Treat $ARGUMENTS as optional scope for the run being closed. Preserve all manual edits, show the operator the exact affected files/issues/docs before consequential actions, and stop on any failed gate or contradictory completion evidence.
Treat $ARGUMENTS as optional scope for this run. Prepare and report a safe execution plan; do not fire any marathon until the operator explicitly confirms the exact dry-run-approved plan and order.
Runs before the agent uses a tool for Bash, Skill, Bash and Bash tool calls, executing relay-xyz-guard.sh, gh177-sandbox-test-guard.sh and gh527-destructive-git-guard.sh via bash (3 commands). From HiQS-Labs/XYZ-forge.
AGENTS.md instructions for HiQS-Labs/XYZ-forge, covering agents.md, danger: commands agents must not run, runtime default, what this file owns and operating principles.
Sweep GitHub issues opened or updated in an age window (default: last 10 days; the window is adjustable, e.g. an 11-14-day slice to avoid overlapping another sweep), verify each one is still valid, reproducible, and not already fixed (fan out subagents to check issue state, comments, and git/commit history for…
Start or join a local XYZ discussion shared by two or more Claude, Codex, or other agent sessions through a compact six-digit ID. Use when a prompt says “Join XYZ AgentChorus #123456 as agent number two… — use the agent-chorus skill” (older invitations omit the trailing clause, and the legacy phrase “Join XYZ…
ATE (Automated Testing Environment): drive long-running (2-3hr+) unattended variation-test suites against the Aider, OpenRouter, and GLM 5.2 pipeline using a local Gemma model in LM Studio as the worker, with periodic frontier-model check-ins for drift control, ending in a single triaged GitHub issue. Use this…
Pressure-test an option set before the operator picks. Catches an AI agent — this one included — conflating two separate problems into one menu, or quietly narrowing the field to a false dilemma that skips the boring option. Borrows debug-mantra's falsify-the-hypothesis rigor (disprove each option before trusting it)…
Diagnose CI health and, when needed, benchmark a config change side by side — without re-deriving either from scratch each time. Use when the operator asks "why is CI slow", "diagnose our CI", "is our CI healthy", "what's the bottleneck", or wants to compare wall-clock across two or more runs-on (or other config)…
One-shot cross-model CONSULT — fan the same question out to Codex and agy in parallel (repo-isolated, advisory), then reconcile their answers into one. Use when the user wants a "second opinion", to "ask Codex and agy", a "panel" or "cross-model" check, or an independent gut-check on a decision/design/doc before…
Read two or more Convergence Maps produced by /dry — across repos, or across re-runs of one repo — and find what repeats. Clusters findings by SHAPE rather than by resource, so "a gateway exists and the files that import it bypass it anyway" surfaces as one portable finding rather than four unrelated ones, and tracks…
Four-mantra debugging discipline — reproduce, trace the fail path, falsify the hypothesis, cross-reference every breadcrumb. Recite the mantra block verbatim at the start of any debugging session, then apply the four steps in order before proposing any fix. Trigger on /debug-mantra and proactively whenever debugging…
Find the subsystems that should be one subsystem. Traces a codebase resource-first instead of subject-first, then reports every place N independent paths reach the same table, file, env var, endpoint, binary, or config key without ever meeting — or where N places define the same decision with different answers — the…
Hotfix fast lane (GH-267) for an explicit /express, "express this hotfix", or "express GH-N" request. Carries a critical, risk-bounded fix, registered regression suite, releases-ledger updates, born-complete PDDA doc, CHANGELOG, landing, and reconciliation in one operator-authorized motion. Requires a fresh full task…
Convert a highly technical document or subject into a plain, accurate, layman-friendly explanation using Feynman-style teaching moves — everyday analogy, a first-principles ladder, jargon stripping, and honesty about the limits. Use this whenever the user wants to "explain this simply", "make this understandable to a…
File a bug against the xyz-3-agents-swarm harness from ANY repo or session — lands a PDDA-compliant capture in that repo's PROJECT/1-INBOX/ (GH- -.md), files the tracking GitHub issue, and parks a one-line ROADMAP pointer, without touching the repo you are standing in. Use when the xyz harness misbehaves while you are…
Walk a repo's front door as a brand-new user would and report whether they can actually get from clone to working — not just whether the docs exist. Use whenever the user asks to audit, review, or critique a repo's onboarding, installation, setup, getting-started, or quickstart experience; asks "can a new user install…
Diagnose and fix the "git push works but gh fails" auth split on macOS — where the GitHub CLI and git consult different credential stores and drift apart — non-destructively, without over-engineering. Trigger when the user says "git push works but gh fails", "gh auth is broken", "token in keyring is invalid", "gh auth…
Produce an honest, defensible ground-truth read of the current repo — how mature the codebase really is, how far along each feature actually is, and what can be safely claimed about it externally. Reads CHANGELOG, architecture/technical docs, and recent git history; never modifies the repo. Offers an Express (10–15…
HQ — the multi-repo command center. Turn a single utterance ("For project Acme, do X") into governance-aware action across every repo on this device. Resolves a project name (fuzzily) to a real repo via the registry ladder (Rebalance projectregistry → XYZ install registry → Git Pulse PDDA registry → filesystem)…