inflowDocsLearning path

Ask [App] in published applications and the Home candidate

Published App Shell Chat is available when its publisher enables it and the user has current access to an approved source or action. Built-in Home Chat is a separate in-progress candidate: current Studio Home mounts Tasks and Applications, but no Home chat panel or launcher. A backend eligibility setting or API source does not create a user-facing Home Chat feature. Ask Inflow is separate from both.

For end users

Built-in Home Chat is not currently available to end users. Current source has no Home Chat feature flag, Home Chat endpoint or launcher on /home. Use the separate Published App Shell Chat feature inside an app when its chat launcher is available.

To chat within a published application, open it from Home Applications and use its launcher when available. Ask a question about that application's approved records, tasks or guidance. Review citations and open their sources through the supplied links; each open rechecks current access. An answer is guidance, and an unavailable or insufficient-evidence response does not establish a business fact.

Ask about your requests or your approval queue

Sara manages a team and also submits her own leave requests. These are two different questions in the same application:

Sara asks Evidence she needs Check before relying on the answer
“What is waiting for my approval?” Open tasks available to Sara in the app's approved task inbox. Compare the cited tasks with her ordinary work view. A partial list is not the whole queue.
“Is my leave request still awaiting approval?” Her intended request in the approved record source. Check the request key, person and current status. Sara's own approval tasks do not establish her request's status.
“Is LR-1005 still pending approval?” The permitted LR-1005 record. Confirm the key and status on its ordinary record page. A known key does not grant access.
“How many requests are awaiting approval?” A permitted record count of the requests in the waiting status. Requests are records: Sara's open tasks are not counted and are not shown as the source. Check the filter and access scope.
“How many leave requests are pending?” A permitted record count, when the source supports approved aggregates. Check the filter and access scope; this is not automatically a count of Sara's task inbox or the whole organization.

An empty task inbox does not mean that none of Sara's submitted requests is awaiting someone else's approval. Conversely, tasks waiting for Sara do not prove anything about her own request. If the question is ambiguous, say which request or whose work queue you mean. The same distinction applies to Arabic questions such as كم من طلباتي بانتظار الموافقة؟: these ask about the person's request records. The phrase بانتظار الموافقة describes a state; writing it does not approve anything.

If you are both a requester and an approver, Chat can still receive your open-task context while answering a request question. The wording of the question does not guarantee that its answer keeps these subjects separate. Before relying on a request status, verify that exact request in its ordinary permitted record view; a task label, count or missing task chip is insufficient.

The word “my” helps identify the question, but does not create an ownership filter or change permissions. Ask about the request's business reference when useful and verify its identity. The publisher's sources and the person's ordinary row, field and task permissions still determine what can be read.

Tell an unanswered question from an answer with limits

The following fictional teaching examples illustrate how to read a response; they are not captured conversations or promised response wording. Assume the app is connected to permitted leave information, with no salary evidence available for the question.

Response situation How to read it Next step
You ask for Huda's salary. The answer explains that the available information cannot answer it, even if it mentions an earlier leave request. App Chat can omit source chips when it recognizes that it cannot answer. A chip that remains, or a mention of an old request, does not turn leave information into salary evidence. Use the ordinary permitted source for that information, or ask its owner. Do not infer a salary or a reason for missing access.
You ask about LR-1005. A fresh permitted read answers that it is pending, then adds that salary is not included. The grounded request answer can retain its request citation despite the caveat. The chip supports the disclosed request facts, not the missing salary. Open the permitted request source and check the fact you need.
A successful permitted search finds no rejected requests under the stated filter. Zero matching results can be a data answer. It is different from a source error, a refusal or proof that no rejected request exists outside your permitted search scope. Check the scope and any incomplete-results notice.
A search finds LR-1005 but no additional matching requests in its scope. “I couldn't find any other requests” can accompany a grounded answer and its citation. It does not promise a complete organization-wide search. Check the returned record and the search limits.
The requested action does not apply to the selected record and no proposal is prepared. The not-applicable answer has no source chips. It does not report that an action ran. Open the ordinary permitted task or form, identify the intended target and review a valid proposal if offered.

Read what the answer actually establishes and what its sources cover. Do not assume every unsupported answer will hide its chips: current recognition does not guarantee the same display for every wording. Do not use the presence or absence of a chip as a task receipt. A grounded answer with a caveat may leave the current record available for a follow-up such as “Prepare an approval for LR-1005,” but that still requires an applicable authorized task, a reviewed target and explicit confirmation. If no clear target is available, name the intended task or use its ordinary form. Check an uncertain operation through its existing Check status path before attempting another action.

