Batch 1 follow-ups

[wishborn]

Loose ends from the v0.101.0 merge wave:

• Systemic tool-argument normalization (upstream issue Systemic tool-call normalization bug: multiple providers crash with raw TypeError on missing or scalar arguments prism-php/prism#1007): coerce string values to declared scalar types when parsing tool calls, so Groq's anyOf relaxation (fix(groq): accept string values for boolean/number tool parameters prism-php/prism#986) doesn't surface string args to typed PHP handlers.
• Backfill a test for the Groq function-name splitting heuristic (fix(groq): handle Llama models embedding arguments in function name prism-php/prism#987).
• Re-add an OpenRouter streaming payload test for parameterless tools with a real fixture (the one in fix(ToolMap): remove static duplicate parameters handling in mapping openrouter prism-php/prism#1004 was unrunnable and was dropped).
• Consider expiring the o3-mini/chatgpt-4o prefixes in StructuredModeResolver once model naming stabilizes.

[wishborn]

Added from Batch 2:

• Backfill fixture-based tests for the Mistral structured tool-call loop (upstream Add tool call support to Mistral structured handler prism-php/prism#965 shipped without tests).
• Evaluate upstream fix(prism): avoid requiring full laravel framework prism-php/prism#1025 (replace laravel/framework with illuminate/* components) properly: audit all Illuminate usage, add a composer-require-checker CI job, and decide whether the Dependabot visibility loss is acceptable.
• DeepSeek promptTokens now excludes cache-hit tokens (upstream fix(deepseek): extract prompt_cache_hit_tokens and reasoning_tokens from usage prism-php/prism#1021) — verify consistency of usage semantics across providers (relates to upstream Usage counts for Gemini vs. OpenAI are inconsistent in completionTokens. prism-php/prism#1017).

[wishborn]

v0.105.0 progress:

• ✅ Tool-argument coercion (string→scalar + BackedEnum) shipped — closes the Feature request: auto-convert string values to BackedEnum instances when tool handler parameters are typed as enums prism-php/prism#1016 item; the coercion half of Systemic tool-call normalization bug: multiple providers crash with raw TypeError on missing or scalar arguments prism-php/prism#1007 is done (crash-hardening for malformed args was already handled by the existing validation layer).
• ✅ Groq name-split heuristic covered end-to-end.
• Remaining: Mistral structured tool-loop fixture tests, OpenRouter parameterless-tool streaming fixture, fix(prism): avoid requiring full laravel framework prism-php/prism#1025 evaluation, provider usage-semantics consistency check (Usage counts for Gemini vs. OpenAI are inconsistent in completionTokens. prism-php/prism#1017).

[wishborn]

Progress (merged to main in 5b6af16):

• ✅ Mistral structured tool-loop fixture tests (both paths: tool-called-then-structured-resend, and stop-without-calling discard path)
• ✅ OpenRouter parameterless-tool streaming payload test with real SSE fixtures (replaces the unrunnable one from upstream fix(ToolMap): remove static duplicate parameters handling in mapping openrouter prism-php/prism#1004)

Analyses (no code changes):

• Upstream fix(prism): avoid requiring full laravel framework prism-php/prism#1025 (illuminate/ split): recommend REJECT.* The proposal is provably incomplete — it misses illuminate/concurrency (used by parallel tool execution), lists unused illuminate/bus, and cannot declare the Foundation helpers (config(), response(), resolve(), app(), event()) at all since they only ship in laravel/framework. Zero practical benefit for a Laravel package (hosts already install the framework) and it degrades security-advisory visibility. If guarding undeclared symbols is the goal, a composer-require-checker CI job on top of the current dependency does it properly.
• Usage cache semantics (upstream Usage counts for Gemini vs. OpenAI are inconsistent in completionTokens. prism-php/prism#1017): convention is promptTokens = non-cached (Anthropic, OpenAI Responses, DeepSeek conform). Found genuine double counting in OpenRouter, Vertex, and Gemini (implicit caching) — they report cacheReadInputTokens while leaving those tokens inside promptTokens. OpenAI's chat_completions path also silently drops cache detail its Responses path reports. Azure/Groq/Qwen/xAI/Z/Requesty ignore available cached-token fields (safe but blind). The double-counting trio is a legitimate-bug fix candidate pending maintainer go-ahead.

[wishborn]

Closing out — everything on this issue is resolved as of v0.109.0:

• ✅ Tool-argument coercion (Systemic tool-call normalization bug: multiple providers crash with raw TypeError on missing or scalar arguments prism-php/prism#1007/Feature request: auto-convert string values to BackedEnum instances when tool handler parameters are typed as enums prism-php/prism#1016) — shipped v0.105.0
• ✅ Groq name-split heuristic test coverage — shipped v0.105.0
• ✅ Mistral structured tool-loop fixture tests + OpenRouter parameterless-tool streaming fixture — merged to main pre-v0.108
• ✅ Usage cache semantics (Usage counts for Gemini vs. OpenAI are inconsistent in completionTokens. prism-php/prism#1017/fix(deepseek): extract prompt_cache_hit_tokens and reasoning_tokens from usage prism-php/prism#1021): double counting fixed for Gemini/Vertex/OpenRouter in v0.108.1, and v0.109.0 completes the sweep — OpenAI chat_completions parity, cache visibility for Azure/Groq/Qwen/xAI, and a double-counting fix for the Z.AI and Requesty streams. All 18 providers now follow the same convention: promptTokens = non-cached input, cacheReadInputTokens reported separately where the API exposes it.
• ❌ Upstream fix(prism): avoid requiring full laravel framework prism-php/prism#1025 (replace laravel/framework with illuminate/* components): rejected permanently per the analysis above — provably incomplete (misses illuminate/concurrency, lists unused illuminate/bus, cannot declare Foundation helpers), no practical benefit for a Laravel-only package, and it degrades security-advisory visibility. The legitimate goal behind it (catching undeclared symbol usage) is now guarded properly: v0.109.0 adds a composer-require-checker CI workflow, and the audit it ran surfaced four genuinely undeclared direct dependencies (ext-mbstring, ext-openssl, psr/http-message, symfony/http-foundation) which are now declared.

Carried as a note in the PM index (not worth an open issue): revisit expiring the o3-mini/chatgpt-4o prefix handling in StructuredModeResolver once model naming stabilizes.