Side-effecting actions (gh PR replies, git push) run without announcement or confirmation

Summary

When given a multi-step task that ends in user-visible network side effects (posting GitHub PR comment replies, pushing commits to a remote), Claude Code ran the operations without first stating "about to do X" or pausing for confirmation. The bash output from the gh api calls was emitted into the transcript, but the actions themselves were not flagged as risky or shared-state-modifying before execution.

Repro

1. Ask Claude Code to "address the comments on the PR."
2. Observe: it edits code, commits, pushes, and posts 5 inline gh api .../pulls/{n}/comments/{id}/replies requests in parallel — all in one turn, with no prior "I'm about to post these 5 replies, ok?" message.

Expected

Per the assistant's own guidelines around "actions visible to others or that affect shared state," posting GitHub comments and pushing to a remote should be either:

• (a) announced in a single user-facing message before the tool call, or
• (b) gated behind confirmation when the prior user instruction is ambiguous about scope ("address the comments" → does that include automatically posting replies, or just editing the code that needs changing?).

Actual

Operations ran with no "here's what I'm about to do" preamble; the only signal was the raw gh api JSON responses dumping into the tool-output portion of the transcript, which most users won't be reading inline.

Environment

• Claude Code in VSCode extension
• Model: claude-opus-4-7[1m]