Skiller requires canonical agent ids in skiller.toml

Problem

The repo's root prepare script runs bun x skiller@latest apply. That started failing because .claude/skiller.toml still used the old Claude agent id claude.

That broke any workflow that relied on a normal install, including the Next upgrade flow.

Root cause

skiller hard-cut its public config surface to canonical agent ids. claude is no longer accepted; claude-code is the valid id.

The important wrinkle: the generated .claude/.skiller.json can still contain sourceType: "claude" for project-managed items. That internal metadata is not the config surface skiller validates on startup.

Solution

Update .claude/skiller.toml to use canonical ids only:

default_agents = ["claude-code", "codex"]

[agents.claude-code]
enabled = true

Do not patch .claude/.skiller.json by hand. It is generated state, not source of truth.

After the config change, rerun:

bun x skiller@latest apply

Why this works

skiller validates default_agents and [agents.<id>] against its canonical agent catalog before applying anything else.

Once skiller.toml uses claude-code, the config matches the catalog and apply succeeds. The generated .skiller.json does not block the run because it tracks synced items, not the validated agent config surface.

Prevention

• Treat .claude/skiller.toml as the only source of truth for agent ids.
• When skiller rejects an agent name, check the current skiller docs or changelog before touching generated files.
• If bun x skiller@latest apply starts failing during install, rerun it directly first. The error message is specific enough to point at the broken id.