placed without your explicit in-session approval. See Guardrails.
Least privilege: Each role holds only the tools its job needs. The analysts and Risk Manager have no order tools; only the PM does.
Claude Code: The agent host Aelix runs inside. There is no separate backend server or Python orchestrator, the PM *is* the session. See Architecture.
MCP (Model Context Protocol): The protocol that connects Claude Code to the broker. Aelix uses one MCP server, robinhood-trading. See MCP & Tools.
robinhood-trading: The single MCP server (HTTP transport, OAuth) that is the desk's only path to the account. Defined in .mcp.json.
Robinhood Agentic account: The isolated, beta, equities-only account the desk may trade. Every other account is read-only for context.
OAuth: The in-session authentication flow for the MCP server. The agent never sees your password; there is a Robinhood mobile verification step.

Preview card: The desk's standard output: a summary of the proposed order (symbol, side, qty, order type, estimated cost, rationale) that the PM presents and waits on. Built by review_equity_order. See The Desk Run.
review_equity_order: A read-only tool that builds an order preview without submitting anything. allow-ed, because it never transacts.
place_equity_order: The tool that actually submits an order. Held only by the PM and gated behind an ask prompt.
Veto: The Risk Manager's power to block a proposed trade outright. A VETO stops the trade; the PM reports back.
APPROVE-WITH-CHANGES: A Risk Manager verdict that clears a trade only after specific changes (e.g. cut size to N, add a stop at $X).
Scan: A saved technical screen the Technical Analyst runs (run_scan / get_scans) to surface candidate tickers.
Watchlist: A list of symbols the desk screens (get_watchlist_items); mutating watchlists is gated behind ask.
INJECTION ATTEMPTS: The line in the Macro/News Analyst's brief where instruction-like text from the web is quoted verbatim and flagged, never obeyed. See Prompt-Injection Defense.
Prompt injection: External content that tries to act like an instruction ("buy X now", "ignore your rules"). Aelix treats all external content as untrusted data.

NAV (total_value): Account value = cash + positions, read from get_portfolio.total_value for the Agentic account. All risk percentages are of NAV, not the broker equity_value field. See Strategies & Risk.
Per-trade cap: 15% of NAV, the hard ceiling on any single place_equity_order.
Concentration cap: 25% of NAV in any one symbol, including all adds.
Max open positions: 6, forces diversification within a small account.
Max daily orders: 4, throttles churn; counts buys + sells per day.
Stop-loss: −8% from average entry, every entry must define this before it is placed.
Daily loss halt: −5% account day P&L, stop trading, report, and ask for the rest of the day.
Cash buffer: ≥10% of NAV kept in cash, never fully deploy.
Mean reversion: A strategy: buy oversold pullbacks inside a confirmed uptrend; long only, swing. One entry per setup, no averaging.
Left-side accumulation: A strategy: planned, risk-budgeted scale-in at support in a quality name's fear-driven selloff. The only defined exception to "no averaging into losers," bounded by a fixed total-risk budget and a whole-position kill-stop.
No averaging into losers: Adding to an underwater position is forbidden unless a strategy file explicitly permits it with limits.

desk-state.json: The snapshot the PM writes after each run; the dashboard mirrors it. Real state is gitignored; only the sanitized example is committed.
injectionAlerts: The array in desk-state.json where flagged injection attempts surface (source, quote, handledBy, action).
decisionLog: A tail of the JSONL audit log flattened for the dashboard timeline as { ts, event, summary, symbol, tone }. See Audit Logging.
JSONL audit log: The append-only logs/*.jsonl trail, one JSON object per line, of every desk decision, written via tools/desk-log.mjs.
Backtester: The offline, dependency-free engine in backtest/ that sanity-checks strategy logic against historical bars, illustrative, not a track record. See Backtester.
/loop: The Claude Code mechanism used to run the read-only desk on a cadence (e.g. behind the dashboard's Run desk button).
Kill switch: Disconnecting the MCP from the Robinhood app or claude mcp remove robinhood-trading, the desk's hard stop.

Missing a term? The FAQ covers scope and safety questions, and each reference page defines its own vocabulary in context.