Aura/CLAUDE.md

aura — project rules

aura is a "game engine for traders": a Rust framework and a playground to author trading nodes, backtest them deterministically and massively in parallel, compose them fractally, validate them (sweep / Monte-Carlo / walk-forward), and freeze a validated strategy into a standalone bot with a broker connection.

This file is the project sittenkodex. It imports the universal discipline from ~/dev/skills/templates/CLAUDE.md.fragment and adds aura's domain invariants. The full architecture lives in the design ledger (docs/design/) and the specs (docs/specs/), not here.

Roles

I am the orchestrator, not the implementer. The skills-plugin agents are my workers: I plan, design, decide, and integrate; they implement, refactor, test, and diagnose. Trivial mechanical edits I may do directly; anything needing broad reading or judgement goes to an agent. Agent reports describe intent, not outcome — I verify the diff and the test output myself before committing.

Commit discipline and main-branch sanctity

• Only the orchestrator commits. No skill agent runs git commit; agents leave their output as unstaged working-tree changes for me to inspect and shape into commits.
• main HEAD is sacrosanct. No git reset/git revert on main. main moves forward only via my commits. Wrong agent output is discarded with git checkout -- <paths>/git stash; a bad landing is fixed forward, never rewound.
• When a commit closes a Gitea issue, reference it in the body: closes #N (or refs #N for non-final work).

Design rationale ≠ implementation effort

Design choices are justified by substance — semantics, structural fit, what the design permits vs forbids, compositional clarity, future-proofing. Implementation effort ("approach A touches 250 sites, B touches 1") is an observation about the current code, not a rationale. Effort is at most a named tiebreaker after substantive reasons tie.

Bug fixes — TDD, always

Bug fixes are RED-first and autonomous: the failing test exists in the working tree before any fix. The debug skill is mandatory for any observable misbehaviour (failing test, panic, wrong output).

Domain invariants (load-bearing — never silently violate)

These are the contracts the whole design rests on. A change that breaks one is a design decision, not a refactor, and belongs in the ledger.

1. Determinism. A backtest is a deterministic, synchronous, non-concurrent event loop that reaches a unique state after each input tick. Same input → same run, reproducibly. Two backtests are fully disjoint → concurrently executable without locking. Parallelism is across sims, never within one.
2. Causality / no look-ahead. A node sees only the past. Look-ahead is made structurally impossible (read-only input windows that end at the cursor; resamplers emit a bar only once it is complete), not merely discouraged.
3. One merge, at ingestion only. Heterogeneous timestamped sources are k-way-merged into a single chronological cycle stream at the ingestion boundary. There is no merge / as-of join inside the graph.
4. The four scalar base types, streamed as SoA. Only i64, f64, bool, timestamp are streamed, as columnar Structure-of-Arrays. Composite streams (e.g. OHLCV) are bundles of base columns. Non-scalars (String, Records, tables, calendars) exist as metadata beside the hot path, never in it.
5. Acyclic dataflow. The graph is a DAG; the only feedback path is an explicit delay/state node (the RTL "register"). The "cycle" of the research workflow is not a dataflow cycle.
6. Record-then-replay determinism boundary. Anything non-deterministic, external, or slow (LLM news agents, web sources) is materialized into a recorded, timestamped stream before it enters the engine. The sim never makes a live external call mid-replay. (See ~/.claude/CLAUDE.md for the IONOS consent rule: external LLM calls happen at the recording/live-source edge, with explicit per-session consent, never inside a backtest.)
7. Strategy output is an intent/exposure stream; position management is a decoupled derived layer. The DAG expresses one state at t (C8: ≤1 record per eval), so a strategy's primary, backtestable output is a signed, bounded exposure f64 ∈ [-1,+1] per cycle, not an equity curve. Signal quality is measured by the sim-optimal broker — a downstream consumer node that integrates exposure·return into a deterministic, frictionless synthetic equity stream in pips. The broker-independent position-event table (event_ts, action[buy/sell/close], position_id, instrument_id, volume) is a decoupled, derivable layer — the first difference of the exposure state, a computed table (not a per-eval output, since one decision instant may yield

   1 event) — that feeds realistic broker nodes (currency equity, real frictions) for viability and deploy. The two broker classes attach at two points: sim-optimal on the exposure stream, realistic on the derived position-event table; brokers are ordinary downstream nodes, never part of the strategy.

8. Deploy artifacts are frozen. Hot-reload (cdylib) is an authoring-loop tool only. The live bot is a statically-linked, versioned, frozen artifact — never hot-swapped (audit trail: this bot = this commit).
9. Engine / project separation. This repo is the reusable engine. Research projects are separate external repos that depend on it. Project-specific signals live in a project's nodes/; cross-project-reusable nodes in shared crates; universal blocks in aura-std (shipped here). Reuse is cargo-native; the hot-reload unit is always the project-side cdylib. No user/project signals in this repo (only examples/ fixtures for the engine's own tests); no multi-project manager or node registry inside aura. A project is always a Rust crate (a cdylib library of node / strategy / experiment blueprints + a static Aura.toml), hosted by aura during research and frozen to a standalone binary for deploy.
10. Authoring surface — all logic is Rust. Nodes, strategies, and experiments/harnesses are authored in native Rust via Claude Code + the skills pipeline, using builder APIs. Declarative config (Aura.toml) carries only static project context, never logic — no experiment/strategy DSL (the RustAst trap). aura ships no embedded coding-LLM; IONOS LLMs are used only as a runtime data source (news bias), with per-session consent, never in the code path.
11. Construction is a bootstrap phase. Blueprints (param-generic graph-as- data from a Rust builder) are bootstrapped into frozen instances (buffers sized, topology fixed) by binding params + data + seed. Params configure and size nodes but never change topology (a topology change is a different blueprint). The harness — sources + strategy + broker node(s) + sinks — is the root sim graph, itself bootstrapped, and is C1's disjoint unit; its structural axes form the experiment matrix, its tuning params the sweep.
12. The World is the product; the playground is a trace explorer. Three ontological tiers: a node is an open, composable fragment (at most one output, C8); a harness is the closed root graph that runs (sources + strategy + brokers + sinks + clock — C1's disjoint unit / root scope), not a node; the World is the program that dynamically constructs and orchestrates families of harnesses (walk-forward / sweep / MC / comparison) — aura's differentiator, not the single backtest. The World is a program: nothing is displayable until it runs, and what is displayable is exactly what a sink (C8) records into the registry (C18). The playground plays any harness and is an execution viewer / trace explorer (structure before, live streams during, recorded traces after) — never a scene editor; topology is grown in Rust + hot-reload, runtime params are UI-tunable.