Namespaced skill plugin for Claude Code and OpenAI Codex CLI agents. All skills load under the skein: namespace (skein:dev-plan, skein:plan-view, etc.) to avoid collisions with third-party skills.
| Skill | Claude | Codex | Description |
|---|---|---|---|
| dev-plan | Yes | Yes | Generate and manage development plans; create runs one fresh-context Explore subagent that gathers verified paths, patterns, dependency versions, and git refs before drafting, and drafts a conditional ## Architecture & Call Flow section (Mermaid diagrams + context-lifecycle table, confirmed at a gate) for plans with 2+ independently-executing components |
| fan-out | Yes | Yes | Parallel agent orchestration via worktrees |
| content-draft | Yes | Yes | Draft content following style guidelines |
| content-review | Yes | Yes | Review content against style guidelines |
| deep-review | Yes | Yes | Multi-lens code review with fresh-context subagents; reconciles findings by structural signature. Opt-in --auto-fix=trivial applies a hard-coded allowlist of mechanical fixes (requires --test-cmd) |
| review-plan | Yes | Yes | Audit a dev plan via five parallel lenses (architecture, sequencing, spec-and-testing, assumptions, codebase-claims; four high-reasoning + one factual) plus one additional pass after reconciliation that detects plan-internal and cross-lens contradictions, before implementation; the architecture lens also audits negative-space topology gaps; reconciles findings by structural signature. A default-on triage-and-clarify loop records decisions into the plan before the marker is written (--batch skips it for CI); grill-eligible findings (architecture/component-boundary, third-party integration, security, rate-limiting, or any Contradiction finding) are delegated to skein:grill's interview protocol instead of the standard option list. Opt-in --auto-fix=trivial applies allowlisted prose edits outside the immutable contract sections |
| grill | Yes | Yes | Relentless, one-question-at-a-time interview over a plan file or freeform idea; splits facts (looked up, never asked) from decisions (one recommendation each, blocks until accept/override/waive); standalone and user-invocable, also used inline by review-plan Step 6.4 for grill-eligible findings |
| rfc-finder | Yes | Yes | Find and link to IETF RFCs and related drafts |
| spec-compliance | Yes | Yes | Check code against RFC/W3C/WHATWG requirements |
| update-docs | Yes | Yes | Audit and update stale docs against branch diffs |
| conduct | Yes | Yes | Walk a reviewed dev plan phase by phase via harness-native clean-context subagents |
| plan-view | Yes | Yes | Generate HTML dashboard and per-plan drill-down pages from a markdown dev-plan corpus; renders Mermaid fences as live diagrams via the Mermaid CDN runtime; --rich mode produces LLM-rendered per-plan views constrained by a widget toolkit; deterministic and rich pages are cross-linked bidirectionally (forward links emitted unconditionally, back-links injected by relink_rich_pages() on every plain run) |
| review-gauntlet | Yes | Yes (partial) | Chain the review gates (adversarial review, deep-review, security-review) into one convergence loop, with fixes applied by an isolated clean-context fixer subagent; opt-in per dev plan via **Review Gates:** none | full on Claude (none | quick | full on Codex — see below), auto-chained from conduct and fan-out. --resume/--fresh pick an interrupted run back up from a target-keyed .gauntlet/ ledger instead of restarting the gate corpus; auto-chained invocations are self-resuming with no extra flags. /code-review is not a gate on the Claude side (the harness blocks Claude from self-invoking it) — run /code-review xhigh --fix yourself; the Claude-side marker has no quick value as a result. On Codex, /code-review remains a native gate (unaffected) and quick still runs it alone; security-review is deferred (no primitive exists), and deep-review is gated pending nested-spawn confirmation — Codex full mode reports these slots explicitly rather than claiming they ran |
| release | Yes (user-invoked only) | Yes (model-invocable; no documented opt-out found; explicit pre-mutation confirmation) | Cut or re-sync a GitHub release from a CHANGELOG.md section: title <repo> vX.Y.Z — <highlight>, body = an optional ## What's New summary paragraph + the section content with its header stripped + a **Full diff:** compare link; an ordinary re-sync preserves the existing highlight and ## What's New presence/content when recoverable, making repeated re-syncs byte-identical unless the user overrides them or malformed existing content requires repair. /release audit scans tags, GitHub releases, and CHANGELOG versions for missing artifacts, releases whose remote tag was deleted, or drifted title/body, then reports a punch list before fixing anything |
Invoke each skill as skein:<name> (e.g. skein:dev-plan, skein:review-plan). Judgment lenses ("high-reasoning" above) vs. mechanical/factual work follow the two-tier model/effort policy in AGENTS.md.
