fix(ai): replay opaque Responses item ids - #56879

Fixes #55357.

Problem

When replaying Responses history, itemID() only kept stored item ids shaped prefix_suffix and dropped everything else. Providers that issue other valid ids lost them on replay:

• Strict gateways that require id on replayed reasoning items (fix(ai): replayed Responses reasoning items are sent without id when the server's item ids have no prefix_suffix shape #55357, bare hex ids like 8a482c218291d43b) failed every follow-up turn with 400 … missing required properties including: 'id'.
• DeepSeek (bare UUIDs), DigitalOcean (bare hex), Novita (bare UUIDs), and Groq Qwen (zdcp8tq5p) had their own ids silently dropped.

The rule was meant to keep foreign ids out after a model switch, but core already handles that: provider metadata, including item ids, is only replayed to the provider and model that produced it.

to-llm-message (core)
  same provider + model, no error?  → keep stored item ids and reasoning state
  otherwise                         → strip them (text/tool calls sent without id)
        ↓
itemID() (packages/ai)
  before: keep only `prefix_suffix` ids
  after:  keep any whitespace-free, non-empty id

Change

 const itemID = (providerMetadata, providerMetadataKey) => {
   const metadata = providerMetadata?.[providerMetadataKey]
   if (!isRecord(metadata) || typeof metadata.itemId !== "string") return undefined
-  const separator = metadata.itemId.indexOf("_")
-  return separator > 0 && separator < metadata.itemId.length - 1 ? metadata.itemId : undefined
+  return /^\S+$/.test(metadata.itemId) ? metadata.itemId : undefined
 }

Ids are opaque and provider-defined, so no shape rule fits every provider:

Empty and whitespace-containing ids are still dropped, since some servers reject "".

Out of scope

• Tool call ids (call_id) are a different path. They are kept across model switches to pair calls with results, and some foreign shapes can still break some targets (bug(llm): cross-provider tool-call IDs can exceed OpenAI-compatible length limit and permanently break sessions #54146, Duplicate function_call_output for reused tool-call IDs makes session unrecoverable #54588). A follow-up will add one shared, collision-safe normalizer per target.
• A gateway that requires id on every replayed assistant message still rejects text that switched in from another provider, because core correctly strips foreign ids.

Verification

• packages/ai: bun test (1822 pass, 0 fail) and bun typecheck pass. Updated the two tests that pinned the old rule, adding real id shapes (UUIDs, bare hex, :-joined, suffix-style) and keeping empty and whitespace ids dropped.
• Live, through the real CLI from source, with request logging, against this branch and dev:
  • 290 sessions and 383 model switches across OpenAI, Azure, xAI, Meta, Anthropic, Google, DeepSeek, Mistral, Bedrock, OpenRouter, Fireworks, Groq, and OpenCode Zen. Every turn forced a tool call, including error, interrupted, and return-to-previous-model turns. No request failed because of replayed state, and no item id or reasoning reached a different provider. Results were the same on both checkouts.
  • Bare-id providers (DeepSeek, DigitalOcean kimi/glm/deepseek, Novita, Groq Qwen) in same-model, switch-away-and-back, and switch-in sessions: all pass on this branch, and their own ids are now replayed verbatim.
  • A local strict gateway reproducing fix(ai): replayed Responses reasoning items are sent without id when the server's item ids have no prefix_suffix shape #55357: every turn fails on dev with the reported 400, and every turn passes on this branch.
• Direct replay probes against about 45 provider and model setups: replaying a provider's own ids verbatim was accepted everywhere.

Added by the opencode bot (AI-generated).

Related: #56776