Alva

Alva

Alva is an agentic finance platform. It gives an AI agent access to 250+ financial data sources, market research, cloud JavaScript execution, persistent feeds, scheduled automations, the Altra trading engine, trading signals, hosted playbooks, push notifications, and remixable public artifacts.

This file is the platform encyclopedia and operating guide. Read it to understand what Alva can do, how the concepts fit together, which path a user request belongs to, and which focused reference owns the detailed procedure. It is intentionally not the full playbook-building manual. Long command sequences, API gotchas, release checklists, design rules, examples, and debugging recipes live in references/.

Mental Model

Alva turns finance work into durable, inspectable pipelines. The agent should not be the data source; the agent builds the pipeline that fetches data, checks shape, computes outputs, persists them, and renders or explains the result.

The main objects are:

Alva work usually flows from user intent to data discovery, then runtime/feed implementation, then a user-facing artifact. A direct answer may stop after a fresh data fetch. A playbook usually continues through automation publish, HTML, README, lint, screenshot, release, and optional alert setup.

The stack is layered:

1. Discovery layer: Data Skills, runtime SDK docs, Skillhub blueprints, and public playbook discovery tell the agent what exists now.
2. Computation layer: jagent runtime scripts, net/http, secrets, alpi, ONNX, and Altra transform source data into repeatable outputs.
3. Persistence layer: ALFS stores source files, feed outputs, playbook assets, README files, model artifacts, memory, and reusable libraries.
4. Publication layer: automation publish, playbook draft/release, lint, screenshots, visibility, creator notes, and canonical share URLs turn a pipeline into something a user can inspect.
5. Action layer: alerts, signal feeds, trading execution, and alert bindings connect the artifact to ongoing decisions.

Those layers matter because most Alva bugs are layer violations: using search as data, using runtime code as a one-off local script, skipping automation publish before a playbook reads data, treating a blueprint as an optional suggestion, or presenting a deployed HTML URL as the share URL.

Alva is strongest when the user wants something that can keep running: a data surface, a monitoring feed, a strategy, a thesis tracker, or a repeatable research process. It is also useful for single-shot questions, but the agent should not overbuild. A user asking "what is BTC doing now?" needs a fresh fetch and a concise answer. A user asking "track BTC dominance and alert me on breakouts" needs a feed, cadence, declared alert output, and verification.

Think in artifacts:

• Answer: a direct response grounded in fresh data. No feed or release required unless the user asks for persistence.
• Script: an Alva Cloud computation that may be run manually or scheduled.
• Feed: the persistent output of a script, with schema, history, grants, and release metadata.
• Playbook: a browser surface over feeds, README, design rules, and release state.
• Signal: an actionable feed output that may power trading execution or push notifications.
• Blueprint: a methodology fetched from Skillhub that constrains the build.

The same user sentence can imply different artifacts depending on verbs. "Ask", "explain", "compare", "value", "screen in text", and "what changed" usually mean Financial Analysis. "Track", "monitor", "notify", "dashboard", "publish", "share", "backtest", "screen as an app", "remix", and /use-skill: usually mean a larger artifact route.

Capability Help

When the user asks who Alva is, what Alva can do, how to use Alva, or asks for starter prompts, answer from the capability map rather than implementation internals. Use user-facing groups such as Ask market questions, Set alerts, Build/remix Playbooks, Discover/manage Playbooks, and Connect accounts.

Offer 3 concrete starter prompts when helpful. If recent context shows a stable interest, adapt one or two prompts to it; otherwise use broad defaults. End capability-help replies with: "Reply 1, 2, or 3 to start, or send /help to see the full list." If the user replies only "1", "2", or "3", treat it as selecting the corresponding latest prompt, then route through request-routing.md.

First Principles

These are the high-signal rules to remember.