Install through the harness plugin CLI on each machine. There are no rsync, promote, or bootstrap scripts — the marketplace files in this repo are the install surface. Both harnesses can install directly from GitHub; no clone required.
Claude Code:
/plugin marketplace add vr000m/skein
/plugin install skein@skeinCodex CLI:
codex plugin marketplace add vr000m/skein --ref main
codex plugin add skein@skeinThe marketplace name skein matches the entries in .claude-plugin/marketplace.json (Claude) and .agents/plugins/marketplace.json (Codex). To pick up upstream changes, re-add the marketplace (or git pull your clone if you installed from a local path) and re-run /plugin install skein@skein / codex plugin add skein@skein — there is no separate sync step.
Pinning to an existing release. Tagged releases already exist through v0.5.3. Prefer pinning over tracking main: for example, pass --ref v0.5.3 to codex plugin marketplace add, and check tagged releases on the GitHub releases page for Claude Code (which resolves the marketplace at add time). Tracking main follows unreleased commits.
If you are developing against a local clone instead, swap the marketplace source for a path: /plugin marketplace add /path/to/skein (Claude) or codex plugin marketplace add /path/to/skein (Codex).
If you are migrating from the older flat layout (skills installed directly under ~/.claude/skills/ and ~/.codex/skills/), use scripts/delete-skills.sh to remove the immutable 11-skill migration-era set after verifying the plugin install loads correctly. It deliberately preserves the post-migration grill, release, and review-gauntlet directories. The cleanup regression test is uvx pytest tests/parity/test_delete_skills.py -q.
skein follows Semantic Versioning; changes are recorded in CHANGELOG.md in Keep a Changelog format.
Tagged releases are published on the GitHub releases page; the latest is v0.5.3. Pin an install to a specific tag with --ref vX.Y.Z instead of --ref main if you want a stable version rather than tracking main.
After merging a release-bearing PR and updating local main, use skein:release (/release X.Y.Z on Claude) to create or re-sync the tag and GitHub release. The skill previews the exact target, title, and body for confirmation before mutating remote state; do not hand-compose the old git tag + gh release create sequence.
- Install tools:
brew install just jq shellcheck shfmt- Repo invariants worth running before pushing:
just check-sync # canonical scripts/ ↔ bundled skill scripts byte-identity
just check-prompt-parity # Claude vs Codex SKILL.md content parity for bundle skills
just check-trunk-snippet-parity # trunk-resolution snippet parity
just parity-tests # bundle + allowlist + orchestration-contract + no-fallback + marker + managed-skill/cleanup-boundary regression coverage
just reconciliation-tests # reconciliation parity + fixture + renderer + determinism
just bundle-appliers # regenerate bundled auto-fix pipeline inside each skillplugins/skein/skills/<name>/ Claude skills (SKILL.md per skill)
plugins/skein-codex/skills/<name>/ Codex skills (mirrored structure; per-harness dispatch idiom)
.claude-plugin/marketplace.json Claude marketplace entry for skein
.agents/plugins/marketplace.json Codex marketplace entry for skein
scripts/ Canonical shell scripts (bundle/check/reconcile/parity/render/auto-fix)
See AGENTS.md for commands, architecture, authority model, contributor workflow, and the model/effort policy.