Batch 3 — features to absorb or build

[wishborn]

Larger features from the upstream backlog, needing real review/design:

• Anthropic adaptive thinking prism-php/prism#998 — Anthropic adaptive thinking
• feat(anthropic): add fine-grained tool streaming support prism-php/prism#988 — Anthropic fine-grained tool streaming
• feat(anthropic): handle pause_turn and refusal stop reasons prism-php/prism#1003 — Anthropic pause_turn + refusal stop reasons
• issue Support for Anthropic Beta Features (Skills, Code Execution, and Custom Headers) prism-php/prism#953 — Anthropic beta features (Skills, code execution, custom headers)
• Feature - Add support to batching and files API's for Anthropic and OpenAI prism-php/prism#951 — Batching + Files APIs (Anthropic + OpenAI)
• Feat: 🛠️ Add client-executed tools & human-in-the-loop tool approval flow prism-php/prism#932 / issue Human-in-the-loop / Tool confirmation prism-php/prism#921 — client-executed tools + human-in-the-loop approval
• feat: add provider-agnostic withReasoning() request method prism-php/prism#1018 — provider-agnostic withReasoning()
• feat(fim): add FIM request and response handling for Mistral AI prism-php/prism#1014 — Mistral FIM
• feat(xai): Add XAI image generation handler and tests prism-php/prism#960 — xAI image generation
• feature(providers): Added Vertex support prism-php/prism#978 / issue Google Vertex support prism-php/prism#795 — Vertex provider; Add Requesty as a provider prism-php/prism#1026 — Requesty provider; issue Cloudflare Workers AI prism-php/prism#520 — Cloudflare Workers AI
• issue 🔭 Sneak Peak - Evented Telemetry Support with support for driver based OpenTelemetry for platforms like phoenix prism-php/prism#935 — evented telemetry / OpenTelemetry
• issues Async HTTP support prism-php/prism#646 / Add Background & Queue Support for OpenAI prism-php/prism#590 / Scaling with blocking io prism-php/prism#919 — async HTTP, queue/background support
• issue OpenAI driver forces /responses endpoint — no /chat/completions fallback for vLLM and LiteLLM backends prism-php/prism#900 — OpenAI chat/completions fallback for vLLM/LiteLLM

[wishborn]

Done in v0.104.0 — 9 of the PR-backed items landed (prism-php#998, prism-php#1003, prism-php#988, prism-php#1018, prism-php#951, prism-php#1014, prism-php#960, prism-php#978, prism-php#1026), each with fork-side fixes where the PRs were stale or under-tested (see release notes). prism-php#932 deferred to #4 (needs a dedicated rebase). Issue-only items (upstream prism-php#935 telemetry, prism-php#646/prism-php#590/prism-php#919 async+queue, prism-php#900 chat/completions fallback, prism-php#953 Anthropic beta features, prism-php#520 Cloudflare) remain candidates for Batch 4 as original dev work.

[vinitkadam03]

Thanks for the fork and maintaining it! Would really appreciate if we could ship telemetry as well. (Arize Phoenix or something similar maybe driver based)
issue prism-php#935 — evented telemetry / OpenTelemetry

[wishborn]

@vinitkadam03 done 🎉 — thanks for the push (and for the original design in prism-php#935).

What: Prism core now emits neutral Laravel telemetry events across the whole generation lifecycle (GenerationStarted / GenerationCompleted / GenerationFailed, StepCompleted, ToolInvoked) — covering text, structured, embeddings, images, streaming, and tool calls. Core takes no OpenTelemetry dependency; a separate driver package maps those events to GenAI-convention OTel spans for Arize Phoenix / any OTLP backend.

Why this shape: keeping OTel in a companion package (not core) avoids the per-handler sync tax that stalled the earlier attempts, and the events carry a stable trace id + explicit step/tool ordinals so the span tree is built deterministically — no fragile ambient context across the tool loop.

Try it (on a branch, in review before merge — PR #7):

"minimum-stability": "dev",
"require": {
  "particle-academy/prism": "dev-feat/telemetry-935",
  "particle-academy/prism-opentelemetry": "dev-main"
}

Set PRISM_TELEMETRY_ENABLED=true and point the OTel SDK at your Phoenix endpoint. Companion package → https://github.com/Particle-Academy/prism-opentelemetry

Feedback very welcome before we merge 🙏