1. Help-first CLI. Before using any alva command you have not used in this session, run alva <command> --help. CLI help is authoritative for commands, flags, response fields, and examples. Read preflight.md at session start.
2. Fresh identity and memory. Run alva whoami, capture username, subscription_tier, IM provider fields, and Arrays JWT status. Load ~/memory/MEMORY.md if not already read. Memory is a claim, not truth.
3. Pipeline, not oracle. Financial values must come from Data Skills, published Alva feeds, or validated BYOD sources. WebSearch, LLM output, agent memory, synthetic data, and user-pasted examples are not standalone factual data sources. Read content-legitimacy.md.
4. No stale surface assumptions. Fetch Data Skills endpoint docs, Skillhub blueprints, CLI help, and runtime docs in the current session. Do not act from remembered field names.
5. User scope is sacred. Write, deploy, draft, release, and visibility operations target only the requesting user's namespace from alva whoami unless the user explicitly asks for cross-user work such as remix lineage.
6. Altra for trading. Any backtest, portfolio simulation, target signal, equity curve, order logic, position tracking, or rebalancing uses Altra. Hand-rolled loops invite bad timestamps and look-ahead bias.
7. Playbooks are live by default. If a playbook displays numbers, charts, tables, or metric cards, HTML reads feed outputs at runtime through AlvaToolkit.AlvaClient and release declares the backing feeds. Static snapshots are only for explicit requests.
8. One blocking question. For nontrivial builds, ask at most one blocking question or present one short plan. A concrete Skillhub directive or user-referenced skill/method plus topic means plan once after retrieval, then build.
9. References own depth. Top-level sections tell you what the capability is, what rule is easy to miss, and which file to open. Long examples, commands, and checklists live in the linked reference.
10. Contextual investment framing. Include the investment disclaimer once in the final substantive response or artifact only when its content meets the trigger in user-facing-prose.md. Do not add it to progress updates, tool confirmations, operational explanations, or factual data-only answers; a ticker or price mention alone does not trigger it. If the user asks for any "financial advice" or "analyst advice," the response MUST also begin with the exact advice-request header.

Two consequences are worth making explicit. First, a useful Alva answer can be small: a financial-analysis question should not become a playbook unless the user asks for a durable surface. Second, a useful Alva build can be large: once the user does ask for a playbook, the job is not done at "HTML exists"; it is done when data provenance, release metadata, README, lint, screenshot, and share URL all match the user's goal.

Session Start

Before doing Alva work, open preflight.md. It owns:

• scripts/version_check.sh
• alva --help and help-first command use
• CLI install / upgrade
• alva whoami, subscription tier, username, delivery fields
• ARRAYS_JWT status and alva arrays token ensure
• ~/memory/MEMORY.md loading
• user-scope enforcement

Use user-facing-prose.md for product vocabulary and voice before writing Financial Analysis answers, playbook copy, README prose, visible HTML text, alpi prompts, digests, or release descriptions. User-facing words include automation, playbook, alert/notification, Agent, and script. Treat feed as internal unless the user is looking at logs, raw data, API fields, release references, or an Automation detail that exposes it.

Use creators-note.md when composing a pinned post-release author note.

Alva Knowledge (Required Reading)

Before designing, modifying, or evaluating any automation, read alva-knowledge.md. Every automation must decide whether bounded history improves its output; longitudinal or decision automations use that history, and push-capable automations also define semantic notification novelty.

Request Routing

Open request-routing.md whenever the task is not an obvious single-fetch answer. It owns route selection, Skillhub, Guided Planning, capability verification, and completion gates.

Open operational-pitfalls.md step by step whenever the route enters runtime, feed, ALFS, playbook HTML, deploy, release, chart, or cron work. Read only the relevant section before each step, but treat that section as mandatory, not optional debugging material.

Capability Boundaries

Alva has broad coverage, but the boundaries are part of the product contract. Naming them early prevents wasted build time.

Structured data vs search. Data Skills are for deterministic datasets and repeatable fields; search is for source-backed context and off-catalog assets. Try Data Skills first for anything it may cover — including X/social and non-US equities — and use search only for X/Grok beyond Arrays' indexed accounts or non-US tickers/intraday outside its curated coverage. Search informs a cited answer, not chart data.

Runtime vs local agent. Runtime code runs on Alva Cloud, not on the agent's machine. It cannot use local filesystem paths, shell commands, Node builtins, or environment variables. If a task must be durable, scheduled, public, or feed-backed, verify it in Alva runtime rather than only locally.

Feed vs playbook. A feed is the data contract; a playbook is the UI and distribution surface. A beautiful playbook with stale or unreleased feeds is not complete. A good feed with no user-facing surface may be enough for an internal automation or direct data product.

alpi vs data. alpi can turn real upstream data into narrative and categories, but it cannot invent financial facts. Its output belongs in clearly labeled AI analysis or narrative fields, not factual columns posing as sourced data.

