[omartelo]

Lets a Claude Code session inside lich run one of its own built-in slash commands when the user asks for it ("run /compact"). This is plugin-only. Nothing goes through lich, whose control refuses a session acting on itself on purpose (lich docs/ceilings.md, "A session cannot control itself").

Design

hooks/self-command.js (registered by hooks/lich.js) works through the Skill tool, the tool the model already reaches for:

• tool.describe on Skill appends one paragraph saying a built-in slash command can be named there when the user asks, and that it runs once the turn ends.
• tool.call on Skill runs the native tool first, so skills and the built-ins it serves as prompts (/init, /review) stay its own. The mod steps in only when that call errored and $.command.list() names the command as builtin. It then answers in the Skill tool's inline shape ({ success, commandName }), with a context note saying the command is queued and the turn should end. It calls $.command.run from $.clock.after(0), because $.command.run rejects inside a hook the turn is waiting on. The command runs once the session is idle, after the turn that asked for it.
• /model and /effort get a deny that points at lich's per-session override, and never run. Through a mod, Claude Code saves them as the default for every new session.
• If a queued command then fails, its error goes to a toast.
• The mod is active only with LICH_SESSION_ID set in an interactive session.

Measured (Claude Code 2.1.295)

• A tool of the mod's own is not possible. Claude Code lists it as mcp__lich__<name>, and $.tool.register refuses with the session already has an MCP server named "lich" in its MCP config; registering would replace it. That covers every lich session.
• Before this change, asked "roda /compact pra mim", the model answered that it could not, since /compact is a CLI command and not a skill. Skill(compact) returns isError with compact is a built-in CLI command, not a skill….
• With this branch (--plugin-dir, lich's MCP server in the config, the same prompt): the model called Skill(compact) and said it would run once the turn ended. The turn ended, and /compact then ran and compacted the conversation. Claude Code accepted the inline result.
• /model, /effort and /cost were not run live.

Known ceilings (in docs/self-command.md)

• The transcript shows the call as "Successfully loaded skill".
• A command that opens a dialog (/cost) holds the command queue until Esc is pressed. $.command.list() cannot tell such commands apart, so none is refused.

Changes

• hooks/self-command.js, registered in hooks/lich.js.
• Docs: docs/self-command.md, plus CLAUDE.md, README.md, docs/providers.md, and CHANGELOG under [Unreleased] › Added.
• Tests: tests/self-command.test.mjs (9) and tests/self-command.engine.test.ts (2). The kit has no command.list, so a test hook stands in for the session's list.

No version bumped.

Tests

• node --test tests/*.test.mjs: 372 pass, 0 fail.
• claude plugin test .: 31 pass, 0 fail.
• claude plugin validate .claude-plugin/plugin.json: passes. The only warnings are ones that already existed, plus a "gating hook without .catch" warning on the new tool.call hook, the same one every other mod already gets.