XYZ Forge is a local-first engineering operations system. It coordinates several AI coding agents — Claude Code, Codex, agy (Google's Antigravity CLI), and others — working on the same repositories, and it carries the surrounding machinery an unattended agent run actually needs: specified work, collision-free lanes, verification gates, and receipts
This repository also configures its own agents. See what XYZ-forge tells them →
Skill Claude CodeCodex
Add a curated set of read-only permission rules -- file reads, directory listings, content search (grep/glob), git inspection, and system checks -- to a Claude Code settings.json allowlist so safe reads stop triggering permission prompts. Local read-only by default; outbound reads (WebFetch, WebSearch, registry…
Skill Claude CodeCodex
Audit a repo's root README as both the artifact under review and the map to everything else — checking accuracy, legibility, understandability, and operator UX, then following every pointer it makes to other docs as a litmus test for the whole repo's doc hygiene. Use when the user asks to "review/audit/check the…
Skill Claude CodeCodex
Trace an existing system end to end — entry points, call paths, every read and write of the state involved, the contracts crossed, the failure and rollback paths — and write it down as a Recon Map before a plan for changing it is drafted. Fires when the user asks to plan, spec, design, refactor, migrate, or size a…
Skill Claude CodeCodex
Tick-backed, self-healing automation for the file-based /relay review loop — a hands-free poll driver (poll.sh), a relay-turn supervisor (relay-drive.sh), a liveness watchdog (watchdog.sh), a verdict-gated runner (runner.sh), and headless cross-model turn-takers (codex-turn.sh, agy-turn.sh). Sibling to the xyz/tick…
Skill Claude CodeCodex
Turn a finished /relay producer↔reviewer thread into a single checklist-style GitHub issue, filed in the repo the relay was actually about. Reviews the relay transcript, summarizes the conclusion (including agreements AND unresolved disagreements), and synthesizes the actionable findings into a markdown checklist…
Skill Claude CodeCodex
Drive an automated /relay review loop on THIS repo with the shipped relay-automation harness (relay-drive.sh + codex-turn.sh / agy-turn.sh / poll.sh) rather than improvising the handoff by hand. Use when the operator wants to "run an automated relay", "have Codex or agy review this end-to-end", "drive a relay to…
Skill Claude CodeCodex
Generate and run a relay thread file in relay-system/ / .md — a turn-based, file-based review loop between two Claude Code agents (a Producer who builds, a Reviewer who critiques and proposes fixes the author applies) so a human stops copy-pasting output between two windows. Use this whenever the user wants to "set up…
Skill Claude CodeCodex
Read, synthesize, diagnose, clean, author, and publish the optional RELEASES.md planning ledger through one routed workflow. Use for /releases; release status or health checks; stale-plan review against merged PRs, commits, and CHANGELOG entries; disciplined release creation or updates; ledger cleanup; historical…
Skill Claude CodeCodex
Deterministic, multi-model code review system and skill for XYZ. Summon frontier reviewers (Command Code -> Qwen 3.8-Max, GLM-5.3, OpenRouter, Codex, Agy) to perform throwaway-worktree isolated code reviews on local diffs or GitHub PRs. Validates [Blocker]/[Should]/[Nit]/[Pass] citations and automatically formats or…
Skill Claude CodeCodex
Review the most recent permission requests in the current session, generalize them into safe allowlist rules, and append them to .claude/settings.local.json to reduce future prompts. Triggers when the user runs "/rpr", "/rpr ", "/rpr session", or says things like "stop asking me", "I keep getting permission prompts"…
Skill Claude CodeCodex
Audit a script-calling skill for CWD-sensitive path-resolution bugs — bundled .sh scripts that run in the session that wrote them but come back "not found" in another session, repo, or install — then write a graded report with a proposed fix. Runs a read-only static path audit plus a live scenario matrix (foreign CWD…
Skill Claude CodeCodex
