ci: install and self-test the skill for every agent the skills CLI supports

.github/scripts/smoke-install.sh

@@ -1,21 +1,70 @@
 #!/usr/bin/env bash
 # Install this repo's skill with the Vercel `skills` CLI into a throwaway HOME, once per
-# agent named on the command line, and run the skill's own self-check from each installed
-# copy. Usage: .github/scripts/smoke-install.sh claude-code codex cursor gemini-cli ...
+# agent, and run the skill's own self-check from each installed copy. With no arguments,
+# the agent list is read from the CLI itself, so agents it gains are covered automatically.
+# Usage: .github/scripts/smoke-install.sh [agent...]
 set -euo pipefail
 
 repo=$(cd "$(dirname "$0")/../.." && pwd)
-[ $# -gt 0 ] || { echo "usage: $0 <agent> [agent...]" >&2; exit 2; }
 
-# Each agent gets its own throwaway home: all agents except claude-code install into the
-# shared ~/.agents/skills/, so a single home would make the one-installed-copy assert below
-# pass for the first agent and then see the previous agent's copy for the rest.
-for agent in "$@"; do
+# Agents the CLI lists but cannot install globally: "<agent>|<one-line reason>" each. They
+# are reported at the end and excluded from the pass count rather than silently skipped.
+exceptions=(
+  "eve|the CLI reports Eve does not support global skill installation"
+  "promptscript|the CLI reports PromptScript does not support global skill installation"
+)
+
+strip_ansi() { sed -E $'s/\033\\[[0-9;?]*[a-zA-Z]//g'; }
+
+reason_for() {
+  local e
+  for e in ${exceptions[@]+"${exceptions[@]}"}; do
+    [ "${e%%|*}" = "$1" ] && { echo "${e#*|}"; return; }
+  done
+  return 0
+}
+
+# Ask the CLI for its own agent list: an invalid -a makes it print "Valid agents: a, b, ...".
+# Output is joined into one line first, so a wrapped list is parsed the same as an unwrapped one.
+list_agents() {
+  local home out
+  home=$(mktemp -d)
+  out=$(HOME="$home" XDG_CONFIG_HOME="$home/.config" XDG_DATA_HOME="$home/.local/share" npx --yes skills add "$repo" --skill interface-matrix -g \
+    -a __invalid__ -y --copy 2>&1 | strip_ansi | tr '\n' ' ' || true)
+  rm -rf "$home"
+  printf '%s' "${out##*Valid agents:}" \
+    | grep -oE '[A-Za-z0-9._-]+(, *[A-Za-z0-9._-]+)+' \
+    | head -1 | tr ',' '\n' | sed -E 's/^ +//; s/ +$//'
+}
+
+agents=()
+if [ $# -gt 0 ]; then
+  agents=("$@")
+else
+  while IFS= read -r a; do [ -n "$a" ] && agents+=("$a"); done < <(list_agents)
+  [ "${#agents[@]}" -gt 0 ] || { echo "could not parse the skills CLI agent list" >&2; exit 1; }
+fi
+
+echo "== skills add --list"
+home=$(mktemp -d); HOME="$home" XDG_CONFIG_HOME="$home/.config" XDG_DATA_HOME="$home/.local/share" npx --yes skills add "$repo" --list; rm -rf "$home"
+
+total=${#agents[@]}
+passed=0
+skipped=()
+
+# Each agent gets its own throwaway home: most agents install into the shared
+# ~/.agents/skills/, so a single home would make the one-installed-copy assert below pass
+# for the first agent and then see the previous agent's copy for the rest.
+for agent in "${agents[@]}"; do
+  reason=$(reason_for "$agent")
+  if [ -n "$reason" ]; then
+    echo "== $agent: SKIPPED — $reason"
+    skipped+=("$agent: $reason")
+    continue
+  fi
   home=$(mktemp -d)
-  echo "== skills add --list ($agent)"
-  HOME="$home" npx --yes skills add "$repo" --list
   echo "== skills add -a $agent"
-  HOME="$home" npx --yes skills add "$repo" --skill interface-matrix -g -a "$agent" -y --copy
+  HOME="$home" XDG_CONFIG_HOME="$home/.config" XDG_DATA_HOME="$home/.local/share" npx --yes skills add "$repo" --skill interface-matrix -g -a "$agent" -y --copy
 
   found=$(find "$home" -path "*/interface-matrix/SKILL.md" -print)
   count=$(printf '%s' "$found" | grep -c . || true)
@@ -27,4 +76,8 @@ for agent in "$@"; do
   done
   ( cd "$dir" && python3 scripts/test_interface_matrix.py && python3 -O scripts/test_interface_matrix.py )
   rm -rf "$home"
+  passed=$((passed + 1))
 done
+
+for s in ${skipped[@]+"${skipped[@]}"}; do echo "exception: $s"; done
+echo "$passed of $total agents installed and tested"

.github/workflows/tests.yml

@@ -51,8 +51,8 @@ jobs:
       - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: "22"
-      - name: skills CLI install smoke test (six agents)
-        run: .github/scripts/smoke-install.sh claude-code codex cursor gemini-cli github-copilot opencode
+      - name: skills CLI install smoke test (all agents)
+        run: .github/scripts/smoke-install.sh
 
   commit-emails:
     runs-on: ubuntu-latest

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## Unreleased
 
+- CI installs and self-tests the skill for every agent the `skills` CLI supports, with the
+  agent list read from the CLI at run time.
 - Copyright holder in `LICENSE` and the README is now `macblackstuff`.
 - Code of conduct (Contributor Covenant 2.1); reports go to `conduct@macblackstuff.com`.
 - README rewritten: what it does, who it is for, a worked example, the

README.md

@@ -125,12 +125,12 @@ python3 scripts/test_interface_matrix.py    # Ran 83 tests ... OK
 
 ### Harnesses tested
 
-CI installs the skill with the `skills` CLI on every push and pull request, once per agent with `--copy`, and
-runs the 83 tests from each installed copy. Six agents are covered: `claude-code`, `codex`,
-`cursor`, `gemini-cli`, `github-copilot` and `opencode`. Five of the six share one user-level skills
-directory (the `skills` CLI decides the target; see its documentation for each agent's path), so
-on a real machine a single installed copy can serve all five. CI still installs and tests each
-agent separately, in its own throwaway home, so a change to any one agent's target is caught.
+CI installs the skill with the [`skills` CLI`](https://github.com/vercel-labs/skills) on every push
+and pull request, once per agent in its own throwaway home, and runs the 83 tests from each
+installed copy. Every agent the CLI supports is covered — 79 at the time of writing (`skills`
+1.7.0), of which 77 are installed and tested; the list is read from the CLI at run time. Two
+agents are excluded with reasons recorded in `.github/scripts/smoke-install.sh`: `eve` and
+`promptscript` — the CLI reports that neither supports global skill installation.
 
 The skill itself is harness-neutral: it is a `SKILL.md` plus standard-library Python, with no
 agent-specific commands.