* Remove the post-templates feature
Removes the browsable post-templates catalog end to end: routes
(app.post-templates.index/apply), controller, form requests, resource,
Registry/PostTemplateData/TemplateNotFoundException services, the
templates:report console command, the templates/ catalog (36 files
across en/es/pt-BR), the templates Index.vue page, the template card
on the post Create.vue choice screen, and its i18n keys across all 16
locales.
Also removes TemplateContextResolver, which fed catalog examples into
PostContentGenerator's AI prompt. The "examples" block (with its
heading) is removed from generator.blade.php rather than left empty;
AI generation no longer references the templates catalog.
The AI content-template system under app/Ai/Templates (image/tweet
card templates for the AI post wizard) and TemplateImageGenerator are
unrelated and untouched.
* Update FormRequest naming examples after the post-templates removal
CLAUDE.md and the Cursor rules cited ApplyPostTemplateRequest and
IndexPostTemplateRequest as naming examples. Both classes were deleted
with the feature, so the guidance pointed at files that no longer exist.
* Share the AiTemplate type between the create screen and the AI wizard
Create.vue and AiPostWizard.vue each declared their own AiTemplate
interface, and they had drifted: the wizard's carried
applies_brand_visuals, the create screen's did not. TypeScript treated
them as unrelated types with the same name, which surfaced as a TS2719
error where the templates prop is passed between them.
The backend sends seven fields (PostController::create), so the shared
declaration follows the payload rather than the union of the two copies.
* Fix OpenRouter support by using laravel/ai default provider config.
PR #216 patched Lab::OpenRouter into every agent match, but laravel/ai
already resolves config('ai.default') — including openrouter — when agents
omit provider(). Those matches also forced unknown providers to Gemini and
BrandAnalyzerRunner checked a non-existent services.openrouter key.
Remove the duplicated provider() overrides, gate availability on
ai.providers.*.key, and document OPENROUTER_API_KEY.
Co-authored-by: Paulo Castellano <hello@paulocastellano.com>
* chore(deps): bump laravel/ai to v0.10.3 for native OpenRouter support
v0.5.1's OpenRouter driver only covered text/embeddings via the legacy
Prism gateway. v0.10.3 ships a native OpenRouter gateway with image,
audio (TTS/STT), and web search support, and drops prism-php/prism as
a dependency. ai-sdk-development skill docs refreshed via boost:update
to match the installed version.
* fix: honor AI_IMAGE_PROVIDER instead of hardcoding OpenAI's gpt-image-2
AiImageClient always passed model: 'gpt-image-2' to Image::of()->generate(),
so AI_IMAGE_PROVIDER silently did nothing for any provider other than
OpenAI — gemini/xai/openrouter would fail against an OpenAI-only model id
and quietly fall back to a stock photo. Usage recording was hardcoded to
provider 'openai' too, so credits were billed against the wrong model
whenever a different provider actually ran.
Drop the hardcoded model so generation falls through to the SDK's own
config('ai.default_for_images') + per-provider default model, and read
the actual provider/model back off the response's meta for usage
recording and source_meta instead of assuming OpenAI.
* refactor: extract AiImageClient into single-purpose steps, fix uncaught exception
generate() built the prompt, called the SDK, and unpacked the response all
in one block, with bytes extraction happening after the try/catch — so a
response with an empty images collection threw an uncaught RuntimeException
from ImageResponse::firstImage() instead of returning null as documented.
Split into cleanKeywords(), buildPrompt(), resolveBrandContext(), and
toResult(), and moved response unpacking inside the try block so any
malformed response is treated as a failure like everything else. Added a
regression test with an empty-images fake response.
* fix: drop hardcoded default_text_model, resolve model per provider
default_text_model was the only per-modality model override in ai.php —
image, audio, transcription, embeddings, and reranking all just pick a
provider and let it use its own default model. Text had a config-pinned
model on top, forced into every agent's model() and into every usage
log's model field regardless of which provider actually ran. Switching
AI_TEXT_PROVIDER (e.g. to openrouter) kept sending OpenAI's model id to
whichever provider ended up handling the request.
