task_plan.md

Task Plan: Upgrade the repository to .NET 10 Windows x64

Current Follow-up: Add OpenAI-Compatible connections

Phase 1: Contract and persistence

• Add a distinct OpenAI-Compatible provider type and defaults
• Persist a Base URL per connection without changing official OpenAI or Codex behavior
• Status: complete

Phase 2: UI and routing

• Expose Base URL, optional Bearer key, and free model ID in the connection editor
• Route tests and Assistant requests through LLMTornado Custom
• Keep the test action aligned with the model column
• Status: complete

Phase 3: Verification

• Add focused provider, persistence, routing, and XAML regressions
• Normalize changed C#/AXAML once and perform all permitted static checks
• Build/test only after explicit user authorization required by repository rules
• Status: implementation complete; build/test awaiting explicit authorization

Completed Follow-up: Discard and connection-editor cancel behavior

Phase 1: Diagnose and correct

• Trace the unsaved-changes dialog result through CanCloseAsync and OnClosing
• Confirm Discard leaves the editor dirty and therefore reopens the dialog on the recursive close
• Clear connection-edit state when Discard is chosen
• Add the incident-specific regression, normalize changed C# files once, then build and test
• Verify in the real UI that Discard keeps Settings open and exits the connection editor
• Add a red Abbrechen button to the right of Speichern, bind it to CancelConnectionCommand, and verify it in the real UI
• Prompt on Abbrechen only after actual user changes; leave a pristine new connection prompt-free
• Status: complete

Completed Follow-up: Add Codex as a connection type

• Add a distinct Codex subscription connection option without conflating it with OpenAI API access
• Route Codex connections through the shared subscription catalog and selected connection in the Assistant
• Hide API-only connection fields for Codex and retain them for API providers
• Enable API connection testing only with an effective key, while preserving keyless local providers
• Show Codex connection testing only while the subscription is authenticated
• Add focused regressions, build against the local LLMTornado fork, and validate the states in the real UI
• Status: complete

Current Follow-up: Align with the LLMTornado PR branch

Phase 1: Verify and adapt

• Verify the clean local LLMTornado PR branch and compare its public Codex API with the prior integration commit
• Identify Toolkit overrides that are now owned by LLMTornado
• Remove the obsolete Toolkit protocol-version override and update focused regression coverage
• Normalize changed C# files once, then build and test against the local LLMTornado project
• Status: complete

Current Follow-up: Compact provider title and Codex model row

Phase 1: Implement and verify

• Locate the provider-list title localization and Codex model item template
• Rename the provider-list title to the localized equivalent of "Providers"
• Render Codex model choices as one display-name line
• Add focused regression coverage, normalize changed C#/AXAML once, then build and validate the UI
• Status: complete

Current Follow-up: Persist OpenAI access mode and keep speed selection visible

Phase 1: Diagnose restart and selection state

• Verify how the saved OAuth credentials, active access mode, and Settings selectors are restored
• Trace the blank Fast selection to synchronous collection rebuilding during ComboBox selection
• Status: complete

Phase 2: Implement and verify

• Persist and restore the selected OpenAI access mode, including silent subscription reconnection
• Defer access-state projection so reasoning/speed selections are not rebuilt reentrantly
• Add restart/selection regressions, normalize changed source files without Perl, then build and test
• Status: complete

Current Follow-up: Subscription connect state, effort, and speed

Phase 1: Inspect available metadata and turn options

• Verify the current LLMTornado Codex model metadata and turn-option surfaces for reasoning effort and speed
• Inspect the live model-catalog schema narrowly for optional speed/service-tier metadata
• Status: complete

Phase 2: Implement and verify

• Make Connect a Primary action that is disabled while already connected
• Add model-dependent effort and speed settings where supported and route selections into subscription turns
• Restore the configured AI connection selector in the Assistant without restoring authentication controls
• Add incident-specific regressions, verify encoding without Perl, then build and test
• Status: complete

Current Follow-up: OAuth subscription model discovery failure

Phase 1: Diagnose the model endpoint

• Trace the exact LLMTornado OAuth model request, response handling, and app-state mutation
• Reproduce the failure without exposing stored credentials and identify the HTTP/API mismatch
• Status: complete

Phase 2: Correct and verify

• Represent authentication versus catalog errors correctly and provide a working catalog retry
• Restore dynamic subscription model discovery for direct OAuth
• Add incident-specific regressions, normalize once, then build and test
• Status: complete