Read a record citation

When a cited record has a short, already-disclosed person name and one readable business key, its citation can be labeled with both so you can distinguish it from another record. For example, a leave request might appear as Leave requests · LR-1003 (Rana Khoury). When its locator is sealed or is a system identifier, the label can use the disclosed name without revealing that locator, such as Leave requests · Huda Nasser. These labels use values the record source already allowed the answer to disclose; they do not reveal hidden fields or grant permission to open the record.

An aggregate such as “3 requests are awaiting review” is a summary, not one named record. Its citation remains a source or summary label rather than borrowing a person's name. If you need to inspect one request, ask about that specific item and open its citation only if a permitted destination is offered. Access is checked again when you open it, so a label that was visible earlier does not guarantee the record is still available.

What you see What it means What to do
Leave request LR-1003 · Rana Khoury A record citation has a readable business key and a short disclosed name (one record reads in the singular). Check that this is the request you meant, then open it if the source link is available.
Leave requests · Huda Nasser The citation has a disclosed name, but its locator is sealed or otherwise not shown. Use the name and surrounding answer as context; do not infer or copy a hidden identifier.
Leave requests beside a count or grouped total The answer cites an aggregate rather than one person's record. Treat it as a summary and ask a narrower question if you need a particular record.

Counts, follow-up questions, and record labels

When App Chat answers a count question, the total alone does not identify which person or request matched. A named record needs current permission-checked evidence. A clear follow-up about the one record currently being discussed can use that context after a fresh check. A previously discussed record may also be named when a fresh permitted read confirms that it matches every supported exact-match filter of a selection read made for this question. It need not be among that read's first returned rows, but an old mention or a count alone is insufficient. If this relationship cannot be established, ask for a list or name the record; do not assume an earlier request belongs to the counted set. If Chat says I couldn't tell which records those are from this question's results. Ask me to list them, or name the record, ask it to list the matching records or mention the request you mean; do not infer a person from the total.

Task summaries omit machine-like internal case references. A linked-record lookup used for a task proposal follows a separate path and may supply its own record key to the answer and the proposal card; the task-summary filter does not guarantee that every such key is hidden. Readable business references such as LR-1005 or INV-2026-11-0001 may appear when the source allows them. If a proposal card shows a generated or unfamiliar key, do not approve based on that label. Open the ordinary permitted task or record view and confirm the target, or ask the application owner to identify it.

If a citation is unclear, unavailable, or no longer opens, use the application's ordinary record list or ask its owner for access. Do not treat the chat title, answer, or source chip as approval to make a business decision.

If App Chat cannot finish checking access to evidence for an answer, it may reply This took too long to check. Nothing was changed. Try again. The answer is withheld; this does not say that your access was removed. Depending on when the check timed out, the AI provider may already have received the turn even though you did not receive its answer. Wait briefly and retry once. If it happens again, open the normal permitted record or task view or ask the application owner to check the source. Do not treat an older answer as current evidence or submit an action based on it. A confirmed access denial is a separate result.

Published App Shells may offer task decisions or other approved actions. First inspect the proposed values, target, required fields and reason. A draft changes no business state. Only your explicit confirmation may execute the ordinary authorized action. If the target changed, review its current state and correct the draft. Keep an uncertain operation identity and use the provided reconciliation/ordinary-form path; do not submit a second fresh operation to guess whether the first completed.

When an approval proposal names a linked record

Suppose Sara asks, “Prepare an approval for Huda Nasser's leave.” App Chat may show a proposed approval card that names the linked leave record as LR-1005 and shows a disclosed field such as 5 days. When the server supplies that readable record reference, the card can show it even if the record's internal locator is sealed; the reference helps identify the target but does not grant permission to open it. The record and the fields shown on that card are part of the evidence for the answer, even if they came from a different read than the first record citation. Before showing the answer, Inflow checks that the same linked record and those fields are still readable and permitted for AI disclosure. The proposal remains a draft: Sara must inspect it and explicitly confirm before the normal task action runs.

After Sara approves, the conversation keeps the proposal's words and its "Update to Leave request LR-1005" line (when the decision is known), and adds a receipt below them. If the decision then fails after a comment was added, the card says the comment was added; do not add it again. When the decision updated the linked record, the receipt lists what the server actually wrote, for example Status changed Pending → Approved on Leave request LR-1005. A protected field is named without its values. A receipt read back later through Check status may omit these lines; open the record to confirm its current values.