Removed model() from all six agents so laravel/ai resolves the model
from the active provider's own default (OpenAI's default is already
'gpt-5.4', so no behavior change there). Usage-recording call sites now
read the actual provider/model back off the response's meta instead of
assuming config('ai.default')/default_text_model. StreamPostContent
needed the then() callback since broadcast()'s StreamableAgentResponse
doesn't expose meta directly.
* refactor: drop AiConfiguration wrapper, use data_get() for array reads
AiConfiguration was a one-line static helper used by only two call
sites, with no laravel/ai equivalent to lean on (confirmed AiManager
and the Provider base class expose no isConfigured()/hasKey() check —
the package's model is try-then-catch, not pre-flight checks). Inlined
the filled(config(...)) check directly into HandleInertiaRequests and
BrandAnalyzerRunner instead of keeping a class around one line of logic.
Also swapped direct array-key reads for data_get() per project
convention across every file touched by the recent AI provider/model
fixes (agents' budget arrays, RunGenerateNode/StreamPostCreation's
humanizer merge, RegeneratePostMediaImage's baseContext/copy/rendered
access). Write/assignment sites (`$x['key'] = ...`) are left as-is —
data_get() only reads.
* feat: enhance AI configuration with new providers and options
Added strict types declaration and updated Azure OpenAI API version. Introduced new 'bedrock' provider configuration with AWS credentials and role assumptions. Enhanced existing providers with additional options, including image deployment for Azure and OpenAI, and updated URLs for Gemini and Ollama. Added support for an 'openai-compatible' driver to broaden integration capabilities.
* fix: remove trailing newline in AI configuration file
* feat: add support for OpenRouter and ElevenLabs API keys in configuration
Updated the production Docker Compose file and example environment file to include commented-out entries for OPENROUTER_API_KEY and ELEVENLABS_API_KEY. This enhances the configuration options for AI providers, allowing for easier integration of additional services.
* feat: allow per-provider model overrides for every AI modality
Adds a `models` array to each provider block, wired to env vars, so
self-hosted operators can pin a specific text/image/audio/transcription/
embeddings/reranking model instead of relying on the package's built-in
default for that provider. Only added for the modalities each provider
actually implements (verified against laravel/ai's Provider classes).
Azure is left untouched — it resolves models via deployment names
(AZURE_OPENAI_DEPLOYMENT etc.), not raw model strings, which was already
wired before this change.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Resolves the AiPostWizard conflict and completes the feature:
- i18n parity: brand_colors_label + brand_colors_description in all 15 locales
(was en/es/pt-BR only, which broke LocalizationParityTest).
- Reworked the two-button toggle into a Switch with an explanatory description
(matches the settings Switch/card pattern).
- Only shown for templates that honor the flag: added appliesBrandVisuals() to
the AiContentTemplate contract (ImageCard=true, tweet cards=false), exposed as
applies_brand_visuals in the create-page DTO, and gated the toggle on it — so
it no longer appears (as a no-op) for tweet-card styles.
- Tests: TemplateContractTest covers appliesBrandVisuals for all templates.
- Brand-analyzer prompt now lists every supported language instead of only
en/pt-BR/es, so onboarding autofill can detect the 12 added languages. The
backtick-formatted list is built in BrandAnalyzer::instructions(), keeping the
Blade clean and the enum free of prompt presentation.
- Translate the delete-confirmation keyword for el/ja/zh/ar (the four locales
that still shipped the English "delete").
- Make the language and font comboboxes RTL-correct (logical ms-* instead of
physical ml-*), and let the language combobox be searched by English name via
a visually-hidden label (ContentLanguage::options() now exposes englishName).
- Correct the ContentLanguage class docblock: the enum is also the source of
truth for the UI locales' text direction.
Tests: SetLocale middleware dir/RTL, isRtl and the full 15-language
englishName/label match arms, LLM language detection beyond en/es/pt-BR, and
store-path persistence of a non-default content language plus rejection of an
unsupported one.