Current Follow-up: OAuth completion must update OpenAI Settings

Phase 1: Diagnose the browser return path

• Reproduce the state transition in credential-free tests and trace LLMTornado OAuth completion
• Determine whether completion is lost in the browser callback, session adapter, shared manager, or Settings projection
• Status: complete; callback and shared state succeeded, presentation is misleading

Phase 2: Correct and verify

• Implement the smallest reliable completion/state-update fix
• Make authenticated state unmistakable and update command enablement
• Remove authentication, subscription model selection, and login/logout controls from the AI Assistant UI
• Add incident-specific regressions, normalize changed source/XAML once, then build and test
• Status: complete

Current Follow-up: ChatGPT subscription access through local LLMTornado

Correction: expose subscription authentication in OpenAI Settings

• Inspect the current OpenAI provider page, Settings viewmodel composition, launcher wiring, and existing Settings tests
• Add an explicit API/Codex authentication section to the OpenAI provider page without changing API-key behavior
• Share the existing app-scoped subscription state, dynamic models, login/logout, and mode switching with Settings
• Add focused Settings regressions, normalize changed source/XAML once, then build and test
• Status: complete

Phase 1: Inspect and establish local dependency

• Verify the LLMTornado checkout, commit, repository rules, solution/project paths, and supported target frameworks
• Attempt the D:\github\LLMTornado\src\LlmTornado.slnx Release build before application changes (blocked by the solution's Docs project because npm run build has no matching package script)
• Trace the consuming app's target framework, LLMTornado package reference, DI, authentication UI, model state, chat/image flows, and tests
• Status: complete with documented upstream Docs build failure

Phase 2: Design and implement

• Add mutually exclusive conditional PackageReference/ProjectReference selection using UseLocalLlmTornado and LlmTornadoProject
• Encapsulate API-key, Codex app-server, and direct OAuth modes behind app-owned capability-aware abstractions
• Keep API-key text/image behavior intact while adding subscription authentication, dynamic models, streaming threads, follow-ups, logout, and correct session disposal
• Update the Avalonia UI so image generation is available only for OpenAI API access and the app-server requirement is clear
• Status: complete

Phase 3: Verify

• Add focused tests for newly introduced app state and routing without real credentials
• Normalize changed .cs, .csproj, and .axaml files once at iteration end
• Build the full app with the local LLMTornado reference and prove the resolved local project is used
• Run all credential-free unit tests sequentially and record any interactive checks still pending
• Status: complete with documented pre-existing localization test-order failure

Constraints

• Do not start a real browser OAuth login without explicit user approval.
• Never log access, ID, or refresh tokens.
• Preserve all pre-existing worktree changes and avoid broad rewrites.
• Use native Git Bash and RTK; invoke dotnet only through the requested documented build/test paths.

