Skip to content

v0.42.0

Choose a tag to compare

@apollo-watcher-release apollo-watcher-release released this 09 Oct 21:21
· 1 commit to main since this release
Immutable release. Only release title and notes can be modified.

The Watcher client's local web app is removed. Review sessions in the Analyzer and answer escalated tool calls in your coding agent's permission prompt.

  • Changed: The client's local web app and its session, decision and settings APIs are removed. Use the Analyzer for session review, the CLI for settings and command rules, and the macOS menu bar for snooze. Linux no longer has a snooze control. Remove container UI relays, but keep editor port forwarding for browser sign-in. Self-hosted deployments must update the server before clients. See the "Local web app removed from the Watcher client" upgrade guide in the customer docs.

  • Changed: Claude Code escalations use its native permission prompt. Unattended claude -p runs deny escalations immediately; use the Auto profile or run interactively. Built-in profiles now allow question and plan tools. Add managed settings overrides before updating the server if those tools must keep escalating.

  • Changed: Local thresholds outside 0 to 11 are ignored. Run watcher doctor settings and correct invalid values.

  • Changed: API integrations sending both id and call_id for an OpenAI Responses local-shell result must send matching values.

  • Added: watcher rules test, watcher decisions list and watcher settings show replace local web app tools. Settings diagnostics can repair rules that disable automatic approvals, with confirmation and backups.

  • Changed: Sign in through the macOS menu bar, browser or watcher login. The client stops storing local session previews and terminal identifiers.

  • Fixed: Codex grading retains context after rewinds. Identical parallel Claude Code escalations keep separate answers, and busy local databases no longer cause bursts of Codex approval prompts.

  • Changed: The red-team drill uses CLI settings and Analyzer results.

  • Fixed: Codex calls without a recorded Watcher review are described accurately.

  • Added: The LiteLLM gateway example streams approved tool calls and asks the model for alternatives after refusals. Requests require a stable session header or native Claude Code or Codex session metadata.