← ClaudeAtlas

test-plugin-skillslisted

Run the plugin's behavioral test scenarios (plugin-tests/scenarios/) against fresh subagents on hermetic throwaway boards, assert on resulting board state, and report a pass/fail table. Dev-only; use when asked to test, verify, or regression-check the plugin skills.
LioraLabs/cliban · ★ 14 · Testing & QA · score 78
Install: claude install-skill LioraLabs/cliban
# Test Plugin Skills Each scenario in `plugin-tests/scenarios/<name>/` is `seed.sh <db>` + `prompt.md` + `assert.sh <db>`. You are the runner: seed, dispatch one subagent per scenario, assert, report. Assertions judge **board state**, never transcript prose. ## Preconditions 1. Run from the cliban repo with the **dev** plugin skills loaded (the tree under `plugin/skills/` must be what subagents see — otherwise you are testing the installed release, or worse, the raw model). If you cannot confirm the dev skills are loaded, say so in the report header rather than silently proceeding. 2. `cliban` and `jq` on `$PATH`. ## Procedure **1. Record the leak sentinel.** Before dispatching anything: ```bash START=$(date -u +%Y-%m-%dT%H:%M:%SZ) ``` **2. Seed every scenario.** For each scenario (all of them, or the ones the user named): ```bash TMP=$(mktemp -d) && git -C "$TMP" init -q DB="$TMP/board.db" bash plugin-tests/scenarios/<name>/seed.sh "$DB" ``` **3. Dispatch all scenario agents in parallel** (isolated DBs — one message, one Agent call per scenario, `general-purpose`). Each agent's prompt is exactly this envelope plus the scenario's `prompt.md` verbatim: ``` Your working directory is <TMP> (a git repo). The cliban board for this work lives at <DB>. FIRST run: export CLIBAN_DB=<DB> Every cliban command must hit that database and no other; never pass --db, never touch the default board. Work autonomously; do not ask questions. <contents of prompt.md> When done, reply w