[BREAKING] FEAT: Add runtime generation configuration

doc/code/registry/0_registry.md

@@ -129,14 +129,16 @@ reconstruction:
 }
 ```
 
-Use `source.params` for overrides; do not also supply `params`. The backend
+Use `source.params` to override constructor defaults; do not also supply `params`. The backend
 keeps the source's other constructor inputs, including authentication, on the
 server. It does not change the source. Missing, changed, or ambiguous sources
 produce an error. If a source was renamed, a unique matching hash can resolve it.
 An optional `effective_hash` checks the reconstructed object's identity.
-Registered descriptors report `reconstructable` and, for targets,
-`supports_temperature_override`. Unsupported inputs can also report
-`reconstruction_error`.
+Converter descriptors expose `reconstructable` and `reconstruction_error`;
+target runtime support is in `capabilities`, with bounds in the separate
+`temperature_range` field, independent of reconstruction. Use
+[runtime generation configuration](../targets/0_prompt_targets.md#runtime-generation-configuration)
+for per-request overrides instead of rebuilding a target.
 Constructor inputs that are not in the registry's build contract are not
 discarded silently: reconstruction is rejected. Such sources can still be used
 as registered objects.

doc/code/targets/0_prompt_targets.md

@@ -57,6 +57,31 @@ Target-side rate limiting remains independent and continues to use `limit_reques
 `_send_prompt_to_target_async(*, normalized_conversation: list[Message]) -> list[Message]` instead of
 overriding `send_prompt_async`.
 
+## Runtime generation configuration
+
+Use `RequestGenerationConfig` in a current request's `MessagePiece.prompt_metadata`
+to override temperature without changing shared target defaults:
+
+```python
+from pyrit.models import MessagePiece, RequestGenerationConfig
+
+piece = MessagePiece(
+    role="user",
+    original_value="Describe this scene.",
+    prompt_metadata=RequestGenerationConfig(temperature=0).to_metadata(),
+)
+```
+
+Check `target.capabilities.supports_temperature` and `target.temperature_range`
+before choosing a value. Bounds are inclusive; `None` means unspecified bounds,
+not lack of support. Missing or null temperature preserves existing defaults.
+Invalid, conflicting, or unsupported overrides fail before provider I/O.
+
+Read saved request and response evidence with
+`RequestGenerationInfo.from_metadata(metadata=piece.prompt_metadata)` to inspect the
+requested temperature, resolved value, and source. Unknown provider defaults remain
+`None`. See [Chat temperature](../../gui/0_gui.md#temperature) for the UI.
+
 ## Chat-style targets vs general targets
 
 A `PromptTarget` is a generic place to send a prompt. With PyRIT, the idea is that it will eventually be consumed by an AI application, but that doesn't have to be immediate. For example, you could have a SharePoint target. Everything you send a prompt to is a `PromptTarget`. Many attacks work generically with any `PromptTarget` including `RedTeamingAttack` and `PromptSendingAttack`.
@@ -249,12 +274,16 @@ A `TargetConfiguration` composes three concerns:
 | `supports_system_prompt`         | The target natively supports a system-role message.                                                                                      |
 | `supports_json_output`           | The target supports a "json" response format that guarantees valid JSON output.                                                          |
 | `supports_json_schema`           | The target supports constraining output to a caller-provided JSON schema.                                                                |
+| `supports_temperature`           | The adapter and configured model mode support per-request temperature overrides.                                                        |
 | `input_modalities`               | The set of input modality combinations the target accepts (e.g. `{text}`, `{text, image_path}`).                                         |
 | `output_modalities`              | The set of output modality combinations the target produces.                                                                             |
 
 Each target class defines defaults; instances can override individual capabilities when they depend on deployment configuration (e.g. `HTTPTarget`, `PlaywrightTarget`).
 
 For well-known underlying models, you can look up a profile with `get_known_capabilities(underlying_model="gpt-4o")` from `pyrit.prompt_target`.
+Profiles alone do not determine runtime temperature support; check the instance's
+capabilities and separate `temperature_range` property as described under
+[Runtime generation configuration](#runtime-generation-configuration).
