chore(deps/api): bump workers-ai-provider from 0.7.5 to 3.1.14 in /packages/api

[dependabot]

Bumps workers-ai-provider from 0.7.5 to 3.1.14.

Release notes

Sourced from workers-ai-provider's releases.

workers-ai-provider@3.1.14

Patch Changes

• #527 cc4b2c0 Thanks @​mkgada! - Rewrite Workers AI tool call IDs exposed to the AI SDK and restore the original provider IDs when building follow-up prompts to avoid collisions across chat turns.

workers-ai-provider@3.1.13

Patch Changes

• #510 dfd2cb4 Thanks @​Specy! - Map inputTokens.cacheRead and inputTokens.noCache from Workers AI's usage.prompt_tokens_details.cached_tokens instead of always reporting them as undefined. This makes prompt-cache hits visible to consumers that compute pricing or telemetry from LanguageModelV3Usage (generateText/streamText result.usage).

  cached_tokens is treated as cacheRead; cacheWrite remains undefined because the OpenAI-style usage shape Workers AI returns does not distinguish cache reads from writes.

  Closes #509.

workers-ai-provider@3.1.12

Patch Changes

• #504 e9b2a9a Thanks @​threepointone! - Forward reasoning_effort and chat_template_kwargs onto binding.run(model, inputs)'s inputs object instead of silently dropping them into the options arg / REST query string. This fixes reasoning models (GLM-4.7-flash, Kimi K2.5/K2.6, GPT-OSS, QwQ) burning the entire output token budget on chain-of-thought with no visible content.

  Both settings-level and per-call usage are supported:

  // Settings-level
  const model = workersai("@cf/zai-org/glm-4.7-flash", {
  	reasoning_effort: "low",
  	chat_template_kwargs: { enable_thinking: false },
  });
  // Per-call (overrides settings)
  await generateText({
  model,
  prompt,
  providerOptions: {
  "workers-ai": { reasoning_effort: "low" },
  },
  });

  reasoning_effort: null is preserved as-is (explicit "disable reasoning" signal). The two fields are also typed directly on WorkersAIChatSettings.

  Closes #501.

workers-ai-provider@3.1.11

Patch Changes

• #494 ed6112a Thanks @​threepointone! - Emit tool-input-end and tool-call events eagerly when streaming multiple tool calls, instead of deferring all of them to stream close. Previously, all tool calls appeared "in progress" simultaneously because tool-input-end was only emitted in flush(). Now each tool call is closed as soon as the next one starts or a null finalization chunk is received, matching the behavior of other AI SDK providers.

workers-ai-provider@3.1.10

Patch Changes

... (truncated)

Commits

• 5b54331 Merge pull request #534 from cloudflare/hotfix/demo-workers-ai-provider-version
• ab1882b fix: keep demos on published workers-ai-provider version
• 498577b Merge pull request #533 from cloudflare/changeset-release/main
• d97d930 Version Packages
• feae676 Merge pull request #508 from cloudflare/hrushikesh/add-semgrep-oss-workflow
• d6977e3 Merge branch 'main' into hrushikesh/add-semgrep-oss-workflow
• a6b7a0e Merge pull request #527 from mkgada/fix/issue-526-toolcallid-collision
• cc4b2c0 Fix Workers AI tool call ID collisions
• 630180b Merge pull request #519 from cloudflare/changeset-release/main
• 8900a5a cleanups
• Additional commits viewable in compare view

Maintainer changes

This version was pushed to npm by GitHub Actions, a new releaser for workers-ai-provider since your current version.

[andrew-bierman]

@dependabot rebase

[dependabot]

Superseded by #2451.