Plan-mode exit can save the plan and offer a gander. Tech lead and on-call comment on one link, and agent connect hands a change back to the session that read the code.
What people write with Claude Code
Pick a doc type to see who reviews it, and what agent connect changes.
How it works
Author writes, reviewer comments, Claude Code acts. The same hosted link covers the whole loop, before the file ever sees git.
Goal: stop losing carrier callbacks during peak shipping season by retrying failed webhooks with exponential backoff, with a kill switch.
1. Enable for staff accounts only. 2. 10% of production traffic. 3. 50%, then 100% if error budget holds.
Open questions: who owns the cleanup two weeks after 100%? [[anchor]]Do we keep the legacy retry path in process, and for how long?[[/anchor]]
Claude Code writes markdown. Gander watches.
Reviewer opens the link. No install. Inline threads on the live doc.
Agent connect reaches Claude Code. It edits the file. The same link refreshes.
Install
The same install works on macOS and Linux. Windows users should use WSL.
$ curl -fsSL https://release.gander.md/install.sh | bash
$ gander signup --email [email protected]
$ gander watch plan.mdOr: install the Gander skill and let the agent save the plan, watch the file, and poll for review. Skill docs.
Homebrew, flags, and other options: Install docs.
The skill lives at ~/.claude/skills/gander. gander mcp install writes the MCP entry that session reads, so agent connect lands without a paste.
Step-by-step setup: Claude Code docs. Every agent: Use with your agent.
Using a different agent? Other agents.
Leave a thread on the line that should change.