ReviewGate CLI 2.0.4

ReviewGate CLI 2.0.4 — local review against your team's own rules.

What's new

• One retry when the stream breaks. OpenAI-compatible providers (DeepSeek) sometimes cut
  the SSE stream mid-answer; the run used to fail straight away with llm_network, and the
  hook let the push through "unchecked". The call is now retried once — and not silently: a
  WARN in the log, retries in usage.calls[] and in the 🔬 block (tokens from the aborted
  attempt are not counted in that line). Timeouts and cancellations are not retried.
• An honest reason for having no judge in --fast. The 🔬 block and the questions: true
  notice no longer claim "no judges configured" — they name the fast mode as the reason, and
  say "nothing to judge" when there are no findings.
• No machine token in the report footer. Instead of License: Community (absent) — just
  Community, with the reason a key did not apply spelled out in words, the same way the bot
  does it (in server mode, after the bot is updated to 0.2.2).
• DeepSeek as the main provider: json_mode: none and max_tokens: 64000. The
  json_mode: object advice from 2.0.3 is withdrawn: on a real review run with a long prompt,
  json_object returns an empty answer — the JSON goes into reasoning, and reasoning spends
  the same ceiling as the answer. The reviewgate init skeleton, help config and the docs
  are updated; 64000 is not above the Sonnet 5 ceiling.
• The pre-push recipe on /docs/cli now survives husky (|| code=$? instead of case $?):
  under set -e the old recipe locked the push even when the review itself failed, losing the
  soft branch.

Full reference: https://reviewgate.dev/docs/cli.

Files

Every platform under two names: reviewgate-2.0.4-* pins the version, reviewgate-latest-* is
a permanent link. Verify with SHA256SUMS. The primary distribution is
https://reviewgate.dev/dl/ — this repository is a mirror.