 
 Tool-call history uses the existing `function_call` and `function_call_output` input
 modalities. These describe acceptance of prior calls and results, not tool execution.

doc/code/targets/6_1_target_capabilities.ipynb

@@ -139,7 +139,11 @@
     "`get_default_configuration(underlying_model=...)` returns a richer profile from\n",
     "`get_known_capabilities` — for example, `gpt-5` gains `supports_json_schema=True`\n",
     "and other models pick up the right modality combinations automatically. Unknown models fall back to\n",
-    "the class default."
+    "the class default.\n",
+    "\n",
+    "Runtime temperature support depends on the configured target instance, not just its model profile.\n",
+    "Check `target.capabilities.supports_temperature`; see\n",
+    "[Runtime generation configuration](./0_prompt_targets.md#runtime-generation-configuration) for usage and bounds."
    ]
   },
   {

doc/code/targets/6_1_target_capabilities.py

@@ -98,6 +98,10 @@
 # `get_known_capabilities` — for example, `gpt-5` gains `supports_json_schema=True`
 # and other models pick up the right modality combinations automatically. Unknown models fall back to
 # the class default.
+#
+# Runtime temperature support depends on the configured target instance, not just its model profile.
+# Check `target.capabilities.supports_temperature`; see
+# [Runtime generation configuration](./0_prompt_targets.md#runtime-generation-configuration) for usage and bounds.
 
 # %%
 class_default = OpenAIChatTarget._DEFAULT_CONFIGURATION.capabilities

doc/gui/0_gui.md

@@ -99,6 +99,16 @@ corresponding Chat route while retaining query parameters and fragments.
 
 ![Chat workspace with a prompt and target response](images/chat.png)
 
+### Temperature
+
+On supported targets, use the thermometer below **Send** to adjust temperature
+for subsequent messages. **Use target default** clears the override; shared
+defaults and saved messages are unchanged. Select a message's temperature footer
+to inspect its saved settings.
+
+See [Runtime generation configuration](../code/targets/0_prompt_targets.md#runtime-generation-configuration)
+for SDK usage.
+
 ## Scanner
 
 Scanner has **Launch** (`/scanner/launch`) and **History** (`/scanner/history`) tabs.
@@ -232,8 +242,7 @@ the original launch configuration is available. It skips completed objectives
 with non-error outcomes, including `failure`. Retrying an unsuccessful technique
 is a separate setting in `technique.execute`.
 
-Compound-attack hierarchy browsing is a separate GUI follow-up. The
-[backend hierarchy API](../../pyrit/backend/README.md#scenario-result-hierarchy)
+The [backend hierarchy API](../../pyrit/backend/README.md#scenario-result-hierarchy)
 provides paged roots and ordered child expansion without clipping run statistics.