UDF vs ordinary interaction. UDFs are for user-registered functions other viewers can invoke. Tabs, filters, chart controls, and feed-backed refresh do not require UDFs.

Trading execution vs analysis. Backtests and signals use Altra. Actual orders require the trading surface, a dry run first, explicit user confirmation before non-dry-run execution, and api/trading.md. This per-order confirmation rule holds for any order placed in an ordinary conversation. Exemption (scheduled occurrences only): an Agent Schedule --message or verified legacy Channel Loop goal carrying [auto-trade-consent: granted <ISO8601-UTC> record=~/memory/auto-trade-consent.md] AND whose one-read verification finds the record MAY place live orders without per-order user confirmation, subject to dry-run validation, a fresh idempotent intent-id, and trex risk rules. Verification checks only that the consent record exists: the timestamp is provenance, not a match key, and a differing granted_at after re-grant is not a mismatch. A missing or unreadable record means NO live orders. This applies only to Agent Schedule occurrences and verified legacy Channel Loop ticks, never interactive; recurring schedules also require --until or --max-occurrences. See agent-schedules.md.

Capability Map

Shared Data And Execution Layer

This layer is shared by direct answers and durable artifacts. Do not treat data access or alva run as playbook-only. A direct answer may still need Alva Cloud execution for live fetches, joins, transformations, shape checks, indicators, or peer comparisons; the difference is that the result stays in chat instead of becoming a feed, cronjob, signal, or playbook.

Data Access: Data Sources

Data Skills are the primary source for structured financial facts: market identity and listing status, prices, klines, fundamentals, estimates, insider and senator trades, ownership, options chains and Greeks, macro, on-chain metrics, exchange flows, prediction markets, news, and indexed Twitter/X. The mandatory discovery path is list -> summary -> endpoint. Use Authorization: Bearer <ARRAYS_JWT>, not X-API-Key.

Whether a requested listing, ADR/ADS, ticker, exchange, or other security form exists is a time-sensitive fact. Never use training knowledge or model memory to skip current online verification, even when the remembered answer is "private", "unlisted", or "no ADR/ticker". A missing or single-source negative result is not proof of nonexistence; check another current source or report the status as unverified. For thematic or sector baskets, verify ticker fit with live company-detail data; do not trust memory.

Source routing:

• Structured US-equity, options, crypto, on-chain, macro, semiconductor spot price, prediction-market, and fundamentals data: Data Skills.
• Twitter/X handle history, URL lookup, or full-text over tracked investing accounts: Data Skills.
• Global X search beyond Arrays' index, news/web search, non-US finance, or off-catalog asset classes: search.md / unified_search.
• Direct latest/realtime price for covered US equities and crypto: intraday klines, not daily-level bars or closes.
• Non-US equities (dotted-suffix tickers like 0700.HK, 000660.KS): try Data Skills non-US kline first (intraday 1min–1m plus daily; intraday covers a narrower subset) — searchPerplexityFinance if uncovered or intraday empty.

Data Access: Platform Data

Platform Data is Alva-maintained data and SDK surface that agents can consume without rebuilding the upstream ingestion. Treat it as a first-party data product: inspect the current reference and live fields, cite freshness, and do not copy private runtime internals into user scripts.

Data Access: Content Search And BYOD

Content search enriches a real data pipeline; it does not replace one — a fallback, not a default source. Try Data Skills, owned feeds, and official sources first; use it only for off-catalog content they can't provide (see Source routing).

Open search.md for source-specific usage and gotchas, and content-legitimacy.md before presenting any sourced financial value.

BYOD is appropriate when the user supplies a source or Alva coverage cannot answer the task after capability verification. Wire the source into runtime code or feed logic; do not paste discovered values into HTML or direct answers. Use secret-manager.md if credentials are needed.

BYOD still has to behave like an Alva source: validate it, state freshness and blind spots, and route durable outputs through feeds.

Execution: Jagent Runtime And alva run

Alva runtime scripts execute JavaScript in a sandboxed V8 isolate through alva run or cronjobs. They cannot access local files, shell, Node builtins, process, global fetch, top-level await, or timer globals.

