@autotranslate/zod@1.0.0-beta.10

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10

@autotranslate/vite@1.0.0-beta.20

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10
  • @autotranslate/cli@1.0.0-beta.20

@autotranslate/react@1.0.0-beta.11

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10

@autotranslate/providers@1.0.0-beta.10

Minor Changes

• bb38cd5
  Thanks @tamimbinhakim! - Add the agent
  provider - translate with Claude Code or Codex

  Drives a headless coding-agent CLI already installed and signed in on the
  machine, instead of calling a model API. No API key, no separate billing.

  provider: {
    name: 'agent',
    agent: 'claude',            // or 'codex'
    model: 'claude-haiku-4-5',  // optional
  }

  The agent is used strictly as a text transformer - tools disabled, sandbox
  read-only, prompt on stdin, JSON out. It cannot touch your repository.

  • Claude Code: claude -p --output-format json --disallowedTools '*'.
    Fenced or prose-wrapped answers are unwrapped by a balanced-brace scan.
  • Codex: codex exec --sandbox read-only --output-schema, so the response
    shape is enforced by the API rather than by prompt discipline.

  Options: agent, model, command, args, timeoutMs (default 300000).

  Each batch is a process spawn, so this suits local development and
  small-to-medium catalogs rather than bulk runs; keep concurrency modest. CI
  runners do not carry your agent login, so prefer ai with a key there.

  Batching, retry with backoff, and re-asking for omitted keys are shared with
  the ai provider.

• bb38cd5
  Thanks @tamimbinhakim! - Fix incremental
  translation: committed catalogs are now the diff baseline

  init gitignored .translations/.cache/, and translate used that cache as
  its only record of what had been translated - it never consulted the committed
  catalogs. On any machine without the cache (fresh clone, new teammate, every
  CI run) the entire catalog was re-sent to the provider despite valid
  translations sitting in the repo.

  Three defects compounded it:

  • Omitted keys deleted committed translations. A key missing from a
    provider response (routine on large batches) was pruned from the catalog on
    write.
  • One failed chunk discarded the run. Catalogs were written only after
    every task settled, so a single 429 threw away every other chunk's work.
  • Reference context was unbounded. Every cached neighbour rode along on
    every request, growing prompts with catalog size - which is what pushed
    large batches past the point where models answer completely.

  Changes

  • Translation state moves to .translations/.state/<locale>/<chunk>.json,
    recording the source hash behind each translation. Commit it - init no
    longer ignores anything under outDir and strips a stale .cache/ line if
    it finds one. Existing .cache/ layouts migrate on the next translate.
  • Switching provider or model no longer invalidates anything.
  • Omitted keys keep their previous translation and stay out of state, so the
    next run retries exactly those.
  • Batches fail independently. TranslateResult gains failures; the CLI
    exits non-zero when any batch failed.
  • Reference context is capped at 8 items.

  Batching moves to the CLI

  translate now plans the whole run, collapses duplicate copy, and splits the
  remainder into uniform batches instead of inheriting whatever landed in a hash
  bucket.

  • New batchSize config option (default 25). This is the lever when a model
    drops items.
  • Identical copy under several keys is translated once and fanned out; copy
    with differing context or description still translates separately.
  • The ai provider retries transient failures with backoff and re-asks for
    omitted keys in smaller slices.

  TranslateProgress reports batch / totalBatches / items instead of
  chunkPath.

  Upgrading

  Run npx autotranslate init once, then commit .translations/. init
  removes the stale .gitignore entry and the next translate migrates
  .cache/ into .state/, so no retranslation is needed.

  Until .state/ is committed, CI still has no diff input and will retranslate
  the full catalog on each run - exactly the old behaviour, no worse. Projects
  that never commit .translations/ at all are unaffected.

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10

@autotranslate/next@1.0.0-beta.20

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10
  • @autotranslate/cli@1.0.0-beta.20
  • @autotranslate/react@1.0.0-beta.11

@autotranslate/eslint-plugin@1.0.0-beta.10

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10

@autotranslate/core@1.0.0-beta.10

Minor Changes

• bb38cd5
  Thanks @tamimbinhakim! - Add the agent
  provider - translate with Claude Code or Codex

  Drives a headless coding-agent CLI already installed and signed in on the
  machine, instead of calling a model API. No API key, no separate billing.

  provider: {
    name: 'agent',
    agent: 'claude',            // or 'codex'
    model: 'claude-haiku-4-5',  // optional
  }

  The agent is used strictly as a text transformer - tools disabled, sandbox
  read-only, prompt on stdin, JSON out. It cannot touch your repository.

