Trail Laravel Package

Product Decisions This Supports

• Multimodal roadmap decisions: v0.4.0 adds first-class tracing for image generation, text-to-speech, and transcription, so PMs can ship multimodal AI features with observability from day one rather than retrofitting instrumentation later.
• Cost attribution and pricing strategy: Image, audio, and transcription calls now count toward run token totals and can be priced from model token rates. This supports per-feature, per-customer, or per-modality cost analysis — but only after you add prices for those models, since no defaults ship and calls remain Unpriced until configured.
• Reliability and support triage: Failures from the Laravel AI SDK’s ImageFailed, AudioFailed, and TranscriptionFailed events are recorded with their own error and duration on SDK 1.2.2+. This helps PMs prioritize fixes by modality and understand whether failures are isolated calls or tool-level failures.
• Usage reporting and export: The runs list, CSV export, and Boost trace listing can be filtered by the new root operations, enabling product reporting on image, audio, and transcription adoption without custom queries.
• Build vs. buy for AI observability: Instead of building custom tracing, cost tracking, and failure capture for each new AI modality, the package extends existing trace/span infrastructure. This is a strong argument for buying/ adopting rather than maintaining bespoke instrumentation.
• Privacy and payload decisions: Texts are stored only when payloads are captured; image and audio bytes are never stored. This supports product decisions around debugging depth versus data minimization.

When to Consider This Package

Consider adoption when:

• Your Laravel application uses the Laravel AI SDK and you are adding or already run image generation, TTS, or transcription features.
• You need trace-level visibility into multimodal AI calls: settings, counts, failures, durations, and token-based cost totals.
• You want to filter and export AI usage by operation for product analytics, customer reporting, or incident review.
• You are on SDK 1.2.2+ and want failed image/audio/transcription calls recorded as Failed rather than left Incomplete or classified by stale rules.
• You can deploy the new version to every process that reads Trail’s tables before recording these new span types — earlier versions cannot read the new span types.
• You are willing to configure model prices for image/audio/transcription models; otherwise those calls will appear Unpriced.

Look elsewhere or defer when:

• You are not on Laravel or not using the Laravel AI SDK.
• You need storage of image or audio bytes for replay/debugging; this package explicitly does not store them.
• You require deep custom APM, distributed tracing, or vendor-specific observability beyond the package’s trace/span model.
• You cannot coordinate a full deployment across all table-reading processes, or you rely on older versions that cannot read the new span types.
• You need out-of-the-box pricing for multimodal models; this release requires manual price configuration.

How to Pitch It (Stakeholders)

For executives: This release turns multimodal AI — image generation, voice, and transcription — into measurable product operations. We can see usage, failures, latency, and token-based cost by modality, and export that data for reporting. It reduces the need to build custom AI observability and helps us decide where to invest, price, or fix reliability. The main operational caveat is that every process reading Trail’s tables must be upgraded before these new span types are recorded, and multimodal models need prices configured to avoid Unpriced totals.

For engineering: v0.4.0 extends the existing trace/span model to image, audio, and transcription calls with no new migrations or configuration. Standalone calls become their own traces; calls inside tools become spans under that tool. Failures are captured from SDK 1.2.2+ events, payload text is captured only when enabled, and image/audio bytes are never stored. Plan a coordinated deploy because older versions cannot read the new span types, and add model prices if cost totals matter.