🤖 chore: upgrade AI SDK to v7 - #3707
Merged
Merged
Conversation
- Bump ai 6.0.175 → 7.0.19 and @ai-sdk/* providers to v7-compatible majors - Normalize AI SDK 7 nested usage (inputTokenDetails/outputTokenDetails) into mux's persisted flat shape at all ingestion boundaries; re-inject Anthropic cache-write tokens into providerMetadata for stable pricing/display - streamText result semantics: usage now spans all steps (old totalUsage); final-step values read from finalStep - allowSystemInMessages for cached system-message prepending (streamManager, advisor) - Migrate devtools middleware + spec gates from LanguageModelV3 to V4 - Pin xai chat-completions API (v7 defaults to Responses API) - Explicitly disable OpenAI reasoningSummary for unsupported models (v7 defaults to detailed) - Preserve MCP redirect-following for user-configured servers (v7 defaults to error) - ToolExecutionOptions is now generic with required context; media tool-result parts removed
- Migrate deprecated stream.totalUsage readers to stream.usage (memory consolidation/harvest, status generator) - Pin streamText<ToolSet> generics so Tool's any-context doesn't infect the inferred result type (no-unsafe-return) - Add behavioral tests for normalizeUsage/withCacheWriteMetadata - Fix tests hit by v7 behavior changes: RFC 8414 canonical issuer for MCP OAuth metadata, provider class-name suffix matching, v7 usage promise name
Member
Author
|
@codex review |
CI's flake-hash-check computed the new offlineCache outputHash after the bun.lock changes from the AI SDK v7 dependency bumps.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c4828eee89
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The 2.9.x line peers with ai@^6; 3.0.0 targets ai@^7 and emits v4-spec models consistent with the other upgraded providers. Addresses Codex review feedback on PR #3707.
Member
Author
|
@codex review |
|
Codex Review: Didn't find any major issues. Hooray! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
AI SDK 7's convertToModelMessages requires FileUIPart.url to parse as a real URL (new URL(part.url)) and inlines data: URLs itself. The v6-era rewrite to raw base64 now throws 'Invalid URL', breaking every image/PDF attachment send. Rebuild the canonical data:<mediaType>;base64,<payload> form instead — still normalizing URL-encoded (non-base64) data URIs, which the SDK's splitDataUrl would otherwise corrupt. Also update flake.nix offline-cache hash for the OpenRouter bump's bun.lock change (hash from CI's flake-hash-check diff). Verified: tests/ipc/streaming/sendMessage.images.test.ts 5/5 pass locally (with bridge OPENAI_BASE_URL unset; the 403s under the bridge URL are a local env mismatch, not a product issue).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Upgrades the AI SDK from v6 (
ai@6.0.175) to v7 (ai@7.0.19), bumping all@ai-sdk/*providers to their v7-compatible majors, and adapts mux's usage accounting, streaming, middleware, and provider configuration to the v7 breaking changes — while keeping mux's persisted data shapes (chat.jsonl, session-usage.json, analytics ETL) fully backward compatible with no data migration.Background
AI SDK 7 shipped 2026-06-25 with several breaking changes that affect mux directly:
usage.inputTokenDetails.{noCache,cacheRead,cacheWrite}Tokensandusage.outputTokenDetails.reasoningTokens; Anthropic cache-write tokens moved offproviderMetadata.anthropic.cacheCreationInputTokensonto usage.result.usagenow accumulates across ALL steps (oldtotalUsage, now deprecated); final-step values moved toresult.finalStep.{ role: "system" }insidemessagesthrows unlessallowSystemInMessages: true— mux prepends a cached system message for Anthropic prompt caching.specificationVersion: "v4"models;LanguageModelV3Middlewaregates no longer match.reasoningSummarydefaults to"detailed"whenever reasoning effort is set, and MCP HTTP/SSE transports reject redirects by default.ToolExecutionOptions<CONTEXT>is now generic with a requiredcontextfield; tool-resultmediaparts were removed fromToolResultOutput.Notably, ESM-only
ai@7is NOT a blocker for mux's CommonJS main-process build:require(esm)works on Node 22 / Electron 40, verified against the compileddist/output.Implementation
src/common/utils/tokens/usageHelpers.ts): newnormalizeUsage()andwithCacheWriteMetadata()convert live v7 usage into mux's persisted flat V2 shape and re-inject Anthropic cache-write tokens intoproviderMetadata.anthropic.cacheCreationInputTokens. Applied at every ingestion point: StreamManager (finish-stepevents,getStreamMetadata,getAggregatedProviderMetadata),SessionUsageService.recordHeadlessUsage(choke point for all headless callers), advisor tool, and/btwside questions. Downstream consumers (pricing, IPC, ETL, historical rows) are untouched;createDisplayUsageadditionally self-heals nested shapes defensively.getStreamMetadatareadsusage(all-steps) +finalStep.usage/finalStep.providerMetadata(context window display); deprecatedtotalUsagereaders migrated (memory consolidation/harvest, status generator).allowSystemInMessages: trueon the two streamText calls that receive prepended cached system messages (StreamManager, advisor transcript).LanguageModelV3*→LanguageModelV4*(structurally identical spec); v3 spec gates inaiService/providerModelFactorynow check"v4".provider.chat();reasoningSummary: nullsent explicitly for models that reject the parameter; MCP transports passredirect: "follow"(user-configured, trusted URLs).ToolExecutionOptions<unknown>+context: undefinedacross ~35 test files and 3 production call sites; media tool-result fixtures cast (production rewrite pathextractToolMedia*already strips them before providers see them).Validation
/btwside question, and a follow-up turn.session-usage.jsonexactly; second turn showed 96.9k cache reads, proving prompt caching still hits across turns underallowSystemInMessages.reasoningTokens), DevToolsrawRequestcapture through the V4 middleware, and headless (title/status generator) spend recording all verified against sandbox session files.require()smoke test ofdist/node/services/streamManager.js+ friends against the ESM-only ai@7 graph.Risks
Pains
Tooltype defaults its context generic toany, which infectsstreamText's inferred result type and tripsno-unsafe-return; fixed by explicitly pinningstreamText<ToolSet>.@ai-sdk/mcpv2 strictly validates OAuth metadata issuers (RFC 8414 canonical form), which surfaced as opaquestartDesktopFlowfailures in tests until the underlying error was extracted._(e.g._OpenAIChatLanguageModel), breakingconstructor.nameequality assertions in tests.Generated with
mux• Model:anthropic:claude-fable-5• Thinking:xhigh• Cost:$79.25