HX107 Rion AI Automation

HX107 teaches how to use Rion without confusing model text, chart context, and a confirmed desktop action.

Surface	Public status
Chart chat and read-only active-chart context	Current
Structured desktop actions	Preview
Legacy Rion Autopilot and Strategy Discovery	Preview; when invoked through a supported legacy path, default cloud-ML upload lacks a dedicated consent prompt
Hyperion Sim actions	Preview, simulation-first
Agent Builder	Preview and entitlement-gated
Managed Rion cloud desktop bridge	Internal; not a third-party API
Hyperliquid or KuCoin live action through the Local Desktop API	Release-gated Preview; capability-discovered in an enabled installed build

Use AI Agent for the current provider list, Rion Autopilot And Strategy Discovery for the generated-code and research workflow, and Local Desktop API for the installed action contract.

Reviewed 1.1.27 source

Default HyperionX Cloud chat uses grounded tools for chart queries, indicators, supported high/low drawings, and task-owned indicator authoring. It has no trading, Discovery, Validator, Optimizer, or Research Lab execution tool. The older workflow examples apply only where the installed provider/build supports that separate route. Check the documentation version map; source review does not guarantee public-release availability or change Preview status.

Lesson 1: Choose Where The Prompt Is Processed

The selected provider determines where prompts and attached context are processed. HyperionX Cloud, OpenAI, and Claude are cloud destinations. Ollama is local only while its configured endpoint is local.

Before sending a prompt:

1. Verify the provider and model shown by the AI settings/header.
2. Confirm that the account and organization are allowed to process the intended market, strategy, and customer context.
3. Remove secrets and unnecessary proprietary data.
4. Read AI Data And Privacy for local storage and provider boundaries.

The parameter-name filter for words such as key, secret, password, and token is a defensive layer, not a guarantee that arbitrary custom text is safe.

In reviewed 1.1.27 source, pasted images require the updated vision-capable Cloud service. Check the preview before sending. Images are saved with protected local history and can be sent again on follow-up turns; text filters do not redact their pixels. Crop private content before attaching it.

Lesson 2: Verify The Active Chart

Rion's read-only snapshot follows the active chart. It can include symbol/timeframe, recent and visible candles, indicator and strategy summaries, Chart Trader state, and selected research-tool state.

For a multi-chart layout:

1. Click the intended chart.
2. Ask Rion to repeat the symbol and timeframe it received.
3. Compare that answer with the visible chart.
4. Only then ask for analysis.

The public context contract does not currently select an arbitrary chart by public chart ID. If the context is stale or points to the wrong chart, stop and reselect rather than correcting the model only in prose.

Lesson 3: Separate An Answer From An Action

Use three evidence levels:

Evidence	Meaning
Model message	The model produced text. Nothing proves a desktop mutation occurred.
Preview tool response	HyperionX accepted or rejected a structured local action. Check status and request ID.
Platform/provider state	Orders, positions, drawings, compiler output, or research state confirm the result.

Never report that a build ran, a drawing changed, or an order was placed based on chat wording alone.

Lesson 4: Practice Read-Only Requests

Begin with tasks that do not mutate state:

• summarize the active chart and indicator context
• explain a strategy parameter without editing it
• report current Validator/Optimizer state only through a surface that exposes it
• inspect the current task-owned Code Lab artifact, or use script inventory where that tool is available
• explain a compiler diagnostic
• summarize upcoming economic-calendar context

Review the returned context for account labels, strategy names, parameters, paths, or other data you did not intend to send.

Lesson 5: Use Preview Desktop Actions

The default grounded Cloud path in reviewed 1.1.27 source can change indicators, draw supported high/low levels from verified evidence, and create or revise and compile a task-owned Code Lab indicator when the required entitlements and profile gates allow it. It does not execute the legacy research pipeline. Separate Local Desktop API routes and legacy workflows expose their own actions.