The new content-language options were hand-duplicated across request
validation, the UI picker, and homepage detection, while the brand
analyzer's structured-output enum and the AI image prompt's language
name still only knew about en/pt-BR/es. That left autofill unable to
detect the new languages and made image text fall back to English for
them.
Introduce App\Enums\Workspace\ContentLanguage as the single source of
truth and derive every site from it:
- Store/UpdateWorkspaceRequest validate against ContentLanguage::values()
- BrandAnalyzer's language enum uses ContentLanguage::values()
- AiImageClient::languageName() resolves via the enum's englishName()
- HomepageMetaExtractor detects through ContentLanguage::fromHtmlLang()
- BrandForm consumes availableContentLanguages from the backend, like
availableFonts/availableImageStyles, instead of a hardcoded list
Also fix two labels: nl "Nederlandse" -> "Nederlands", zh -> "中文".
Replace the 'single'/'carousel' string format threaded through the AI
generation path with a typed GeneratorFormat enum (PostContentGenerator,
PostContentHumanizer, PostContentStreamer, StreamPostCreation,
RunGenerateNode, deriveFormat).
Remove generatorFormat() from AiContentTemplate and the three templates —
it conflated style and format (returning 'tweet_card'). The tweet-card
check now lives on ContentStyle via isTweetCard(), and the humanizer skip
on ContentStyle::humanizes(), so the style declares its own behaviour
instead of callers comparing magic strings.
- TemplateContext gains bool $isCarousel (set from format === CAROUSEL_FORMAT in StreamPostCreation)
- AiContentTemplate interface: promptView() now takes TemplateContext
- ImageCardTemplate handles both single and carousel via context.isCarousel; schema() returns the correct shape per mode; assemble() routes to assembleSingle/assembleCarousel
- TweetCardTemplate handles both single and carousel; promptView() returns tweet_card_carousel view when isCarousel; schema() returns {tweet_text} or {caption, slides[]}; assemble() calls forTweetCard (single) or forTweetCardCarousel (carousel)
- CarouselTemplate deleted — carousel is now a format dimension each style handles, not a standalone style
- AiTemplateRegistry: [ImageCardTemplate, TweetCardTemplate] only
- PostImagePipeline: adds forTweetCardCarousel — renders one renderTweetCard per slide text, returns N media items
- StreamPostCreation: carousel detection restored to $format === CAROUSEL_FORMAT (like main); style dispatched via AiTemplateRegistry::find($this->template); humanize skips tweet_card (single and carousel)
- PostContentGenerator: promptView() call updated to pass templateContext
- New prompt: resources/views/prompts/post_content/tweet_card_carousel.blade.php
- Tests: TemplateContractTest and TweetCardTemplateTest updated for promptView(context) signature; AiTemplateRegistryTest asserts [image_card, tweet_card]; StreamPostCreationTest adds tweet_card carousel test; TweetCardTemplateTest covers both schema shapes and both promptView return values
- Add AiTemplateRegistry (resolves, keys, default)
- Add template/templateContext params to PostContentGenerator; schema() delegates to template when both are set; instructions() uses template.promptView()
- Refactor StreamPostCreation.handle() to resolve template via registry, build TemplateContext, delegate assemble() to template; add createPostFromGenerated()
- Remove handleSingle(), handleCarousel(), resolvedContentType() (no other callers)
- PostAiCreateController passes template param from request (default image_card)
- All existing AI tests green; registry test added
Replace free-text brand_tone/brand_voice_notes with a single structured
brand_voice_traits JSON column backed by the BrandVoiceTrait enum, exposed
as choice-chip pills in the brand settings UI and autofillable from a site.
Brand voice and visuals become per-automation toggles on the Generate node.
Unify the image controls into one 0-10 picker (0 = text-only, 1 = single,
2+ = carousel) and feed the generator the most restrictive selected network
so copy fits every platform. Pass that same platform context through the
humanizer pass — extracted into a shared ResolvesPlatformCopyBudget trait —
so the rewrite can no longer drift past the character cap the generator
respected, in both the automation and manual creation flows.