Errors Encountered

Error	Attempt	Resolution
LLMTornado Release solution build failed in LlmTornado.Docs.csproj because npm run build reported a missing script	1	Retained the full log and did not run an unauthorized alternate build; inspect the documented Docs target and continue read-only architecture analysis
Combined ClientApp/Codex path check did not identify which expected path was absent	1	Stop the combined shape; verify each path independently and enumerate the Docs package files narrowly
RTK-proxied sed failed with Bad locale: LC_ALL=C.UTF-8 while resuming plan files	1	Do not retry the same command shape; use RTK-proxied head for the required short plan refresh
RTK could not resolve Git Bash awk while reading a source range	1	Do not retry with awk; compose the bounded range from RTK-proxied head and tail
Git Bash path conversion stripped the leading slash from MSBuild /p: arguments, causing MSB1008 before build execution	1	Rebuild the argument array with MSYS path conversion disabled for the native dotnet invocation
The corrected MSBuild invocation combined native properties with a POSIX solution path, causing MSB1001 before compilation	1	Use native Windows paths for the solution, local project, and log while path conversion is disabled
A broad type: project assets search produced excessive unrelated output	1	Query the exact LlmTornado.csproj entry and its surrounding schema only
A multi-target configuration search included an unverified, absent LLMTornado Directory.Build.props path	1	Verify each path first; no such LLMTornado props file exists, so inspect the verified project and repository props separately
Piping the full LLMTornado log search into head caused an expected RTK broken-pipe diagnostic	1	Use exact output-path patterns without truncating an active RTK pipeline
AdditionalProperties=Configuration=Release evaluated correctly but the external solution reference still reused Debug output	1	Also set the ProjectReference's explicit SetConfiguration metadata used by solution builds
Querying a solution with MSBuild -getProperty failed with MSB1063	1	Inspect the verified SDK targets and project-level evaluated items; solution files do not support that query
Explicit SetConfiguration metadata was overwritten and LLMTornado still resolved to Debug	1	The SDK target shows solution builds unset parent configuration for unassigned external references; disable that behavior only in the local-reference app project
The app project built LLMTornado in Release, but the test project's transitive graph invoked the external project again in Debug	1	Apply the conditional parent-configuration retention repository-wide through Directory.Build.props so every local-reference graph node uses Release
The first full test invocation could not write its C:\tmp log inside the sandbox	1	Rerun the same authorized test entry point outside the sandbox
Full test runs each failed one existing localization test because global culture state leaked between tests	2	Both exact failing tests passed in isolation; all 17 focused AI tests passed, so leave unrelated localization behavior unchanged
The first correction full-suite run timed out in two DNS tests under restricted networking and hit the known order-dependent localization assertion	1	The localization test passed alone, the DNS tests passed with network access, and the repeated full network-enabled suite passed all 220 tests
Read-only LLMTornado Git verification hit sandbox dubious-ownership protection	1	Repeat the same status and HEAD checks under the repository owner's approved context
Disabling MSYS conversion also left the POSIX solution path unchanged, causing MSB1001 before compilation	2	Use only verified native Windows paths in the conversion-disabled argument array
Combined English/German localization patch assumed the wrong capitalization for the English test-connection value	1	No localization file was changed; use the exact verified key line as the only patch context
Combined OAuth UI correction patch included BOM-sensitive first-line context in AiAssistantPanel.axaml.cs	1	No file was changed; split the correction into smaller file-scoped patches and start code-behind edits after verified non-first-line context
Combined findings/progress note patch used progress lines as findings context	1	No file was changed; update each planning file with its own verified section context
Final combined planning update again used progress lines as findings context	1	No file was changed; apply final findings and progress updates as separate file-scoped patches

Current Follow-up: Editable searchable connection model picker and required name

Phase 1: Inspect

• Trace the existing model field, provider model source, free-text persistence, and save/test paths
• Verify Avalonia 12.1 AutoCompleteBox filtering and popup APIs
• Trace connection-name validation and dirty/save state
• Status: complete

Phase 2: Implement

• Keep arbitrary model IDs editable while opening/filtering the known model list incrementally
• Limit the suggestion popup to 20 visible model rows
• Preserve model badges, keyboard selection, provider changes, testing, and persistence
• Give an empty required connection name a red error border
• Status: complete

Phase 3: Verify

• Add focused regression coverage
• Perform a critical source review
• Normalize changed .cs and .axaml files once
• Run permitted build/tests and restart only when the running app no longer locks output
• Status: complete

Validation note

• The first build exposed an Avalonia 12 event-argument mismatch; the focus handler now uses the routed event type.
• Final build: 16 projects, 0 warnings, 0 errors.
• Final tests: 209 passed.

Current Follow-up: Disabled KI-Gen tooltip and tooltip/delete contrast

Implementation

• Show the existing connection-required KI-Gen tooltip while the button is disabled
• Keep per-row delete buttons on the semantic error color during hover
• Give application tooltips contrasting Daisy neutral surface/content colors
• Preserve dynamic theme switching and accessibility metadata
• Status: complete

Verification

• Normalize changed .cs and .axaml files once for this iteration
• Build with 0 warnings and 0 errors
• Run focused AI Assistant and Settings regression tests
• Start the current win-x64 output without monitoring or CUA
• Status: complete

Validation note

• The first build attempt was blocked by the previously launched app holding output DLLs. After the user closed it, the rebuild succeeded.
• The full suite passed 207 of 208 tests; the same unrelated localization event test remains order-dependent. All 23 focused tests passed.

