Work in a neighbour repo from here: reading via --add-dir, editing via a background session in that repo steered with SendMessage. Triggers: "/we:sideload", "load context for", "cross-repo".
Install
npx skillscat add weside-ai/claude-code-plugin/sideload Install via the SkillsCat registry.
/we:sideload
A basename resolves against the sibling directories of this repo; a path is used as-is. Pick the
mode by what the task does in the other repo.
Reading only (answer a question, compare code): the other repo's instructions must load as
instructions, not as a file dump. Start the session withCLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 claude --add-dir <repo>. Measured (2.1.283,
27.09.2026): this loads the repo's CLAUDE.md and .claude/rules/, but not its AGENTS.md, soRead <repo>/AGENTS.md yourself when the repo has one. Without the variable, --add-dir grants
file access only. Mid-session, when no restart is possible: Read the instruction file and the
rules the task touches.
Editing (a change lands in the other repo): never edit it from this session, because its
path-bound rules, hooks and worktree conventions do not apply here. Start a native session there:cd <repo> && claude --bg "<task, with the context a human would give>". It prints an id;claude agents lists it. Steer it with SendMessage to its name from ListAgents, and passnotify_when_idle: true to hear when it finishes instead of polling. A session in a different
permission mode holds cross-session messages for its user's approval, so a send is not proof it
read the message. The other session commits, pushes and reports under its own repo's rules.