 Conversation reads expose main and pruned conversations only; adversarial and
 preparation contexts remain internal. Transcript bodies load only for a selected

frontend/src/App.labels.test.tsx

@@ -351,7 +351,7 @@ describe('Shared new run labels', () => {
   it('hosts Chat controls below the labels and above the objective, and removes them when navigating away', async () => {
     const user = userEvent.setup()
     jest.mocked(targetsApi.listTargets).mockResolvedValue({
-      items: [{ ...TARGET, supports_temperature_override: true }],
+      items: [{ ...TARGET, capabilities: { ...TARGET.capabilities, supports_temperature: true } }],
       pagination: { limit: 200, has_more: false },
     })
     renderApp('/attack/chat')
@@ -366,9 +366,13 @@ describe('Shared new run labels', () => {
     await waitFor(() => expect(targetPicker).toBeEnabled())
     await user.selectOptions(targetPicker, 'test_target')
     expect(targetPicker).toHaveValue('test_target')
-    const temperature = within(settings).getByRole('spinbutton', { name: 'Temperature' })
+    expect(within(settings).queryByRole('spinbutton', { name: 'Temperature' })).not.toBeInTheDocument()
+    const generation = screen.getByRole('button', { name: /^Adjust temperature/ })
+    await user.click(generation)
+    const temperature = screen.getByRole('textbox', { name: 'Temperature' })
     await user.type(temperature, '0.7')
-    expect(temperature).toHaveValue(0.7)
+    expect(temperature).toHaveValue('0.7')
+    await user.click(generation)
     expect(within(settings).getByRole('button', { name: 'Edit Conversation' })).toBeEnabled()
     expect(within(screen.getByTestId('chat-area')).queryByRole('group', { name: 'Chat controls' }))
       .not.toBeInTheDocument()
@@ -382,12 +386,15 @@ describe('Shared new run labels', () => {
 
     await user.click(screen.getByRole('button', { name: 'Home', exact: true }))
     expect(screen.queryByRole('group', { name: 'Chat controls' })).not.toBeInTheDocument()
+    expect(screen.queryByRole('button', { name: /^Adjust temperature/ })).not.toBeInTheDocument()
     expect(screen.getAllByTestId('labels-bar')).toHaveLength(1)
     await user.click(screen.getByRole('button', { name: 'Scanner', exact: true }))
     expect(screen.queryByRole('group', { name: 'Chat controls' })).not.toBeInTheDocument()
     await user.click(screen.getByRole('button', { name: 'Attacks', exact: true }))
     expect(screen.getAllByRole('group', { name: 'Chat controls' })).toHaveLength(1)
     expect(within(screen.getByRole('group', { name: 'Conversation actions' })).getByRole('switch')).toBeChecked()
+    await user.click(screen.getByRole('button', { name: /^Adjust temperature/ }))
+    expect(await screen.findByRole('textbox', { name: 'Temperature' })).toHaveValue('0.7')
   })
 
   it('preserves an edit made before the backend defaults arrive', async () => {

frontend/src/components/Chat/ChatInputArea.test.tsx

@@ -121,6 +121,30 @@ describe("ChatInputArea", () => {
     expect(onToggleConverterPanel).toHaveBeenCalledTimes(1);
   });
 
+  it("should place the compact temperature control immediately after Send without losing converter or Undo actions", async () => {
+    const user = userEvent.setup();
+    const onUndo = jest.fn();
+    render(<TestWrapper><ChatInputArea {...defaultProps} initialText="Hello"
+      activeTarget={makeTarget({ target_registry_name: "temperature-target", capabilities: buildCapabilities({ supports_temperature: true }) })}
+      generationDraft={{ temperature: "0" }} onGenerationDraftChange={jest.fn()}
+      undoLastSend={{ isUndoing: false, onUndo }} convertedValue="Converted"
+    /></TestWrapper>);
+    const prompt = screen.getByTestId("prompt-editor");
+    const send = within(prompt).getByRole("button", { name: "Send message" });
+    const temperature = within(prompt).getByRole("button", { name: "Adjust temperature: 0" });
+    const clearConversion = within(prompt).getByRole("button", { name: "Clear conversion" });
+    expect(send.compareDocumentPosition(temperature) & Node.DOCUMENT_POSITION_FOLLOWING).toBeTruthy();
+    expect(temperature.compareDocumentPosition(clearConversion) & Node.DOCUMENT_POSITION_FOLLOWING).toBeTruthy();
+    act(() => send.focus());