Current Follow-up: Resizable and independently filtered provider model lists

Phase 1: Inspect

• Trace the two search bindings, projected collections, row templates, and current layout rows
• Verify existing selection and persistence paths remain shared only where intended
• Status: complete

Phase 2: Implement

• Add a real GridSplitter between the text and image model areas
• Ensure each search property filters only its matching collection
• Show each row's delete action only on row hover or keyboard focus
• Preserve row selection, clickability, automation metadata, and persisted model data
• Status: complete

Phase 3: Verify

• Add or adjust focused source/viewmodel regression coverage
• Normalize changed .cs and .axaml files once at iteration end
• Perform static validation and a critical self-review without CUA
• Build successfully and run focused regression tests
• Status: complete

Validation note

• The full 207-test run had one unrelated failure in ToolkitLocalization_CultureChanged_FiresEvent; the same test passed in isolation, and all 13 focused Settings tests passed.

Current Follow-up: Use DaisyButton throughout Settings

Phase 1: Inspect

• Verify the actual Flowery DaisyButton API and theme variants
• Inventory XAML, navigation, footer, page, and code-created dialog buttons
• Confirm the final requirement: explicit suitable non-Ghost variant on every button
• Status: complete

Phase 2: Implement

• Use DaisyButton for all Settings navigation actions
• Retain all existing page/footer DaisyButtons and make every variant explicit
• Replace code-created save/discard/cancel buttons with DaisyButton
• Preserve commands, enabled states, accessibility IDs, selection, and layout
• Add focused regression coverage
• Status: complete

Phase 3: Verify

• Normalize changed .cs and .axaml files
• Run documented build and tests sequentially
• Start the updated win-x64 application and confirm its window appears
• Status: complete

Current Follow-up: Fix unresponsive AI Settings toolbar button

Phase 1: Diagnose

• Inspect the actual toolbar enablement and hit-testing state
• Trace the routed Click handler through MainWindow to SettingsWindow
• Inspect the initial Settings navigation selection
• Status: complete

Phase 2: Correct

• Replace the custom themed toolbar control with a normal Avalonia Button
• Make enablement, hit-testing, focus, and automation semantics explicit
• Select AI Connections through the Settings ListBox itself
• Strengthen focused regression coverage
• Status: complete

Phase 3: Verify

• Normalize changed .cs and .axaml files once at iteration end
• Validate XAML parsing, scoped diffs, and source invariants
• Run build and tests after the user permits closing/rebuilding the running app
• Status: complete within the running-app constraint

Current Follow-up: Always-available AI Settings toolbar button

Phase 1: Inspect

• Locate the image-editor toolbar and existing AI Settings dialog path
• Confirm the Settings navigation target for AI connections
• Identify unrelated worktree changes that must remain untouched
• Status: complete

Phase 2: Implement

• Add an icon-only AI Settings button between AI Gen and Sessions
• Keep the button enabled independently of images, providers, and connections
• Open Settings directly on AI Connections
• Add localized tooltip/accessibility text and focused source coverage
• Status: complete

Phase 3: Verify

• Normalize changed .cs and .axaml files once at iteration end
• Build the solution and run all tests sequentially
• Start the updated application and verify the new toolbar icon is visible
• Status: complete

Current Follow-up: Modernize Settings and separate provider model catalogs

Phase 1: Analyze current UI and prototype

• Inspect the prototype and record its navigation/model-catalog structure
• Trace the current Settings XAML, viewmodel state, localization, and tests
• Identify unrelated worktree changes that must be preserved
• Status: complete

Phase 2: Implement focused Settings redesign

• Replace the top tab strip with clear Settings page navigation
• Present text models and image-generation models as separate sections on KI-Anbieter
• Give each model section independent search/filter state
• Preserve provider, connection, and model-selection behavior
• Update localization and focused regression coverage where supported
• Status: complete

Phase 3: Verify

• Normalize changed .cs and .axaml files once at iteration end
• Run user-permitted build and tests sequentially
• Visually verify every Settings page and the two model searches
• Status: complete

Current design findings

• The prototype uses persistent left navigation with a single page content surface.
• KI-Anbieter contains distinct Textmodelle and Bildmodelle cards, each with its own search.
• Current on-disk work already modifies Settings XAML and viewmodel files; those edits must be separated from this task before patching.

Goal

Increase the application version, move every project from .NET 8 to .NET 10 Windows, constrain runtime assets and execution to win-x64, and keep text-to-image chat usable without a loaded source image, with restore/build/test closure.

