> ## Documentation Index
> Fetch the complete documentation index at: https://docs.flowdeck.studio/llms.txt
> Use this file to discover all available pages before exploring further.

# Validation

> Choose how FlowDeck agents check their work and share the results

Writing the change is only part of finishing it. Validation is the work an agent does
afterward to show that the result builds, passes the right tests, and behaves as expected.

FlowDeck always requires the smallest relevant tests and any checks the repository explicitly
defines as gates. The Validation settings control the extra runtime checks and evidence an agent
gathers after those required checks, so you can choose between a faster handoff and a more
comprehensive review.

## Why FlowDeck validates more

FlowDeck is built to let agents do more than edit files. An agent can use the project's saved
target to build and test, run the real app, exercise changed flows, inspect logs, and capture
evidence from simulators and devices. That makes it possible to catch problems a code diff or
clean build cannot reveal, such as clipped text, a broken interaction, or state that fails only
after the app is launched.

Balanced is the default: tests, build/launch, then one check of the changed flow and its
visible result. It checks both light and dark only for appearance-dependent styling changes;
otherwise it uses the current appearance. It checks changed loading/empty/error/permission
states, but does not add rotation, resizing, adjacent flows, or extra devices.

## Choose a starting point

The preset at the top of **Settings ▸ AI Workflows ▸ Validation** sets every option below it.
You can then change any option independently. When you do, the preset reads **Custom**. Choosing
a named preset again resets all five options to that preset.

> [card] Lean
>
> Runs relevant tests, confirms compilation and successful launch, then stops.
> Uses the saved target, no screenshots or recordings, and a brief report.

> [card] Balanced
>
> Adds one check of the changed flow and visible UI. Captures useful visual evidence.
> No video, rotation, or extra devices. This is the default.

> [card] Thorough
>
> Adds relevant edge cases, supported orientations/window sizes, and one additional
> target with a meaningful compatibility difference. Captures evidence during those
> checks and reports a compact results matrix.

## Customize each method

### Run the app

Controls how much runtime behavior the agent exercises after the required checks pass. The other controls independently select artifacts, targets, and reporting; preset names add no extra instructions. For documentation-only work, validate the documents. For libraries/CLIs, use their test or executable entry point rather than launching an unrelated app.

| Option | What the agent does |
| - | - |
| **Launch only** | After relevant tests pass, confirms compilation and successful startup. Does not navigate UI flows. |
| **Changed features** | Confirms compilation/startup, exercises the shortest changed flow once, and visually inspects affected UI. Checks light/dark only for colors, materials, shadows, or appearance-dependent styling; otherwise uses current appearance. Checks changed loading/empty/error/permission states. No rotation, resizing, content-size sweeps, or adjacent flows. |
| **Extended checks** | Adds applicable failure, empty, long-content, and permission states. Repeats an interaction once if lifecycle/state changed. For UI, checks light/dark and supported portrait/landscape on mobile or compact/wide windows on desktop. Lists applicable checks, avoids a Cartesian product, and stops when they pass. |

### Screenshots

Controls still-image evidence. It does not decide whether the agent runs tests or checks the app.

| Option | What the agent does |
| - | - |
| **No screenshots** | Does not capture screenshots and describes any visible result briefly. |
| **Only when necessary** | One final image for meaningful visual changes; one per appearance only when both were checked. Skips copy-only work unless layout changes. Never adds runtime checks. |
| **Always** | One final image per visually distinct configuration already checked. Before/after only when needed to explain the result. |

### Screen recordings

Controls video evidence for behavior that a still image cannot explain well.

| Option | What the agent does |
| - | - |
| **No recordings** | Does not record video. |
| **Only when necessary** | Records selected checks only when motion or timing cannot be proven with stills. Does not add or repeat flows. |
| **Always** | One short clip per motion or multi-step flow already selected. Captures during the check; skips when no such flow exists. |

### Additional devices

Controls whether the agent repeats relevant runtime checks beyond the saved target.

| Option | What the agent does |
| - | - |
| **Saved target only** | Uses only the target selected in FlowDeck unless the feature or task requires another one. |
| **Only when necessary** | Adds one relevant simulator or device when there is a meaningful compatibility risk. |
| **Multiple targets** | Saved target plus one additional relevant screen-size, OS, or platform. Repeats only difference-sensitive checks. More targets require explicit task scope; unavailable or irrelevant targets are reported as a limit. |

### Final report

Controls how much validation detail appears in the agent's final response. It changes the report,
not the checks that are required.

| Option | What the agent does |
| - | - |
| **Brief** | Names the checks that passed, reports the test count, and includes only essential evidence. |
| **Standard** | Summarizes tests, runtime checks, evidence, concrete outcomes, and any remaining gap. |
| **Detailed** | Provides a target-by-target matrix with test counts, flows and states exercised, artifacts, results, and remaining gaps. |

## What never changes

Validation preferences never let an agent skip required tests, repository gates, safety rules,
permission requirements, or honest reporting. An agent must not claim a check or artifact it did
not produce. A setting can reduce extra runtime work or evidence, but it cannot turn a failed
check into completed work. Test-driven development controls whether a test is written and run
before implementation; it does not turn that focused red-green cycle into a full test suite,
UI sweep, or device matrix.

Validation stops when the selected checks pass and the changed behavior is proven. If a check
fails, the agent investigates that failure and repeats the smallest affected scope. A failure is
not permission to start a general compatibility sweep.

Changes take effect when a session is next configured or restored, and for Team workers hired
after the change. They do not rewrite the instructions of an agent that is already running.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.