+    await user.tab();
+    expect(temperature).toHaveFocus();
+    expect(temperature).toHaveTextContent("0");
+    await user.click(within(prompt).getByRole("button", { name: "Undo last send" }));
+    expect(onUndo).toHaveBeenCalledTimes(1);
+    await user.click(within(prompt).getByRole("button", { name: "Toggle converter panel" }));
+    expect(defaultProps.onToggleConverterPanel).toHaveBeenCalledTimes(1);
+  });
+
   it("should place Undo directly after the converter button inside the prompt box without opening the panel", async () => {
     const user = userEvent.setup();
     const onUndo = jest.fn();

frontend/src/components/Chat/ChatInputArea.tsx

@@ -3,13 +3,14 @@ import { useState, useCallback, useEffect, useLayoutEffect, useRef, forwardRef,
 import { Button, Caption1, Popover, PopoverSurface, PopoverTrigger, Tooltip, Text } from '@fluentui/react-components'
 import { SendRegular, DismissRegular, InfoRegular, CopyRegular, WarningRegular, SettingsRegular, ArrowSyncRegular } from '@fluentui/react-icons'
 
-import type { AttackTargetResolutionStatus, ChatSendOutcome, ConvertedFileChip, MessageAttachment, MultiSendOptions, PieceConversion, TargetInstance, UndoLastSendAction } from '@/types'
+import type { AttackTargetResolutionStatus, ChatSendOutcome, ConvertedFileChip, MessageAttachment, MultiSendOptions, PieceConversion, RequestGenerationDraft, TargetInstance, UndoLastSendAction } from '@/types'
 import { isTargetResolutionBlocking } from '@/utils/targetIdentity'
 
 import { usePromptEditorStyles } from './PromptEditor.styles'
 import PromptEditor from './PromptEditor'
 import SystemPromptSetup from './SystemPromptSetup'
 import MultiSendSettings from './MultiSendSettings'
+import GenerationSettings from './GenerationSettings'
 import { attachmentsFromFiles, PIECE_TYPE_TO_DATA_TYPE, withDraftIdentity } from './converterTypes'
 
 const DISABLED_WARNING_HOVER_DELAY_MS = 700
@@ -107,6 +108,8 @@ export interface ChatInputAreaHandle {
 }
 
 interface ChatInputAreaProps {
+  generationDraft?: RequestGenerationDraft
+  onGenerationDraftChange?: (value: RequestGenerationDraft) => void
   undoLastSend?: UndoLastSendAction
   repeatDisabledReason?: string
   onSend: (
@@ -160,7 +163,7 @@ const ChatInputArea = forwardRef<ChatInputAreaHandle, ChatInputAreaProps>(functi
   convertedValue, onClearConversion, onClearAllConversions, onConvertedValueChange,
   converterOutputDataTypes = [], mediaConversions = [], onClearMediaConversion, convertedFileChip,
   onClearConvertedFileChip, showSystemPrompt = false, supportsSystemPrompt = false, systemPrompt = '', onSystemPromptChange,
-  repeatDisabledReason, undoLastSend,
+  repeatDisabledReason, undoLastSend, generationDraft = {}, onGenerationDraftChange,
   initialText = '', initialAttachments = [], initialTextMetadata,
 }, ref) {
   const styles = usePromptEditorStyles()
@@ -325,6 +328,8 @@ const ChatInputArea = forwardRef<ChatInputAreaHandle, ChatInputAreaProps>(functi
                 <Button className={styles.sendButton} appearance="primary" icon={<SendRegular />}
                   onClick={() => { void handleSend() }} disabledFocusable={!canSend} aria-label="Send message" data-testid="send-message-btn" />
               </Tooltip>
+              {onGenerationDraftChange && <GenerationSettings value={generationDraft}
+                onChange={onGenerationDraftChange} target={activeTarget} disabled={disabled} />}
               {convertedValue != null && <Tooltip content="Clear conversion" relationship="label">
                 <Button appearance="subtle" className={styles.clearConversionButton} icon={<DismissRegular />}
                   onClick={onClearConversion} data-testid="clear-conversion-btn" />

frontend/src/components/Chat/ChatWindow.styles.ts

@@ -151,16 +151,6 @@ export const useChatWindowStyles = makeStyles({
     alignItems: 'center',
     gap: tokens.spacingHorizontalXS,
   },
-  temperatureField: {
-    gridTemplateColumns: 'max-content 5rem',
-    columnGap: tokens.spacingHorizontalNone,
-    alignItems: 'center',
-    flexShrink: 0,
-  },
-  temperatureInput: {
-    width: '5rem',
-    minWidth: 0,
-  },
   sharedTarget: {
     maxWidth: '240px',
     [NARROW_VIEWPORT_QUERY]: {