Carry cead's own scoping skill, slice, for Codex and Claude Code

[1zeroone0]

The repo carries its own skill, so a clone anywhere (a remote box, a phone session's sandbox) scopes cead's work the way it is done here, for Codex and Claude Code alike.

What this PR does

• .agents/skills/slice/SKILL.md: slice, cead's own scoping skill, named for CODE.md's Method step 2 ("a PR takes a slice of the spec"). It settles a PR's slice (the spec's actions and properties) and its description before any spec, Lean or Rust. Each round covers what applies: whether the spec changes (then the first commit), which functions a skeptic must trust (Lean), the shape before the skeleton, nouns into CODE.md's table with their type homes and terms from OSTEP or ocap, the RLM thesis, dependencies one crate at a time, structure only when required. It opens with the principle from Run one job end to end as a recursive language model #13: slicing is a conversation, not a step toward automation.
• .claude/skills → ../.agents/skills: one link for every repo skill, the pattern of CLAUDE.md → AGENTS.md. .agents/skills is where Codex reads repo skills; .claude/skills is Claude Code's.
• .gitignore: .claude/* stays local (worktrees, handoffs, settings) except .claude/skills.
• AGENTS.md says where skills live.

What this PR does not do

• Touch the general scope skill in ~/dev/skills: it stays Roone's for scoping any work; slice is cead's.
• Add other skills.

Merge requirements

• In a fresh clone from GitHub, Codex lists slice (codex exec, read-only) and Claude Code lists slice (claude -p), both through the symlinked directory; worktrees and handoffs stay ignored.
• No name collision: Claude Code lists slice from the repo and scope from Roone's user skills.
• .gitignore conflicts with Run one job end to end as a recursive language model #13's (spec/.lake/); resolved in Run one job end to end as a recursive language model #13 by merging main after this lands, not here.

🤖 Generated with Claude Code

[1zeroone0]

974f5ea: Carry the scope skill in the repo, readable by Codex and Claude Code alike

1. Built: the skill at .agents/skills/scope/SKILL.md; the symlink .claude/skills → ../.agents/skills; .gitignore re-including only .claude/skills; one line in AGENTS.md.
2. Why: a clone anywhere should scope the way this machine does. .agents/skills is Codex's repo location (per the Codex docs) and .claude/skills is Claude Code's; linking the directory mirrors CLAUDE.md → AGENTS.md.
   • Rejected: a link per skill, which is more links to keep; and keeping the skill only in ~/dev, which a clone never sees.
3. Bloat: none.
4. Drift: the skill text is copied from ~/dev/skills/scope. Until Roone decides, two copies exist, and the repo's has the new paragraph.
5. Trust surface: untested. Whether each tool follows a symlinked skills directory is a merge requirement to check in a fresh clone.
6. How it breaks:
   • A tool that doesn't follow directory symlinks sees no skills.
   • In this repo Claude Code may list both the personal and the repo scope.
7. Not confident: which copy Claude Code prefers when both exist.
8. Verify yourself: the .gitignore negation.
9. Next steps: finished once verified in a fresh clone.

[1zeroone0]

23ffdbc: Make the repo's skill cead's own: slice, the method's first step

1. Built: scope → slice, rewritten for cead: target and reads (spec files included); the slice named in the destination; a checklist per round (spec, Lean scalpel, shape, nouns and types, thesis, dependencies, structure); write-back with a Spec: line.
2. Why: Roone keeps the general scope for any work. A distinct name means no collision, and CODE.md's own word for a PR's part of the spec was already "slice".
3. Bloat: one skill, not two copies of scope.
4. Drift: the handoff for Run one job end to end as a recursive language model #13 now says /slice 13.
5. Trust surface: a fresh clone from GitHub was verified with both tools (see the description).
6. How it breaks: a tool that stops following directory symlinks. Checked for both today.
7. Not confident: whether the round checklist is too long for small PRs. "What applies" is meant to let it shrink.
8. Verify yourself: the Rounds checklist.
9. Next steps: finished; ready for review.