#2235 077c680 Thanks @AndreyHirsa! - Bump figlet from 1.9.4 to 1.11.4.

#2230 2893846 Thanks @AndreyHirsa! - Bump @markdoc/markdoc from 0.5.4 to 0.5.10 and csv-stringify from 6.6.0 to 6.9.0.

#2208 8dfdf49 Thanks @cherkanovart! - Clear two npm audit advisories that surfaced through transitive dependencies.

csv-parse moves from 5.6.0 to 7.0.2, which patches GHSA-8cw4-87c7-c6xx (prototype replacement reachable through the columns option). None of the options renamed in csv-parse 6.0.0 were in use, so the CSV loaders are unchanged.

The AI SDK packages move to the latest release of the major they were already on, so that @ai-sdk/provider-utils resolves at or above 4.0.33 and clears GHSA-866g-f22w-33x8 (uncontrolled resource consumption). The SDK packages pin @ai-sdk/provider-utils to an exact version, so bumping them is the only way to reach it: ai goes to 6.0.280 and @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/groq, @ai-sdk/mistral and @ai-sdk/openai follow on their 3.x line.

The system prompt moves from a { role: "system" } entry in messages to the system option on generateText. Since ai 6.0.170 the former makes the SDK print a prompt-injection warning on every BYOK run.

The basic translator also stops wrapping its system prompt in a serialized { role, content } envelope before handing it over, so the provider now receives the prompt text on its own rather than a line of JSON quoting it. This is the one change here that alters what reaches the model.

#2208 8dfdf49 Thanks @cherkanovart! - Clear two npm audit advisories that surfaced through transitive dependencies.

csv-parse moves from 5.6.0 to 7.0.2, which patches GHSA-8cw4-87c7-c6xx (prototype replacement reachable through the columns option). None of the options renamed in csv-parse 6.0.0 were in use, so the CSV loaders are unchanged.

The AI SDK packages move to the latest release of the major they were already on, so that @ai-sdk/provider-utils resolves at or above 4.0.33 and clears GHSA-866g-f22w-33x8 (uncontrolled resource consumption). The SDK packages pin @ai-sdk/provider-utils to an exact version, so bumping them is the only way to reach it: ai goes to 6.0.280 and @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/groq, @ai-sdk/mistral and @ai-sdk/openai follow on their 3.x line.

The system prompt moves from a { role: "system" } entry in messages to the system option on generateText. Since ai 6.0.170 the former makes the SDK print a prompt-injection warning on every BYOK run.

The basic translator also stops wrapping its system prompt in a serialized { role, content } envelope before handing it over, so the provider now receives the prompt text on its own rather than a line of JSON quoting it. This is the one change here that alters what reaches the model.

Updated dependencies [8dfdf49]:

• lingo.dev@0.138.8

#2208 8dfdf49 Thanks @cherkanovart! - Clear two npm audit advisories that surfaced through transitive dependencies.

csv-parse moves from 5.6.0 to 7.0.2, which patches GHSA-8cw4-87c7-c6xx (prototype replacement reachable through the columns option). None of the options renamed in csv-parse 6.0.0 were in use, so the CSV loaders are unchanged.

The AI SDK packages move to the latest release of the major they were already on, so that @ai-sdk/provider-utils resolves at or above 4.0.33 and clears GHSA-866g-f22w-33x8 (uncontrolled resource consumption). The SDK packages pin @ai-sdk/provider-utils to an exact version, so bumping them is the only way to reach it: ai goes to 6.0.280 and @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/groq, @ai-sdk/mistral and @ai-sdk/openai follow on their 3.x line.

The system prompt moves from a { role: "system" } entry in messages to the system option on generateText. Since ai 6.0.170 the former makes the SDK print a prompt-injection warning on every BYOK run.

The basic translator also stops wrapping its system prompt in a serialized { role, content } envelope before handing it over, so the provider now receives the prompt text on its own rather than a line of JSON quoting it. This is the one change here that alters what reaches the model.

#2197 9db8613 Thanks @cherkanovart! - Fix run --key, which matched nothing and could overwrite unrelated lockfile entries.

--key filtered with a raw glob match, but flat buckets join nesting with /, so a prefix like auth/login matched no key at all and the run reported everything as cached. It now matches on exact key, on a prefix that ends at a /, or on a glob. auth/login selects auth/login/title and leaves auth/login_url and sign-in-error alone.

A --key run also wrote checksums for every source key, not just the translated subset, which marked untouched keys as translated in i18n.lock. --key now suppresses the checksum write, as --target-locale already did, and no longer computes the discarded checksums at all.

The help text for --key documented dot-separated paths and an auth.login example, neither of which matched real keys. It now states the / separator, that a prefix must end at one, and that a glob does not cross one.

The --frozen failure message told the user to run lingo.dev lockfile, which does nothing once the lockfile section is populated. It now points at lingo.dev run, which localizes what is pending and updates the lockfile.