Persist the trigger node's schedule editor fields on save (they were
silently dropped by validated() for lacking validation rules).
- Replaced the `regenerate_image` boolean with a `change_mode` string parameter in the PostImageRegenerator schema, allowing for more granular control over regeneration options.
- Updated methods in RegeneratePostMediaImage to handle the new `change_mode` values: `image_only`, `text_only`, and `both`.
- Enhanced documentation and comments to clarify the new regeneration logic.
- Added tests to verify correct behavior for different regeneration modes, ensuring that text and image changes are handled appropriately.
- Introduced a new `regenerate_image` boolean parameter in the PostImageRegenerator schema to control background image regeneration.
- Updated relevant methods in RegeneratePostMediaImage and TemplateImageGenerator to utilize the new parameter for reusing existing backgrounds when appropriate.
- Enhanced documentation and comments to clarify the usage of the `regenerate_image` option.
- Added tests to ensure correct behavior when reusing background images.
- Added guidelines for AI agents to avoid embedding prompts in PHP and to store them in Blade files under `resources/views/prompts/`.
- Updated `PostImageRegenerator` to render prompts using Blade instead of heredocs, improving maintainability and consistency with project standards.
- Enhanced documentation in `CLAUDE.md` and `.cursor/rules` files to reflect these changes.
Let users adjust AI-generated slides in place via async job and Echo, while applying workspace brand, background, and text colors to image prompts. Autofill swaps site text/background colors for the image palette, and regeneration is blocked on finalized posts with safer job cleanup.
Co-authored-by: Cursor <cursoragent@cursor.com>
Two more pieces in app/Ai/ that were registered but never reached by any
production caller, same audit as the PlatformRules cleanup.
DebugGeminiRequest middleware: zero references — not registered in the
laravel/ai middleware pipeline, no agent imports it. Pure dev-time
scaffolding that got left behind.
ExtendedGeminiProvider: registered via Ai::extend('gemini', ...) so the
class IS resolvable when AI_DEFAULT=gemini, but its only override —
defaultImageOptions — is unreachable in practice:
- All text agents (PostContentGenerator, PostContentHumanizer,
PostContentReviewer, PostContentStreamer, BrandAnalyzer) only call
text generation, never defaultImageOptions.
- Image generation (AiImageClient) hardcodes OpenAI's `gpt-image-2`
model, so the Gemini provider isn't on the image path either.
Removed:
- app/Ai/Middleware/DebugGeminiRequest.php
- app/Ai/Providers/ExtendedGeminiProvider.php
- tests/Unit/Ai/Providers/ExtendedGeminiProviderTest.php
- Ai::extend('gemini', ...) wiring + configureAi() method and its four
imports (ExtendedGeminiProvider, Ai, GeminiGateway, Dispatcher) in
AppServiceProvider.
The whole app/Ai/PlatformRules/ system was registered in AppServiceProvider
but never read in production code. Confirmed via grep across app/, resources/,
config/, database/, tests/ — only consumers were:
- AppServiceProvider::configurePlatformRules() registering everything in the
Registry (which nothing then called)
- tests/Feature/Ai/PlatformRules/RegistryTest.php asserting the registrations
it had just performed
The actual AI text agent (PostContentGenerator) reads the per-platform caps
straight from the Platform enum (maxContentLength / recommendedAiContentLength)
and feeds them into the prompt template — bypassing this layer entirely.
Removed:
- app/Ai/PlatformRules/ (12 files: Contract, Registry, 10 Rule classes)
- tests/Feature/Ai/PlatformRules/RegistryTest.php
- 11 use imports + boot() call + configurePlatformRules() method in
AppServiceProvider
- Platform::contentOverflow returns int (0 = fits) via max(), drops the
nullable-int + ternary pattern; HasSocialHttpClient and the validation
rule updated to compare against 0 instead of null.
