v3.6.0

Added

• Atlas::batch() submits large request sets as deferred jobs for ~50% lower cost — OpenAI (text, vision, embeddings), Anthropic and Google/Gemini (text, vision). With persistence on, submit() auto-tracks the job, atlas:batch-poll pulls results in keyed by custom_id, and atlas:batch-prune bounds history (atlas.batch.retention_days, default 90); with it off, batching is stateless and you poll the provider yourself. Tools and per-request middleware aren't supported and are rejected up front.
• ->reasoning(ReasoningEffort) on text and agent requests (and an Agent::reasoning() default) enables extended thinking at one effort level (Minimal/Low/Medium/High), mapped to each provider's format. Optional budgetTokens: and includeSummary:.
• ->countTokens() on text and agent requests returns input-token count before sending — exact on Anthropic/OpenAI/Google, estimated on xAI/Ollama/LM Studio. Returns a TokenCount with an estimated flag.

Fixed

• Extended thinking now works through multi-step tool calls and persisted-conversation reloads.
• Google (Gemini) tool calls no longer 400 on non-object results (number, boolean, array, quoted string) — wrapped automatically. Plain strings and JSON objects are unchanged.
• Google (Gemini) withProviderOptions() no longer corrupts nested generationConfig fields — extra keys (e.g. topP) merge alongside Atlas's values, and an overridden key (e.g. temperature) now replaces the value instead of becoming a list Gemini rejects with a 400.
• Google (Gemini) now honors temperature: 0.0 instead of silently dropping it, so deterministic output works.
• Streaming Google (Gemini) tool calls now surface FinishReason::ToolCalls and keep every call when several arrive in one chunk — previously a streamed, tool-terminated response lost its finish reason and could drop a parallel call.
• Chunked embeddings now work when persistence runs on a separate Postgres connection (atlas.persistence.connection) and the app default isn't Postgres — pgvector is detected on the persistence connection, so chunk writes and similarity search no longer fail.
• OpenAI (and xAI) document inputs now send the correct input_file content part instead of input_image, so attaching a Document (PDF, Word, Excel, CSV, text, Markdown, …) works — previously non-file-ID documents were sent as images and rejected with a 400. URLs pass through as file_url; base64/path/storage/upload inline as file_data.
• xAI streaming errors are now attributed to xai, not openai — a mid-stream error through the shared Responses parser previously carried $provider === 'openai', misattributing xAI failures for anything branching on the provider.
• Providers on the shared chat_completions/responses drivers (Ollama, Groq, LM Studio, custom OpenAI-compatible endpoints) now report their configured provider key in exceptions, events, and middleware — previously failures were attributed to the generic driver type (chat_completions), making two such providers indistinguishable.

Migration

No breaking changes — drop-in upgrade. Stateless batching (submit, then poll the provider yourself) needs no setup. Tracked batching (auto-polled, persisted results) adds new tables — publish and migrate:

php artisan vendor:publish --tag=atlas-migrations
php artisan migrate

PR Summary List

• Add reasoning support and APIs for text/agent requests by @timothymarois in #49
• Add pre-flight token counting API and documentation by @timothymarois in #50
• Add Batch API (jobs, polling, persistence) by @timothymarois in #51
• Handle pgvector, Anthropic thinking, OpenAI batch by @timothymarois in #52
• Map Document inputs to OpenAI input_file by @timothymarois in #53
• Pass provider through SSE parsing; add tests by @timothymarois in #54
• Use configured provider name and fix parsers by @timothymarois in #55
• Extract shared Responses handlers, xAI adapters by @timothymarois in #56

Full Changelog: v3.5.0...v3.6.0