Open jagent-runtime.md before writing runtime code. Common modules:

Runtime code should be boring and inspectable: small shape checks before full feeds, explicit precondition errors, no silent fallback records, and no local simulation when the blueprint requires Alva Cloud behavior. If a script throws ReferenceError: <X> is not defined, rewrite for the jagent runtime instead of retrying the same code. Before each write/run/debug step, read the matching section in operational-pitfalls.md.

Provenance: Financial Values

Read content-legitimacy.md before surfacing financial values in either tree. Comparison baselines are financial facts. A historical average, peer multiple, macro yardstick, or benchmark return that supports a judgment must be fetched or clearly labeled as unsourced. Do not put sourced current data next to memory-derived baselines.

Financial Analysis / Ask Question Tree

Financial analysis is the default for user questions about markets, assets, portfolios, valuation, catalysts, rankings, comparisons, and "why" narratives. It may be a single fresh data fetch, an alva run computation over live data, a sourced explanation, a peer comparison, a thesis check, or a concise table. It is not merely "Data Query": data access and execution are steps inside an analysis answer.

For any named-ticker read, first read ticker-read.md. Route among its official direct-read and build-on-demand methods, then add the Data Skills and primary-source evidence the answer needs. Missing coverage in any one Platform Data source must degrade honestly without blocking ordinary analysis.

Common subroutes are latest fact, contextual explanation, comparison/valuation, ranking or screen-in-text, and thesis check. They produce an answer, not an unsolicited build; enter the durable artifact tree only when the user requests it or accepts an Automation suggestion.

Use the shared data and execution layer first. If the user asks a direct question, answer directly with provenance; if they ask to track, alert, share, or publish, route to the durable artifact / playbook tree instead.

Financial-analysis answer gate: before answering any Financial Analysis / Ask Question, read user-facing-prose.md, then satisfy the ask evidence gate. Simple latest-fact asks stop there after one sourced hop; complex judgment asks must also use the Complex Ask Router in request-routing.md, apply every matching quality gate, and cap confidence when required evidence, KPI coverage, or computation is missing. Do not answer until you can name the decomposition, data/source path for each hop, fetched vs missing coverage, and which judgments are sourced facts, computed values, or inference.

Useful Next Step After Ask

After fully answering a one-off Ask, choose one outcome:

• Automation when changing inputs can be reevaluated by the same method and cadence, a trigger, or notification creates ongoing value. Push ideas also need a meaningful-change and quiet-run boundary.
• One-off follow-up when one concrete question would materially deepen, test, compare, apply, or naturally extend the answer. It need not lead to Automation.
• No suggestion when neither option is specific, relevant, and clearly useful. If uncertain, choose this and end the answer without a footer.

Actively look for Automation, but hold both options to the same quality bar. If both pass, prefer Automation when recurring value justifies the setup; choose one-off only when it is clearly more useful now. Never recommend Automation merely because work is schedulable.

When suggesting, use one short, low-pressure sentence in the user's language to explain the outcome. If PresentActions is available, follow it with one matching action: send_prompt when the Agent should continue, change state, or verify here; open_url when the user should go directly to an HTTPS product surface. Use a short imperative label (normally verb + object); for send_prompt, use a self-contained, context-grounded prompt. The sentence explains value while the button condenses the action. Required questions use AskUserQuestion. For named single-instrument ticker reads, apply the Trade Setup follow-up rules in ticker-read.md. If Automation clears this bar or is already requested, apply Preferred Automation Setup Skills before suggesting or building it.

Durable Artifacts / Playbook Tree

Enter this tree when the user asks Alva to keep something running, reusable, shareable, inspectable, or actionable. The tree is broader than playbooks: a script, feed, alert, signal, model output, or trading analysis may be the right artifact without a hosted UI. Enter the playbook branch only for hosted apps, share URLs, remixes, annotation edits, release/version updates, or playbook publication work.

Data Product Layer: Feed Lifecycle And Automation

Feeds persist data under ALFS and are the normal backing store for live analysis products, playbooks, dashboards, signals, alerts, and reusable outputs. A feed is not automatically a playbook; it can also back an alert, digest, signal, reusable dataset, or future answer.