If access to LR-1005 or AI disclosure of a shown field is revoked before that check, the answer and proposal are withheld. If Sara later opens the conversation after access has been revoked, the saved answer can appear as refused/withheld instead of repeating the earlier approval wording. This is an access safeguard; it does not reverse an action Sara already confirmed. To learn the task's current business state, open the ordinary permitted task or record view and verify it there. A hidden or refused old answer is not evidence that the task was approved or rejected.

Use a clear question and check both the target task and the linked record before confirmation. For example, if the card says Huda Nasser · LR-1005 · 5 days, make sure that is the request Sara intended; if the name, key or days are missing or do not match, stop and open the ordinary task form. Never infer a missing link from a name alone.

The proposed Home Chat action set includes task completion or comments, with a fresh task-authority check and explicit confirmation. This Home host has not been delivered in current Studio, so these are not actions available from Home Chat today.

Do not paste credentials or unrelated sensitive information into chat. App Chat does not automatically include unsaved form values. If your question depends on values that have not been saved, use the application's ordinary form or save through its normal workflow first; only ask about information from sources the publisher approved and you are currently allowed to read.

The published app remains usable when AI is off. Its chat launcher is omitted when that App Shell does not enable Chat or you lack an admitted source/action; an app link does not grant access. Built-in Home currently has no Chat launcher. The candidate Home setting does not affect published App Shell Chat.

For publishers

For INTERFACE_HELP, first save and publish the help screen. Then open the intended APPSHELL in Interface Designer and expand App Chat. Chat is disabled by default; select Enable App Chat. Optionally set Chat launcher label. Under Approved App Sources, choose Source type and Permitted source, optionally enter a Display label, then select Add source. Choose only from the available permitted options; the target is not free-form text. For Interface Help, the picker reflects the current designer document and Screen View permissions, not publication status, so verify separately that the selected help screen is published. Available source types are Record Type, Task Inbox, Guidance Document, Process Guidance, and Interface Help.

For an INTERFACE_HELP source, after publishing the help screen, select it from the available permitted options. Confirm independently that it is published. Save the App Shell and publish that new draft revision. Verify the intended app’s published settings and ordinary-reader access independently; importing JSON or saving a draft does not enable chat in a published application. These App Shell settings are independent of the /home operator setting.

When the App Shell itself is an INTERFACE_HELP source, citations identify an individual canvas page when that page has a valid URL stub. Opening a citation first rechecks current source access, then opens only that exact page if it is still among the reader's visible pages. A malformed, hidden, removed or stale page citation does not fall back to the default page. Existing shell-level citations retain their root behavior, and help from a bound screen continues to open the containing App Shell page.

Configure only the minimum relevant sources. A source binding adds a restriction; it grants no record, task, workspace or document permission. Records and source lineage must remain currently readable. Indexing and the environment's approved AI policy are separate prerequisites. A published help page alone does not prove retrievable evidence.

Under Approved App Actions, choose Action type and Permitted action, optionally enter a Display label, then select Add action. Add only an action with a supported target and handler, such as an approved task, record, process, or published-form action. The saved binding must require confirmation. Do not enable an action merely because its type is listed: qualify the target and handler for this App Shell. Save and publish the App Shell; draft-only changes do not change the published app.

Enable and review each intended application independently. Never blanket-enable every shell. Recheck ordinary reader behavior, denied sources, revoked source/citation/history, EN/AR, explicit confirmation, uncertain results and AI-off ordinary operation. Use the importable training example as a starting point.

For platform operators

Current source recognizes home-app-chat as a distinct provider purpose, but has no Home Chat feature flag, Home Chat API/controller or panel host. The purpose name alone does not wire Home Chat into the application. Do not change the published App Shell app-chat or Ask Inflow studio-agent defaults in an attempt to enable Home. If a Home Chat candidate is delivered later, provider credentials must use the server-side credential vault; never place them in custom_config.properties, a tracked file, browser state or evidence.

Support and current integration status

The published App Shell host uses /api/app-chat/apps/{appShellKey} and selectors based on its actual ${agentBase}.app-chat.toggle, .panel, .prompt, .send and .proposal.{id}.confirm controls. A Home-targeted route family and source bindings are a proposed design; current master has no /api/app-chat/home routes or Home panel selectors. Home sources cannot be authored in App Shell JSON.

No provider, Home runtime setting, or App Shell application was enabled by writing this manual. Published App Shell qualification and the Home host, coordinated database chain, indexed-source/provider qualification, publication/configuration, and ordinary-user evidence remain pending in the Home program. Consult the reuse contract and current execution record; source presence does not mean deployed or VERIFIED.

Inflow documentation · Published 19 November 2026

Screenshot captions retain their recorded build and capture context.

Type to search articles and their contents.