- Drop the hardcoded 63206-char limit to 10000 across all entry points
(Platform enum, MCP CreatePostTool/UpdatePostTool, FacebookRules, all
three FormRequests). Facebook's API accepts up to 63206 but nobody
writes 63k-char posts and emoji-heavy content risks overflowing the
TEXT column's 65535-byte ceiling.
- Compiled i18n JSON regenerated with the content_exceeds_platform key.
Replaces the vague "intro → development → conclusion" guidance in the
generator prompt with an explicit per-slide arc: open with a specific
real urgent problem, develop the idea while creating swipe motivation,
prove with concrete evidence (result/before-after/BTS/learning), and
close with a single specific next action. Includes a slide_count to
role-distribution table (2 through 6+ slides).
Adds a `role` enum field to each slide in the carousel JSON schema
(hook, development, proof, cta) so the structure is enforced via SDK
not just prompt text. Field is additive — handleCarousel still reads
only title/body/image_keywords, so consumers are unaffected.
- TemplateImageGenerator->render() now returns a typed array shape
{path: string, source_meta: array} instead of a one-off RenderedSlide
DTO. Single internal callsite, no need for a dedicated class.
- Drop the `?? 'en'` fallback on $workspace->content_language in 6
callsites: the column has a NOT NULL default of 'en' at the DB level,
so the null coalesce was dead code.
- WorkspaceFactory now seeds content_language, brand_tone, brand_font
and image_style explicitly so make() (no DB persist) produces a
complete model — DB defaults aren't applied until create().
Mentions in post comments
- @mention autocomplete (workspace members, current user excluded) with
marker syntax @[uuid] persisted, display names rendered via CommentBody
chips; live edit replaces markers with names and converts back on save.
- NotifyMentions action with workspace-scoped membership check, dedupes
same user, only newly-added mentions on update.
- Email + in-app via SendNotification job, respecting per-user
notification_preferences.mentioned_in_comment.
- Heartbeat-based presence (Cache, 60s TTL, 30s ping) so online recipients
get only the in-app notification — no email noise.
- Real-time bell on workspace.{id}.user.{id} private channel
(NotificationCreated event), scoped channel name avoids client-side
filtering and lays out a convention for future workspace channels.
- Mailable localized via lang/{en,es,pt-BR}/mail.php; Maizzle source
template for the email is committed and built into resources/views/mail.
AI generation refactor (Action layer + MCP)
- Extracted Actions/Ai/Generate{Image,Video} with QuotaExhaustedException
so agent tools and MCP tools share a single domain entry point.
- Mcp/Tools/Ai/Generate{Image,Video}Tool registered in TryPostServer; both
return MediaResource payloads.
- Orientation::imageApiSize maps non-OpenAI ratios to 1:1/2:3/3:2.
- config/ai.php is now the single source of truth driven by env, removing
the trypost.ai shim. Default text/image providers flipped to OpenAI.
Settings/UX
- /settings/workspace split into shadcn Tabs (Workspace / Brand / Users)
with three components.
- /assets and the in-editor MediaPicker open the ImagePreviewDialog
lightbox on image click while preserving action button behaviour.
- Comments tab landed via ?tab=comments&comment=<id> from notification
click (scroll-to + temporary highlight).
- Mention autocomplete popover flips above when near the viewport bottom.
- Real social platform PNGs replace Tabler brand glyphs in schedule
pills and post list, with hover tooltip carrying display_name + handle.
Bug fixes
- AcceptInvite: controller now passes workspace + role payload that the
Vue page expects; login/register CTAs preselect the invite email.
- WorkspaceInvite mailable: stopped referencing nonexistent
$invite->workspace and $invite->role; column added to the migration,
Invite model casts role to WorkspaceRole, CreateInvite persists it.
- PostCommentCreated: added broadcastAs so .PostCommentCreated actually
matches the Echo listener; payload now includes mentioned_users so
receivers render the chip correctly without a refetch.
- Preview components for X/Pinterest/Threads/Bluesky/LinkedIn/Mastodon/
TikTok/YouTube switched from item.type === 'image' to
!isVideoMedia(item) so media without a persisted type still renders.