Interrogate authority before planning a feature, refactor, migration, rewrite, backfill, event-sourcing design, dual-write setup, new authoritative store, or system-of-record change. Fire whenever a proposal might introduce, move, duplicate, or replace authoritative state — or involves event sourcing or "event…
Skill Claude CodeCodex
Session-scoped triage: what did I leave open, what is rotting, and is the plan still right — answered in under a screen. Runs in seconds over the session, the working tree, open PRs, and the two ledgers (ROADMAP + RELEASES DB), then emits a ranked list capped at 7 items plus a capped strategic read. Use when the…
Skill Claude CodeCodex
Draw an interactive system architecture or Git-history diagram for a repo. Architecture maps start from existing rendered knowledge (graphify, codebase-memory, RAG, architecture docs) before raw code. Git-history maps render commits, branch cuts, and merges as stacked lanes from local refs. Produces a JSON spec and…
Skill Claude CodeCodex
Apply software-engineering standards when authoring or reviewing project plans, build/spec/PRD documents, architecture RFCs, or agent governance. Use for "write a plan", "review this build doc", "apply our SWE standards", or "is this plan ready to build"; also apply before drafting a project plan. Grades grounded…
Skill Claude Code
Apply three lenses to one problem in a fixed order — recon (what is actually there), debug-mantra (is what I believe actually true), ponytail (what is the smallest thing that works) — with an explicit rule for how much evidence to buy before shrinking. Use when a task tempts you toward opposite errors at once…
Skill Claude CodeCodex
Install the full swarm stack into a target repo: vendor the XYZ harness (relay/marathon/consult/self-improve + tick) into /.xyz/, then optionally install the PDDA doc-governance runtime (PROJECT/PDDA.md + utils/pdda/.sh). Both installers self-locate, are idempotent, and register the target into their own per-user…
Skill Claude CodeCodex
Give a git repo a stable, distinct VS Code background tint so windows are tellable apart at a glance. The color is scoped to the workspace (.vscode/settings.json under workbench.colorCustomizations), so it applies every session with no per-session step. Two modes: deterministic (SHA-256 of the repo's GitHub slug -> a…
Skill Claude CodeCodex
Summarize what shipped to main over the last week as a concise, human- sounding end-user report. Use when the user wants "this week's shipped improvements", "top things that landed", or a weekly release-style recap from git history, especially when they want the answer framed around everyday user impact instead of…
Skill Claude CodeCodex
Coordinate two (or more) AI coding agents working CONCURRENTLY on non-overlapping, path-scoped lanes of ONE shared repo, via the tick CLI — for parallel builds and parallel codebase recon, with collision-free claims, liveness heartbeats, and an honest concurrency metric. Use when the user wants to "run two agents in…
Skill Claude CodeCodex
Give an AI agent a real cloud browser (Browserbase) for web research, scraping, page fetching, web search, form-driving, and site monitoring, with the API key loaded by path from /secrets/browserbase.txt and never written into the repo. Routes each task to the lightest capability (Fetch -> Search -> browse.sh skill…
Skill Claude CodeCodex
Consolidate multiple Git worktrees and clones of a repository into a single clean checkout, safely merge associated PRs in topological dependency order, reconcile post-merge governance state (PDDA and RELEASES DB), and safely remove leftover worktrees/clones without data loss or interrupting active sessions. Strictly…
Skill Claude CodeCodex
Start one task or a batch of GitHub issues and carry it through fresh full clones, repo-governed intake, grounded recon, surgical DRY planning, Codex relay plan QA, execution, verification, final relay QA, and ready PRs. Use for /start-task, "start these issues", or requests to clone, register, execute, QA, and make a…
Skill Claude CodeCodex
End-to-end disciplined problem resolution ladder: establishes ground truth via /debug-mantra, engineers the leanest safe solution via /ponytail (reusing and extending existing subsystems with zero code sprawl), enforces AGENTS.md, SOP.md, and GUIDING PRINCIPLES governance, checks cross-repo CHANGELOG cohesion…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: