fix(providers/openai): emit item_reference for reasoning when store=true

[ibetitsmike]

Disclaimer: opened by a Coder Agent on behalf of @ibetitsmike.

Fixes the gpt-5.5 xhigh (and other Responses API reasoning models) follow-up turn failure when an assistant turn produces reasoning + a provider-executed web_search_call (or any item that requires a preceding reasoning item).

Problem

When Store: true is enabled and an assistant turn contains both a reasoning item and a following provider-executed item (e.g. web_search_call), the OpenAI Responses API requires the reasoning item to be replayed alongside the following item on the next turn. Otherwise it rejects the request with:

Item 'ws_xxx' of type 'web_search_call' was provided without its
required 'reasoning' item: 'rs_xxx'.

Previously toResponsesPrompt unconditionally skipped reasoning items. The skip was introduced to fix charmbracelet#181 (inline OfReasoning replay was rejected when no following item was present), but it produces a dual failure mode for any conversation that combines reasoning with a provider-executed tool call or function call.

Fix

• store=true: replay the reasoning item via item_reference using the persisted ItemID from ResponsesReasoningMetadata. Mirrors how provider-executed tool calls are already replayed and matches the OpenAI documented contract for stored items.
• store=false: server-side reasoning IDs are ephemeral and cannot be referenced. Provider-executed tool calls are likewise skipped, so there is nothing to pair with. Behavior is unchanged.
• Defensive fallback: if ItemID is missing from metadata, the reasoning part is skipped (rest of the assistant message still replays).

Inline OfReasoning replay is intentionally not used to avoid re-introducing the failure mode that charmbracelet#181 fixed.

Tests

Updated and added in providers/openai/openai_test.go:

• TestResponsesToPrompt_ReasoningWithStore — now asserts store=true emits item_reference for the reasoning item; covers the no-ItemID skip path; keeps the store=false skip assertion.
• TestResponsesToPrompt_ReasoningWithWebSearchCombined — full reasoning + web_search_call replay with order assertions (user, item_reference(rs), item_reference(ws), text, user).
• TestResponsesToPrompt_ReasoningWithFunctionCallCombined — reasoning + function_call + function_call_output round-trip with order assertions.

Full suites pass:

• go test ./providers/openai/...
• go test ./...

Related

• Closes the failure mode that closed PR fix(providers/openai): emit item_reference for reasoning items when store=true #7 (this PR is a smaller, focused replacement).
• Builds on fix(openai): skip reasoning items in Responses API replay charmbracelet/fantasy#181 (skip inline reasoning) without reverting it.
• Companion coder/coder PR will bump the pin and add chatd integration tests + a defensive sanitizer.

Implementation plan and decision log

Follow-up to closed PR #7. Plan kept the change minimal:

1. Red: extended TestResponsesToPrompt_ReasoningWithStore and added two combined-replay tests; all failed against the previous unconditional skip.
2. Green: replaced the unconditional continue with a store-gated item_reference emission using GetReasoningMetadata. Build + tests green.
3. Refactor: not needed — pattern only appears once.

Decisions:

• Chose item_reference over inline OfReasoning to avoid re-introducing fix(openai): skip reasoning items in Responses API replay charmbracelet/fantasy#181's failure mode.
• Defensive fallback for missing ItemID keeps existing behavior (skip) instead of erroring.
• store=false path unchanged.