- UpdatePostRequest now accepts media.*.{type,mime_type,size,...} so the
posts.media JSON keeps the metadata that the previews need.
- Removed throttle:6,1 from social connect routes (was 429ing legitimate
OAuth retries).
- Used MediaType enum cases instead of literal 'image'/'video' strings
when creating media rows.
Tests
- MentionParser unit tests, NotifyMentions feature tests including
online/offline channel selection and preference gating, MCP AI tool
happy paths, MentionedInComment mailable rendering, AcceptInvite +
search-members + index mentioned_users path. 1229 passing.
When config('trypost.ai.text_provider') points to a provider whose
API key is populated (services.gemini.api_key or
services.openai.api_key), the autofill action pipes the homepage
markdown through a BrandAnalyzer agent with structured output to
produce higher-quality values for:
- brand_description — 2-3 sentences summarizing the company based on
the actual page content, not the raw meta description (which is
often generic SEO boilerplate)
- brand_tone — classified into one of our seven enum values from the
writing style on the page
- content_language — detected from the actual content (more reliable
than <html lang> which is often wrong)
- brand_voice_notes — concrete writing guidelines inferred from the
site's style, written in the detected language
When the LLM provider is NOT configured (open-source self-hosted
deploys without API keys), the action keeps the existing
deterministic meta-tag-only flow — no crash, no noise, no LLM cost.
When the LLM fails mid-request, we log a warning and fall back to
the meta-tag values so the user still gets something useful.
Stack additions:
- league/html-to-markdown ^5.1 converts the main body to clean
markdown for the LLM input (truncated to 4000 chars).
- BrandAnalyzer agent (Agent + HasStructuredOutput) with schema
enums matching our allowed tones and languages.
- resources/views/prompts/brand_analyzer.blade.php holds the
instructions, including explicit enum lists and examples of good
voice_notes.
Frontend: Brand.vue now also fills brand_tone and brand_voice_notes
from the response when present.
Tests (+3): LLM-configured happy path, no-credentials fallback
(asserts BrandAnalyzer is never prompted via preventStrayPrompts),
and LLM-exception fallback (meta tags win, brand_tone stays null).
Users can now pick the language AI uses for captions, hashtags,
descriptions, AND on-image / on-screen text — independent of the
app UI locale. A user with pt-BR UI who writes for an English
audience gets English content, and vice versa.
- New content_language column on workspaces (added to the existing
create_workspaces_table migration since we're pre-production),
defaulting to 'en'. Fillable on the model.
- Dropdown in Settings → Workspace → Brand, side-by-side with Tone.
Options: English, Português (Brasil), Español. Short helper copy
explains what the setting controls.
- Validation: sometimes|string|in:en,pt-BR,es — optional on update
so existing callers that don't post the field keep the current
value (DB default handles fresh workspaces).
- System prompt now uses $content_language (sourced from the
workspace) instead of app()->getLocale(). The rule was also
strengthened: the AI must always write in this language regardless
of what language the user types instructions in, and must instruct
image/video tools to render on-media text in this language too.
- Image and video Blade prompts received a new line forcing any text
rendered inside the generated media to use $content_language.
- GenerateImage tool and VideoGenerationService pass the workspace's
content_language into their respective prompt templates.
Previously the caption would be in Portuguese (since it followed the
user's input language) but the image could render with English text
because there was no explicit constraint. Now the whole pipeline is
aligned to one per-workspace language.
Critical fixes:
- Quota enforcement regressed when the regex flow was replaced with
tools. The LLM was trusted to respect [Session state] quota hints,
which is vulnerable to prompt injection / hallucination. Each tool
now checks Pennant feature limits at the top of handle() and returns
a short quota-exhausted string instead of calling the provider.
Audio shares the video quota (there is no AiAudioLimit by design).
- N+1 on aiMessages: combined the two separate ->get() scans that
produced imagesInThread and videosInThread into a single query.
- postPlatforms lazy load: controller now loadMissing('postPlatforms')
before constructing SocialMediaAssistant, which accesses the relation
in activePlatformRules().