For indicator authoring, ask Rion to build the requested HyperionX indicator, inspect the actual compilation result and source, then test its plotted behavior. A successful compile does not establish runtime correctness or successful historical validation. The scoped HTTP Code Lab route still supports only open/list/compile/status; it is not the grounded conversation's artifact-authoring tool.

For any mutation:

1. State the intended action and target.
2. Prefer a scoped action over a natural-language general command.
3. Use dry-run when the selected interface supports it.
4. Require the expected confirmation path.
5. Check the structured response.
6. Verify the result in the owning platform window.

Action names and permissions are a Preview contract. Discover capabilities from the installed build rather than copying a route/action table from a lesson.

Lesson 6: Keep Trading Simulation-First

Strategy Discovery cloud consent limitation

The legacy Strategy Discovery workflow uses cloud ML by default and can upload the full research CSV plus metadata without a dedicated consent prompt when that workflow runs. This is a public-release blocker for unrestricted use. Add local only or no ml on a supported legacy request whenever the data must remain local, then verify that the ML stage was skipped. These phrases do not enable Discovery in the default grounded Cloud engine. See Rion Autopilot And Strategy Discovery before running it.

The reviewed 1.1.27 default Cloud conversation does not execute trades. Use this exercise only through a separate supported trading interface. That path can prepare or perform only the provider actions exposed by the installed build after all global and provider-specific gates pass. Hyperion Sim is the simulation-first path; guarded Hyperliquid and KuCoin live actions require separate capabilities, settings, entitlements, active-context matching, connection state, and exchange permissions.

Before a simulated action:

• select LocalPaper explicitly
• verify symbol, side, type, price, and quantity
• keep auto-execute off while learning
• confirm the action when required
• inspect the returned result and then the Orders/Positions panels

Agent Builder fields and a live_trade entitlement value do not independently enable live HTTP order routing. A newer release-enabled build can add guarded Hyperliquid and KuCoin actions only when the provider-specific capability and setting, active account/instrument, connection, and exchange-permission checks all pass. Follow Local Agent Live Trading API before testing either path.

Lesson 7: Understand Agent Builder Limits

Agent Builder can create specialized profiles, but it is not a hard trading-risk engine.

• Selected context/tool toggles are enforced for non-baseline specialized profiles.
• The built-in agent and rion-scout IDs bypass that specialized-profile filter.
• Maximum quantity, daily loss, position/trade limits, allowlists, approval text, routing text, and several other fields remain advisory metadata in the reviewed implementation.
• Agent Builder profiles do not scope the global Local Desktop API token.
• Telemetry and recent-action summaries in the Agent Builder UI are placeholders.

Use a new non-baseline profile, start read-only, and review Agent Builder before enabling any tool.

Lesson 8: Audit What Can Be Audited

Token-authenticated Local Desktop API calls write daily JSON Lines audit records. Those records describe the request and local result; they do not prove an exchange fill. Ordinary model conversation and every in-app click are not automatically equivalent to a Local Desktop API audit record.

For an incident:

1. Preserve the request ID and timestamp.
2. Check the local API response.
3. Check the owning desktop window.
4. Check provider/account state for any trading concern.
5. Export logs through Support And Diagnostics.

To start a fresh chat in reviewed 1.1.27 source, right-click the chat panel and choose Delete all. It clears saved messages and images, pending input and confirmations, and local conversation tool context. Cleanup of the previous cloud context is attempted when a valid session is available; this is not provider-retention deletion. Generated files and separate chart memory remain. Older builds may label the command Clear Chat.

Completion Check

• You can identify the selected model destination before sending context.
• You can prove which chart supplied the snapshot.
• You can distinguish model text, a tool response, and confirmed platform state.
• You keep all AI trading exercises on an explicitly selected LocalPaper account.
• You can explain which Agent Builder controls are enforced and which are advisory.
• You do not use the Internal cloud bridge or Preview local API as an undocumented public integration.