  • Claude Code: claude -p --output-format json --disallowedTools '*'.
    Fenced or prose-wrapped answers are unwrapped by a balanced-brace scan.
  • Codex: codex exec --sandbox read-only --output-schema, so the response
    shape is enforced by the API rather than by prompt discipline.

  Options: agent, model, command, args, timeoutMs (default 300000).

  Each batch is a process spawn, so this suits local development and
  small-to-medium catalogs rather than bulk runs; keep concurrency modest. CI
  runners do not carry your agent login, so prefer ai with a key there.

  Batching, retry with backoff, and re-asking for omitted keys are shared with
  the ai provider.

• bb38cd5
  Thanks @tamimbinhakim! - Fix incremental
  translation: committed catalogs are now the diff baseline

  init gitignored .translations/.cache/, and translate used that cache as
  its only record of what had been translated - it never consulted the committed
  catalogs. On any machine without the cache (fresh clone, new teammate, every
  CI run) the entire catalog was re-sent to the provider despite valid
  translations sitting in the repo.

  Three defects compounded it:

  • Omitted keys deleted committed translations. A key missing from a
    provider response (routine on large batches) was pruned from the catalog on
    write.
  • One failed chunk discarded the run. Catalogs were written only after
    every task settled, so a single 429 threw away every other chunk's work.
  • Reference context was unbounded. Every cached neighbour rode along on
    every request, growing prompts with catalog size - which is what pushed
    large batches past the point where models answer completely.

  Changes

  • Translation state moves to .translations/.state/<locale>/<chunk>.json,
    recording the source hash behind each translation. Commit it - init no
    longer ignores anything under outDir and strips a stale .cache/ line if
    it finds one. Existing .cache/ layouts migrate on the next translate.
  • Switching provider or model no longer invalidates anything.
  • Omitted keys keep their previous translation and stay out of state, so the
    next run retries exactly those.
  • Batches fail independently. TranslateResult gains failures; the CLI
    exits non-zero when any batch failed.
  • Reference context is capped at 8 items.

  Batching moves to the CLI

  translate now plans the whole run, collapses duplicate copy, and splits the
  remainder into uniform batches instead of inheriting whatever landed in a hash
  bucket.

  • New batchSize config option (default 25). This is the lever when a model
    drops items.
  • Identical copy under several keys is translated once and fanned out; copy
    with differing context or description still translates separately.
  • The ai provider retries transient failures with backoff and re-asks for
    omitted keys in smaller slices.

  TranslateProgress reports batch / totalBatches / items instead of
  chunkPath.

  Upgrading

  Run npx autotranslate init once, then commit .translations/. init
  removes the stale .gitignore entry and the next translate migrates
  .cache/ into .state/, so no retranslation is needed.

  Until .state/ is committed, CI still has no diff input and will retranslate
  the full catalog on each run - exactly the old behaviour, no worse. Projects
  that never commit .translations/ at all are unaffected.

@autotranslate/cli@1.0.0-beta.20

Minor Changes

• bb38cd5
  Thanks @tamimbinhakim! - Add the agent
  provider - translate with Claude Code or Codex

  Drives a headless coding-agent CLI already installed and signed in on the
  machine, instead of calling a model API. No API key, no separate billing.

  provider: {
    name: 'agent',
    agent: 'claude',            // or 'codex'
    model: 'claude-haiku-4-5',  // optional
  }

  The agent is used strictly as a text transformer - tools disabled, sandbox
  read-only, prompt on stdin, JSON out. It cannot touch your repository.

  • Claude Code: claude -p --output-format json --disallowedTools '*'.
    Fenced or prose-wrapped answers are unwrapped by a balanced-brace scan.
  • Codex: codex exec --sandbox read-only --output-schema, so the response
    shape is enforced by the API rather than by prompt discipline.

