Changelog

All notable changes to spaanproductions/laravel-ai-usage are documented in this file.

v0.1.8 - 2026-10-19#

• content_limit can be set with AI_USAGE_CONTENT_LIMIT.
• Traces have their own limit per string, trace.content_limit (AI_USAGE_TRACE_CONTENT_LIMIT). It defaults to content_limit, so nothing changes unless you set it.

v0.1.7 - 2026-10-19#

• Separate retention for parts of a row, so the rows (tokens and cost) can be kept longer. model:prune now also clears the prompt and response with content_retention_days, the execution trace (and its attachments) with trace_retention_days, and only the attachment files with attachment_retention_days. All are off by default.
• The record page no longer links to attachments the retention has deleted.

v0.1.6 - 2026-10-17#

• Publishing the migrations again after an update only adds the new migrations. Before, vendor:publish --tag=ai-usage-migrations copied every migration again under a new date, so migrate failed on tables that already exist. If you followed the v0.1.4 upgrade note, delete the duplicated create_ai_usage_records_table and create_ai_usage_budget_alerts_table migrations that are not yet migrated.

v0.1.5 - 2026-10-17#

• Applications that use immutable dates (Date::use(CarbonImmutable::class), as the Laravel starter kits do) no longer get a TypeError on the dashboard and record list, and their budget alerts fire again. Dates are typed as CarbonInterface, and the dashboard chart no longer relies on mutating its date cursor.

v0.1.4 - 2026-10-15#

Publish the ai-usage-migrations again and migrate: a new migration adds the trace columns. Update to v0.1.6 first, or delete the duplicated older migrations this publish adds (see v0.1.6).

• Execution traces. Every row links to its trace and parent call (trace_id, parent_invocation_id, parent_tool_invocation_id), so sub-agents and calls made inside tools form a tree on the record page.
• trace.enabled (AI_USAGE_TRACE, off by default) stores the full trace of agent calls on laravel/ai 1.x: instructions, input, output, every model step (tokens, cost, text) and every tool call (arguments, result), with timings. The record page shows it as a waterfall with expandable spans.
• trace.attachments (AI_USAGE_TRACE_ATTACHMENTS, off by default) copies prompt attachments to a disk, downloadable from the record page and pruned with the rows.
• redaction (on by default) masks values under sensitive keys (*_token, password, api_key, …) and optional regex patterns in stored prompts, responses and traces.
• Failed calls get an error_kind (rate limited, overloaded, connection, credits, tool error, exception), with a filter on the record list and in the CSV export.
• AiUsage::withoutRecording() skips recording for a block of code.
• Model ids with a version suffix (e.g. claude-sonnet-5-5-v2) are priced with their base model.

v0.1.3 - 2026-10-12#

• Configurable table names (ai-usage.tables), used by the migrations and the models, for applications that already have a table with one of the default names.

v0.1.2 - 2026-10-10#

• A "Models" page next to Dashboard and Records: every synced model with its summary, modality icons, context and prices per million tokens, filterable by provider, pricing and search, sorted newest first with a sort dropdown, and a "Sync models" button that runs the same sync as ai-usage:sync-models.
• The stylesheet URL carries a version hash, so browsers load the new stylesheet after an update instead of keeping the cached one for a year.

v0.1.1 - 2026-10-10#

• "Today" and "Yesterday" periods on the dashboard and the record list.
• The dashboard chart shows hours for periods of up to two days.
• The dashboard chart can show cost, tokens or requests.

v0.1.0 - 2026-10-10#

First release.

Added#

• Records the tokens and cost of every laravel/ai call from its events: agent prompts and streams, embeddings, images, audio, transcriptions, reranking, failovers and failed agent calls.
• Support for laravel/ai 0.9 – 0.11 and 1.x. Token counts are normalized, so cached tokens are never charged twice on 1.x.
• Latency per call. Start times of calls that never complete are dropped after an hour and capped at 1,000, so long-running queue workers do not leak memory.
• Attribution to a feature, subject and user via AiUsage::context() and the ProvidesAiUsageContext agent interface. The HasAiUsage trait for subject models.
• Pricing from the ai_models table, synced from OpenRouter by ai-usage:sync-models, with the rates used stored on each record. ai-usage:reprice recalculates historical rows.
• Budgets per day, week or month, optionally per feature, with the BudgetThresholdReached event (once per period and threshold) and an optional mail.
• Retention through model:prune and retention_days.
• Optional storage of the prompt and response text (off by default).
• Dashboard, record list (filters, sorting, CSV export) and record detail as full-page Livewire components, with a standalone layout and compiled stylesheet. Routes, middleware, gate and layout are configurable.
• Publish-only migrations (ai-usage-migrations, ai-usage-models-migrations), a demo seeder and a Laravel Boost skill.