Current Phase

Settings persistence regression protection: complete

Current Follow-up: Protect settings from destructive overwrite

Phase 1: Diagnose

• Confirm the overwritten file contains defaults and zero AI connections
• Trace direct non-atomic writes, fallback-to-default behavior, and startup locale saves
• Identify stale-process overwrite risk
• Status: complete

Phase 2: Implement

• Replace direct writes with atomic file replacement
• Add three rotating backup slots with UTC timestamps
• Restore the newest valid backup after invalid JSON
• Preserve corrupt/conflicting snapshots and reject stale-process writes
• Prevent AI synchronization errors from discarding loaded settings
• Status: complete

Phase 3: Verify

• Add backup rotation, truncated-file recovery, older-backup fallback, first-start, locale-save, stale-writer, concurrent-writer, and abandoned-mutex regression tests
• Normalize changed C# files once
• Build all projects and run all tests sequentially
• Status: complete

Current Follow-up: Correct chat bubble layout and actions

Phase 1: Correct diagnosis

• Compare the rendered screenshot with the outer responsive-width implementation
• Confirm the outer control is constrained but PART_Bubble still overflows internally
• Status: complete

Phase 2: Implement

• Apply the responsive width directly to Flowery's internal PART_Bubble through a permitted global descendant style
• Strengthen regression coverage so it checks the internal template border, not only the outer control
• Constrain the internal content width by the bubble's horizontal padding
• Use 8 px padding for user bubbles and render their timestamp outside Flowery's broken end-footer layout
• Give Copy and Delete the same background while coloring only the Delete glyph red
• Status: complete

Phase 3: Verify

• Compile the corrected XAML successfully
• Run all tests sequentially
• Normalize changed .cs, .csproj, and .axaml files once per correction iteration
• Launch and visually validate the real Windows application with CUA
• Status: complete

Current Follow-up: Wrap, copy, and delete individual chat messages

Phase 1: Inspect

• Inspect every chat-message template and its effective width constraints
• Trace existing per-message copy handling and clipboard command wiring
• Trace message persistence/change notifications needed for individual deletion
• Status: complete

Phase 2: Implement

• Make every message wrap within the available chat-panel width
• Add a visible copy action to every individual message
• Add a visible delete action to every individual message
• Add focused regression coverage
• Status: complete

Phase 3: Verify

• Normalize changed .cs, .csproj, and .axaml files once
• Build and run all tests sequentially
• Status: complete

Current Follow-up: Show loading shade while switching strip images

Phase 1: Inspect

• Trace the image-strip selection command through workspace image loading
• Identify that no existing state narrowly spans the thumbnail decode/load operation
• Locate the main-editor image container and current overlay composition
• Status: complete

Phase 2: Implement

• Expose a narrowly scoped image-switch loading state
• Add a theme-aware, input-blocking shade over the main image
• Add focused regression coverage
• Status: complete

Phase 3: Verify

• Normalize changed .cs, .csproj, and .axaml files once
• Build and run all tests sequentially
• Status: complete

Current Follow-up: Fix late-bound AI chat clipboard

Phase 1: Inspect

• Verify toolbar and per-message commands produce non-empty text
• Verify panel DataContext is supplied by a binding
• Identify clipboard wiring is limited to OnLoaded
• Status: complete

Phase 2: Implement and verify

• Wire from the parent view's DataContext lifecycle
• Await and flush clipboard writes through async commands
• Add wiring, complete-message, and whole-chat regression coverage
• Normalize, build, and run all tests
• Status: complete

Current Follow-up: Fix OpenRouter Gemini image request

Phase 1: Inspect

• Trace the exact request path and serialized payload
• Compare it with current official OpenRouter and Google image-generation contracts
• Check current LLM Tornado support for the required fields
• Status: complete

Phase 2: Implement

• Correct the OpenRouter image-generation request path and parameter mapping
• Add focused regression coverage for Gemini/OpenRouter payloads
• Status: complete

Phase 3: Verify

• Run the iteration's single source normalization
• Build and run all tests sequentially
• Status: complete

Current Follow-up: Fix AI error clipboard action

Phase 1: Inspect

• Trace the error bubble click handler and message command
• Find where CopyToClipboardAction is wired to the active TopLevel clipboard
• Status: complete

Phase 2: Implement

• Wire clipboard access from the view layer using the active TopLevel
• Add regression coverage for the actual copy path
• Status: complete

