v0.8.0
Minor Changes
-
cf92a17: Add
permissionModetoclaudeCode()andapprovalsReviewertocodex()— provider-level options for AI-mediated per-tool approval, an alternative to full bypass for AFK host runs (noSandbox()+run()).claudeCode({ permissionMode: "auto" })emits--permission-mode autoinstead of--dangerously-skip-permissions. Accepts any of Claude's permission modes:default,acceptEdits,plan,auto,dontAsk,bypassPermissions.codex({ approvalsReviewer: "auto_review" })swaps--dangerously-bypass-approvals-and-sandboxfor-a on-request -s danger-full-access -c approvals_reviewer="auto_review"so Codex's reviewer agent evaluates each approval prompt.
Patch Changes
- 932302b: Bump the Codex default model from
gpt-5.4-minitogpt-5.4insandcastle initscaffolding and the interactive agent picker. The previous default was underpowered for implementation work. - c6c3026: Fix
opencode()interactive sessions (and theinitscaffold's opencodesetupCommand) seeding the prompt with-p, which is theopencode run/attachbasic-auth password flag, not a prompt seed. Use--prompt(the TUI's long-form-only seed flag) instead. The TUI pre-fills the textbox but does not auto-submit (see sst/opencode#3937).