Menu
Lazy MCP discovery via mcporter. Hard-trigger before any curl, gh, or HTTP fallback.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Kai's "start a sidequest" command - file a Forgejo issue from a described piece of work, then dispatch it through ward into its own session.
Tier and prune an issue backlog - P0-P4 tier definitions (P4 is the icebox/lowest tier), a target distribution as ranges, percentile-cut assignment that lands within the bands, pruning, and a second orthogonal automation-mode axis (headless/interactive/consult). Triggers - prioritize, prioritization, triage the backlog, P0/P1/P2/P3/P4, backlog ratio, icebox, burn down the backlog, tier the issues, automation mode, eligibility to dispatch.
gpg-ssm is a GPG wrapper that pulls signing material from AWS SSM at sign time instead of caching on disk. Use when wiring up a fresh machine for signed commits, bootstrapping the shared signing key, rotating a key, or debugging signing failures. Triggers - gpg-ssm, gpg signing, signed commits.
Git + GitHub umbrella. Hard rule - never use gh api graphql without double-confirming, REST is default. Owns the full PR lifecycle (branch, commit, open, monitor CI, auto-fix, merge) and routes to ward-passthrough and git-workflow siblings.
git workflow, forgejo, default tracker, commit, push, PR, gh, fj, issue, todo, gish.
Kai's technical preferences - Go CLI defaults (urfave/cli), Charm TUI stack, dead-repo check, doc conventions, default license. Triggers - cli, tui, library, dependency, recommend tool.
| name | tooling-mcp-servers |
| description | Lazy MCP discovery via mcporter. Hard-trigger before any curl, gh, or HTTP fallback. |
mcp, mcporter.
The lazy-loaded MCP layer. Configured servers live in <personal-os-repo>/config/mcporter.json (symlinked from the workspace root so mcporter finds them via its default ./config/mcporter.json lookup). Typed headers per server live in <personal-os-repo>/mcp-servers/*.d.ts.
The point of this layout: discovery is cheap (this skill + the per-server index) and schema is paid only for the server actually needed. This is the lazy path - the only path on Codex, and the fallback anywhere. Note the harness split: on Claude Code the same merged inventory is also registered natively into user-scope (mcp__<name>__* tools) by sync-claude-mcp.py, since Claude's 1M window + on-demand schema deferral make native registration near-free. Both paths reach the same servers. See Design notes ("Eager vs lazy is now a per-harness split").
If the agent is about to reach for anything that smells like an MCP-shaped capability (search a corpus, drive a browser, query an observability backend, hit a service over HTTP that has an MCP wrapper, etc.), this skill fires first. Falling back to curl, gh, raw HTTP, or "I'll just shell out" without checking the mcporter inventory is the bug. The inventory (or mcporter list when the inventory drifts) enumerates everything flat - read it before improvising.
mcporter list discovery commands and .d.ts regeneration..d.ts / mcporter call / output flags, plus the add-a-server checklist..mcp.json, and mobile / cloud MCP separation.