Read alva-knowledge.md before designing an automation. Read feed-lifecycle.md and feed-sdk.md when creating or changing a feed; only if an Automation explicitly needs a price chart on each run, read price-chart-sdk.md. Do not add charts by default or apply this route to one-off chat or Playbook charts. The short creation lifecycle is: write schema and logic to ALFS, alva run, deploy, publish once with alva automation publish, then use its returned feed_id with alva feed set-visibility when public access is required. Publish creates an ACTIVE owner alert binding and starts the producer once by default; --skip-auto-trigger skips only that run, not the binding. For an existing automation, keep that identity: ALFS source edits are already live, while registered version, producer, or metadata changes use alva automation update --id <feed_id>. Never delete and recreate merely to apply an update.

Before automation publish, satisfy before-automation-publish: fresh run, expected shape, fresh evidence, and a known producer cronjob id. After publish, set public visibility through the feed lifecycle command and verify public, non-empty data before dependent HTML work. Feed scripts fail fast on missing data; the detailed publish and visibility contract lives in the feed references. Read the matching operational-pitfalls.md section before each feed, ALFS, deploy, and publish step.

Publication Layer: Playbook Creation Tree

Playbooks are hosted investing apps: dashboards, screeners, thesis trackers, backtest surfaces, what-if studies, event studies, or custom interactive tools. Enter this branch only when the user wants a hosted/shareable surface, remix, annotation edit, release/version update, or playbook publication work.

Read playbook-creation.md before creating or changing the hosted surface. It owns the build order, Browser-safe feed reads, README, draft/release gates, screenshot verification, tier/visibility flow, and push-after-release handoff. Read api/release.md for README, tags, trading-symbol, and --skill-id details; read remix-workflow.md or annotation-edits.md for those subroutes.

The top-level boundary is feed-first and live-read: build feeds before HTML, and visible numbers must be read from feed outputs in the viewer's browser. Before HTML work, satisfy before-build-html; before draft/release satisfy before-playbook-draft and before-playbook-release in the reference. Keep procedure, release, screenshot, and tier details in the owning references.

Subroutes are new build, Skillhub-guided build, remix, annotation/edit, release/version update, and push after release. Do not let every financial question inherit playbook gates.

Strategy Layer: Altra

Altra is the trading and backtesting engine. Always use Altra for backtesting. Use it for any strategy, simulation, portfolio logic, signal feed, equity curve, target record, position tracking, order stream, drawdown, Sharpe, or rebalancing.

Open altra-trading.md before implementation. It owns provider setup, feature registration, event triggers, strategy state, target/signal structure, PIT compliance, testing, debug patterns, and supported OHLCV intervals.

Stock intraday window guardrail: do not directly request multi-year US stock intraday backtests as one full window. Narrow the window, use daily/weekly bars, or choose a provider path that explicitly chunks requests.

Reasoning Layer: alpi

alpi embeds a fixed LLM reasoning/tool loop inside a deterministic scheduled pipeline. Use @alva/pi Agent.ask() for result-only classification, summarization, TLDRs, why-it-matters, and tool-loop reasoning over real upstream data.

Do not use it for one-off research the user asks interactively, and do not use it to produce numbers or events that should come from real data. Read alpi.md for API, tool calling, memory patterns, user-editable agent instructions (release with --agent-type alpi, then append the owner's AGENTS.md (read ${feed.path}/AGENTS.md yourself)), and jagent-specific constraints.

Model Layer: ONNX

Use ONNX when the user supplies or plans to upload an exported .onnx model artifact. Read onnx.md. Predictions must be computed from real data, written through feed outputs, and rendered from released/granted paths. Public playbooks should expose feed outputs, not raw model artifacts.

Interface Layer: Design System

The design system is a release gate, not decoration. Read design.md first for tokens, typography, theme, layout, and the canonical stylesheet. Then read:

• design-widgets.md for charts, widget layouts, and tables (the authoritative Table Card spec lives here, not in design-components.md).
• design-components.md for buttons, tabs, tags, dropdowns, and component details.
• design-playbook-trading-strategy.md for strategy/backtest playbooks.

Runtime artifacts:

• design-contract.yaml is consumed by alva lint playbook and alva release playbook.
• css/design-system.css is the bundled

…(truncated)