Phase 3: Verify

• Run the iteration's single source normalization
• Build and run all tests sequentially
• Status: complete

Current Follow-up: Exclude Settings from remembered sidebar selection

Phase 1: Inspect

• Trace Flowery sidebar selection and persistence order
• Identify startup restore and in-session highlight behavior
• Status: complete

Phase 2: Implement

• Restore the last navigable item after Settings is invoked
• Prevent persisted settings state from being restored at startup
• Add Test Connection to the connection editor
• Disable Save after a successful save and re-enable it for every edit-field change
• Add regression coverage
• Status: complete

Phase 3: Verify

• Run the iteration's single source normalization
• Build and run all tests sequentially
• Status: complete

Current Follow-up: AI settings navigation and connection selection

Phase 1: Inspect

• Trace connection selection, edit, save, and collection replacement
• Identify why the saved connection remains visible but the editor becomes empty
• Status: complete

Phase 2: Implement

• Add the AI Assistant settings cog and route it through the existing Settings dialog
• Preserve or restore the edited connection selection after saving
• Add regression coverage
• Status: complete

Phase 3: Verify

• Run the final one-time BOM/CRLF normalization for this iteration
• Build the solution and run all tests sequentially
• Status: complete

Current Follow-up: Accessible AI error messages

Follow-up Phase 1: Inspect and design

• Inspect chat bubble templates, message model, and existing clipboard handlers
• Identify the exact layout cause of clipped error text
• Status: complete

Follow-up Phase 2: Implementation

• Make provider errors wrap, scroll, and remain selectable
• Add a visible per-error copy action
• Status: complete

Follow-up Phase 3: Verification

• Add focused regression coverage
• Run build and tests sequentially
• Normalize changed .cs, .csproj, and .axaml files once
• Status: complete

Current Phases

Phase 1: Inventory and design

• Find the authoritative application-version source
• Inventory every target framework and runtime-related MSBuild setting
• Attribute the large runtimes directory to exact packages/assets
• Status: complete

Phase 2: Implementation

• Bump the application version
• Change all projects and documentation from .NET 8 to .NET 10 Windows
• Enforce win-x64 runtime/platform selection
• Status: complete

Phase 3: Verification

• Fix AI-chat enablement for text-to-image without an input image
• Add regression coverage that the AI panel remains enabled without an image
• Restore, build, and test sequentially
• Verify output layout and runtime contents
• Normalize changed source/project files once and verify the diff
• Status: complete

Phase 4: Delivery

• Review scope and results
• Report version, framework, RID, output size, and validation results
• Status: complete

Previous Completed Task: Hugging Face provider and warning cleanup

Phase 1: Requirements and discovery

• Capture user requirements and repository constraints
• Inspect provider/configuration architecture and warning call sites
• Verify current Hugging Face API contracts from official English documentation
• Status: complete

Phase 2: Technical design

• Choose supported Hugging Face tasks and endpoint/authentication mapping
• Identify UI, persistence, provider-routing, and test changes
• Record decisions in findings.md
• Status: complete

Phase 3: Implementation

• Replace obsolete Avalonia and SkiaSharp calls
• Add Hugging Face provider across shared and desktop AI layers
• Add/update tests and documentation
• Status: complete

Phase 4: Verification

• Run restore/build/tests sequentially
• Fix failures without workaround builds
• Run one final BOM/CRLF normalization for changed .cs/.csproj/.axaml files
• Verify diff and provider/model exclusions
• Status: complete

Phase 5: Delivery

• Review changed scope and validation results
• Report supported HF behavior and remaining limitations
• Status: complete

Key Questions

1. Which Hugging Face API surface supports both chat and image generation with a stable OpenAI-compatible contract?
2. How should models be discovered and classified for the existing combined model dropdown?
3. What exact replacement signatures remove the Avalonia and SkiaSharp warnings without behavior changes?