Project conventions:
- \RuntimeException and \Throwable are now imported at the top of the
controller instead of inlined per CLAUDE.md.
- Tool handle() methods use data_get($request, 'prompt') instead of
direct array access, matching the data_get convention for service
classes.
- enrichContent() pluralizes attachment types via Str::plural so
'2 images' reads naturally instead of '2 image'.
Tests:
- Added Storage::fake('public') in tool tests so runs don't pollute
the local disk with fake-generated files.
- Restored quota enforcement coverage (deleted when the flow changed)
as three new tests — one per tool — that fill the monthly bucket
and assert the tool refuses and nothing is generated.
- Removed inline \App\ references in tests in favor of imports.
SocialMediaAssistant::activePlatformRules() reads post_platforms on the
attached Post, maps them to Platform enum values, and resolves rule
classes via Registry::forMany(). The rendered instructions include only
the summaries of platforms actually selected on this post — so when a
user is posting to X only, the agent doesn't waste tokens thinking
about Instagram Reels.
The PLATFORM KNOWLEDGE block in system.blade.php is replaced with
@include('prompts.assistant.platforms') which iterates the passed rules
and renders each summary. When the post has no platforms, the section
renders nothing (a @if guards it).
Each platform has a rule class exposing specs() (char limits, aspect
ratios, media limits, format-specific constraints) and summary() (a
short concise description the agent can render into instructions).
Registered in AppServiceProvider::configurePlatformRules() mapping
Platform enum values to rule classes. Both InstagramFacebook and
LinkedInPage share rules with their non-business siblings.
Contract + Registry follow the lookup-map pattern. Registry is a
static registry seeded at boot — rules are cheap enough to new up
per request.
Tests cover all 12 platforms via Pest dataset and verify the forMany
fan-out, clear/register behavior, and representative summaries.
The agent now implements HasTools and returns three tool instances
(GenerateImage, GenerateVideo, GenerateAudio) each wired with the
workspace + post + userId context.
Replaced the MEDIA GENERATION RULES block in the system prompt. Tools
replace the old [GENERATE_IMAGE:vertical] regex convention — the LLM
now invokes tools directly with typed parameters (prompt, orientation),
which eliminates the fragile text parsing in the controller.
Each tool implements Laravel\Ai\Contracts\Tool with description(),
handle(Request), and schema(JsonSchema). They share a common pattern:
- Constructor receives Workspace, optional Post, userId, and an optional
AttachmentCollector (resolves from container if not injected).
- handle() does the work (Image::of(), Audio::of(), or calls our custom
VideoGenerationService for Veo), persists media + usage log, pushes
the full attachment shape into the collector, and returns a short
text summary to the LLM.
- schema() exposes typed parameters with enum constraints and doc
strings so the LLM selects valid inputs.
Also fix the Ai::extend closure signature — MultipleInstanceManager
passes (app, config) not just (config) to custom creators.
The agent implements Laravel AI SDK's Agent + Conversational contracts:
- instructions() renders the existing Blade system prompt with workspace
brand context (name, description, website, tone, voice notes, locale)
- messages() reads directly from the AiMessage model scoped to the post,
so our existing conversation persistence stays the single source of
truth — no duplicate SDK-managed storage.
- Assistant messages with attachments are enriched inline with a summary
like [This assistant message attached: 2 image] so the model tracks
progress through carousel-style multi-image generations.
- provider() maps the trypost.ai.text_provider config to Lab::Gemini or
Lab::OpenAI, preserving the existing provider-switching behavior.
The stock provider only exposes 1:1, 2:3, 3:2 via defaultImageOptions()
match statement — any other ratio gets silently dropped to null.
ExtendedGeminiProvider adds 9:16, 16:9, 4:3, 3:4, 4:5, 5:4, 21:9 so
Image::of()->size('9:16') now works for Reels/Stories/TikTok/Shorts
and size('16:9') for X/LinkedIn/YouTube landscape content.
Gemini's native API supports all these ratios; the SDK restriction was
purely in the provider's match statement.