  Options: agent, model, command, args, timeoutMs (default 300000).

  Each batch is a process spawn, so this suits local development and
  small-to-medium catalogs rather than bulk runs; keep concurrency modest. CI
  runners do not carry your agent login, so prefer ai with a key there.

  Batching, retry with backoff, and re-asking for omitted keys are shared with
  the ai provider.

• bb38cd5
  Thanks @tamimbinhakim! - Add a git merge
  driver for the generated catalogs

  Now that translation state is committed, two branches that both add copy will
  both write to .translations/. Git's line-based merge conflicts on adjacent
  keys and writes conflict markers into JSON, which breaks every consumer.

  autotranslate init registers a merge driver and the matching
  .gitattributes rules:

  • Catalogs merge to the union of both sides' keys; deletions are honoured.
  • State drops any key both branches retranslated, so the next translate
    re-derives that one key instead of picking an arbitrary winner.
  • index.ts is regenerated from the union of both sides' chunk files, and
    now emits one import per line so it stays diff-friendly.

  Conflicts in your own source still surface normally; only generated files
  resolve automatically.

  .gitattributes is committed, but the driver command lives in .git/config,
  which git does not share. Teammates run autotranslate init once to enable
  it.

  Also adds autotranslate merge-driver <base> <ours> <theirs> [path], invoked
  by git rather than by hand.

• bb38cd5
  Thanks @tamimbinhakim! - Fix incremental
  translation: committed catalogs are now the diff baseline

  init gitignored .translations/.cache/, and translate used that cache as
  its only record of what had been translated - it never consulted the committed
  catalogs. On any machine without the cache (fresh clone, new teammate, every
  CI run) the entire catalog was re-sent to the provider despite valid
  translations sitting in the repo.

  Three defects compounded it:

  • Omitted keys deleted committed translations. A key missing from a
    provider response (routine on large batches) was pruned from the catalog on
    write.
  • One failed chunk discarded the run. Catalogs were written only after
    every task settled, so a single 429 threw away every other chunk's work.
  • Reference context was unbounded. Every cached neighbour rode along on
    every request, growing prompts with catalog size - which is what pushed
    large batches past the point where models answer completely.

  Changes

  • Translation state moves to .translations/.state/<locale>/<chunk>.json,
    recording the source hash behind each translation. Commit it - init no
    longer ignores anything under outDir and strips a stale .cache/ line if
    it finds one. Existing .cache/ layouts migrate on the next translate.
  • Switching provider or model no longer invalidates anything.
  • Omitted keys keep their previous translation and stay out of state, so the
    next run retries exactly those.
  • Batches fail independently. TranslateResult gains failures; the CLI
    exits non-zero when any batch failed.
  • Reference context is capped at 8 items.

  Batching moves to the CLI

  translate now plans the whole run, collapses duplicate copy, and splits the
  remainder into uniform batches instead of inheriting whatever landed in a hash
  bucket.

  • New batchSize config option (default 25). This is the lever when a model
    drops items.
  • Identical copy under several keys is translated once and fanned out; copy
    with differing context or description still translates separately.
  • The ai provider retries transient failures with backoff and re-asks for
    omitted keys in smaller slices.

  TranslateProgress reports batch / totalBatches / items instead of
  chunkPath.

  Upgrading

  Run npx autotranslate init once, then commit .translations/. init
  removes the stale .gitignore entry and the next translate migrates
  .cache/ into .state/, so no retranslation is needed.

  Until .state/ is committed, CI still has no diff input and will retranslate
  the full catalog on each run - exactly the old behaviour, no worse. Projects
  that never commit .translations/ at all are unaffected.

Patch Changes

• Updated dependencies
  [bb38cd5,
  bb38cd5]:
  • @autotranslate/core@1.0.0-beta.10
  • @autotranslate/providers@1.0.0-beta.10

@autotranslate/zod@1.0.0-beta.9

Patch Changes

• Updated dependencies
  [e0aa5ca]:
  • @autotranslate/core@1.0.0-beta.9

@autotranslate/zod@1.0.0-beta.8

Patch Changes

• Updated dependencies
  [d90aae1]:
  • @autotranslate/core@1.0.0-beta.8