Decisions Made

Decision	Rationale
Use official English vendor documentation only	User explicitly prohibited Chinese websites
Preserve all existing uncommitted task changes	They belong to the current requested update set
Use https://router.huggingface.co/v1 for HF chat	Official OpenAI-compatible endpoint minimizes duplicated chat code
Use native HF text-to-image HTTP contract	Official docs state the OpenAI-compatible endpoint is chat-only

Errors Encountered

| Combined localization patch used unstable translated-line context | 1 | Patch was atomic and wrote nothing; reapplied using only the stable localization key line | | RTK-filtered rg --files output added line numbers to JSON paths | 1 | Replaced the pipeline with direct find -print0 enumeration | | Provider-section regression looked for AutomationId instead of XAML's dotted attached-property name | 1 | Corrected the test to match AutomationProperties.AutomationId | | Post-normalization test patch introduced mixed CRLF/LF in the test file | 1 | Restored that file to UTF-8 BOM with CRLF and verified all three iteration source files |

Error	Attempt	Resolution
First AI-panel regression test constructed Avalonia UI on a non-owner MSTest thread	1	Parse the XAML source and assert the named AI panel has no IsEnabled binding, avoiding dispatcher and persisted-session side effects
Test search for AppSettings.Current returned no matches with unhandled rg exit 1	1	Treat the valid no-match result explicitly and avoid relying on AppSettings-based test setup
WMIC status query with /value produced an invalid GET expression under Git Bash	1	Drop the formatting switch and use the previously verified plain WMIC column query
Post-normalization build could not overwrite DLLs locked by running app PID 32488	1	Inspect process attribution first; do not terminate because the assistant did not start it
tasklist.exe filter switch was path-converted by Git Bash/MSYS	1	Do not retry that command shape; use WMIC read-only process and owner queries instead
First test run aborted all 172 tests because explicit Win32/Skia startup had no text shaping system	1	Add UseHarfBuzz() to both explicit Avalonia bootstrap chains before rerunning build/tests
First .NET 10 build reported SYSLIB0060 for the PBKDF2 constructor	1	Replace the constructor/stateful reads with one static 48-byte PBKDF2 derivation and split it into the same 32-byte key and 16-byte IV
.NET 10 restore reported NU1510 for explicit framework package references	1	Remove only the four package references identified as redundant: System.Text.Json in three projects and System.Net.Http in the AI project
Combined HF patch targeted SupportedProviders in the wrong file	1	No changes were written; split the patch by verified file location
Standalone AI build failed with three missing StringComparison references	1	Qualified the references with System.StringComparison; solution builds had not compiled this omitted project
First-line import patch did not match the BOM-prefixed file	1	Avoided first-line context and patched the verified method body instead
Final cog-button build was blocked by a user-running app instance (PID 45580)	1	Preserved the process and deferred build/test until the app is closed
Scoped rg call placed --glob options after paths, so Windows treated them as file paths	1	Do not repeat that shape; use options before the assigned target path in any follow-up search
Flowery interface search repeated the same invalid post-path --glob placement	2	Stop using that command shape; inspect the already identified concrete interface file directly
Local Flowery source and test-project paths were inferred incorrectly	1	Resolve exact paths from project.assets.json and rg --files before access
First follow-up build used a non-public Flowery SelectItem method	1	Use the public SelectedItem property and StateStorageProvider.SaveLines, preserving collapsed categories
Clipboard wiring patch included BOM-prefixed first-line context	1	Patch imports from the verified second line, then patch method bodies separately
Combined web-open call failed to parse before execution	1	Split official OpenRouter API and documentation opens into separate calls
Endpoint-capability projection assumed a .data[] array	1	Inspect the exact endpoint response schema before projecting provider capabilities
Assumed a standalone OpenSourceToolkit.AI.Tests directory existed	1	Use rg --files to resolve actual test projects before further access
OpenRouter test patch included BOM-prefixed first-line context	1	Patch imports from the verified second line and add test methods separately
Clipboard ownership patch included BOM-prefixed first-line context	1	Remove imports from verified non-first-line context and patch parent/panel separately
Repository-wide helper search expanded rg --files into a second rg argument list	1	Stop that command shape; inspect the already identified concrete scripts path directly
Assumed the default localization file was under OpenSourceToolkit.NET/Localization	1	Do not retry the inferred path; resolve the exact .resx location with rg --files first
Assumed the previously documented flowery junction still existed	1	Stop using that path; derive the installed Flowery.NET package root from the current project.assets.json package folders
xmllint is not installed in native Git Bash	1	Do not retry that tool; rely on the authorized Avalonia XAML compilation during the normal solution build

Notes

• Repository shell policy requires native Git Bash and RTK.
• User explicitly authorized test execution during verification.
• BOM normalization applies only to .cs, .csproj, and .axaml and runs once at the end of the iteration.