fix(agents): provider robustness — OpenAI strict schema, Anthropic refusals, retries, real request timeout

[epeicher]

Stacked on #466.

Testing

What it does

Fixes four provider failures hit while testing #466:

1. The 400 every agent chat hit on the OpenAI connector (below).
2. Unexpected Anthropic API response: Missing the "content" key. — a model REFUSAL the Anthropic provider plugin crashes on before its own refusal handling.
3. Agent stopped after 8 turns without a final answer. — the model spending the whole turn cap on tool calls.
4. cURL error 28: Operation timed out after 30007 milliseconds — the AI Client's own 30-second request timeout, which feat(agents): enhance agent capabilities and HTTP timeout management #466's http_request_timeout raise cannot reach.

OpenAI: strict structured output 400

Bad Request (400) - Invalid schema for response_format 'response_schema': In context=('properties', 'call_to_actions', 'items'), 'required' is required to be supplied and to be an array including every key in properties. Missing 'style'.

Rationale

The OpenAI provider sends the answer schema verbatim with strict: true, and strict structured output has no optional fields: every key in properties must also appear in required. The agent answer schema listed style as optional in the call-to-action items (and call_to_actions as optional at the root), which OpenAI rejects before generating anything. #466 already covered the other strict rule (additionalProperties: false) and fixed the Copilot/drafts schemas; the agent answer schema's required lists were the remaining gap.

Implementation

• desktop_mode_agent_answer_schema() declares full required lists: [id, label, style, reply] on the items, [text, call_to_actions] at the root. The lenient parse/sanitize path is untouched — a bad or missing style still defaults to secondary for pre-filter runtimes and non-strict providers.
• desktop_mode_ai_normalize_response_schema() (the safety net feat(agents): enhance agent capabilities and HTTP timeout management #466 introduced for additionalProperties) now also repairs partial required lists in the same recursive walk, so a plugin-filtered schema addition cannot reintroduce the 400. Provider-safe: Google's provider already strips response-schema keys its API rejects.

Testing instructions

npm run test:php -- --filter='Tests_DesktopMode_AiResponseSchemaNormalization'   # 16 tests

Manual, with only the OpenAI provider plugin active (deactivate the Google/Anthropic providers to force routing): chat with any agent asking for a proposal that needs approval. Before this fix the answer is the 400 above; after, a structured answer arrives with working call-to-action buttons. Verified live against the OpenAI connector.

Anthropic: refusal handling, retry, forced final answer, real timeout

All reproduced live with only the Anthropic provider active. A raw-body trace of the failing "Missing content" case settled what it actually is:

TURN: status=200 stop=refusal content=[]
BODY: {"model":"claude-sonnet-5", ..., "content":[], "stop_reason":"refusal",
       "stop_details":{"type":"refusal","category":"bio", ...}}

A model refusal (a safety-classifier false positive over innocuous demo content), returned as a 200 with an empty content array — which the provider plugin throws on ("Missing the "content" key") before reaching its own refusal handling. Not a flap: retries usually reproduce it.

• Refusal translation: desktop_mode_agent_humanize_generate_error() maps that parse error to an actionable message ("its safety system most likely declined this request — rephrase, or switch the provider in Settings → Connectors"), preserving the provider's original text in the error data.
• Bounded retry: one retry per generate turn for transient signatures — a failed models-list fetch (No models found …), gateway 5xx, transport timeouts, and the empty-content case (kept because borderline refusals are stochastic and one request is cheap). desktop_mode_agent_generate_error_is_transient() is the pure, unit-tested classifier; deterministic rejections (schema 400s, rate limits) are never retried.
• Forced final answer at the turn cap: when the model spends all DESKTOP_MODE_AGENT_RUNNER_MAX_TURNS turns calling tools, the runner makes one last TOOL-LESS generate over the transcript — with nothing to call, the model can only answer from what it already gathered. Falls back to the original max-turns error if even that fails; turns reports MAX + 1. The desktop_mode_agent_runner_generate filter doc notes it can now be invoked more than once per turn.
• The 30s timeout: Core's WP_AI_Client_Prompt_Builder constructor pins an explicit 30-second timeout via the SDK's RequestOptions, bypassing the WordPress HTTP default that feat(agents): enhance agent capabilities and HTTP timeout management #466's wrapper raises — hence "timed out after 30007 milliseconds" despite the 180s allowance. The wrapper now also raises Core's wp_ai_client_default_request_timeout filter, with identical raise-only + finally-removal scoping.

Verified live with only the Anthropic provider active: the question that produced the 8-turn error now completes (turns 7 and 8 across runs), and a refused translation request surfaces the explained refusal instead of the parse error.