[agentHits]

What changed

Dependency-only upgrade in apps/dokploy and packages/server (same versions in both, resolved with a 3-day minimum release age):

Package	From	To
ai	6.0.300	7.0.127
@ai-sdk/anthropic	3.0.127	4.0.71
@ai-sdk/azure	3.0.133	4.0.90
@ai-sdk/cohere	3.0.67	4.0.54
@ai-sdk/deepinfra	2.0.85	3.0.62
@ai-sdk/mistral	3.0.70	4.0.56
@ai-sdk/openai	3.0.124	4.0.83
@ai-sdk/openai-compatible	2.0.81	3.0.62
ai-sdk-ollama	3.8.8	4.4.0

Compatibility of the set: every provider and ai resolve to a single shared @ai-sdk/provider@4.0.21 / @ai-sdk/provider-utils@5.0.53; ai-sdk-ollama@4.4.0 has peerDependencies.ai: ^7.0.103; all @ai-sdk/* peer on zod ^3.25.76 || ^4.1.8 (we use zod 4.6). New transitive deps: @workflow/serde, undici 7 inside provider-utils; ai no longer depends on @opentelemetry/api.

No source changes were needed.

Release notes / migration guide (v6 -> v7)

Checked the official 6 -> 7 migration guide against the three call sites (select-ai-provider.ts, services/ai.ts, routers/ai.ts):

• Node 22+ and ESM only: we run Node 24 and both packages are already "type": "module" with ESM esbuild output.
• experimental_output removed: we already use output: Output.object(...) and result.output.
• system -> instructions, system messages in prompt rejected by default: we only pass a plain prompt string.
• usage / content / toolCalls now aggregate all steps, finalStep added: we only read result.text and result.output, single step.
• Request/response bodies excluded from results by default, telemetry moved to @ai-sdk/otel, lifecycle callback renames: not used.
• OpenAI Responses reasoning summary defaults to detailed (only when reasoning is set): not used.
• Anthropic providerMetadata.anthropic.cacheCreationInputTokens removed: not used.

The codemod was not needed: none of the patterns it rewrites appear in the code.

Verification

• bun run typecheck: all 4 workspaces pass.
• AI tests: __test__/ai/* and __test__/placement/deploy-target-boundary.test.ts: 25/25 pass.
• Full apps/dokploy suite: 2055 passed; failures are only *.real.test.ts (need Docker) plus better-auth-schema-compat and cloud-queue-availability timeouts, which fail the same way on AgentHits-Dev without this change.
• bunx biome check on the changed manifests: clean.
• Node 24 imports ai and every provider as ESM.
• Runtime smoke test (temporary script, not committed): the real selectAIProvider for all 14 provider branches (OpenAI, Azure, Azure /v1, Anthropic, Cohere, Perplexity, Mistral, Ollama, DeepInfra, Gemini, OpenRouter, Z.ai, MiniMax, custom) with fetch intercepted, calling generateText plain (as in test connection) and with Output.object using the suggestVariants schema. Request URL, auth headers, body keys, response format type and the parsed output were identical on v6 and v7; the only difference is the model specificationVersion (v3 -> v4).
• Not built locally: next build (left to CI).

Manual check on dev (AI settings in the panel)

• Settings -> AI: add or edit a provider, load the models list (uses the provider list endpoints, not the SDK, but worth one click) and Test connection with at least one OpenAI-compatible provider and, if available, Anthropic/Ollama.
• Templates / compose AI suggestions ("generate"): run a prompt and confirm suggestions with docker compose, env variables and domains come back and can be deployed.
• Analyze logs with AI on a build log and a container log.

🤖 Generated with Claude Code