Skillshare 技能验证

Run isolated E2E tests in devcontainer. $ARGUMENTS specifies runbook name or "new".

Flow

Phase 0: Environment Check

1. Confirm devcontainer is running and get container ID:

   CONTAINER=$(docker compose -f .devcontainer/docker-compose.yml ps -q skillshare-devcontainer)
   

   • If empty → prompt user: docker compose -f .devcontainer/docker-compose.yml up -d
   • Ensure CONTAINER is set for all subsequent docker exec calls.

2. Confirm Linux binary is available:

   docker exec $CONTAINER bash -c \
     '/workspace/.devcontainer/ensure-skillshare-linux-binary.sh && ss version'
   

3. Confirm mdproof is installed:

   docker exec $CONTAINER /workspace/.devcontainer/ensure-mdproof.sh
   

   This auto-installs from GitHub release, or falls back to /workspace/bin/mdproof (local dev binary).

4. Check for lessons learned from previous runs:

   test -f /workspace/.mdproof/lessons-learned.md && cat /workspace/.mdproof/lessons-learned.md
   

   If the file exists, read it before writing or debugging runbooks — it contains known gotchas and assertion patterns.

Phase 1: Detect Scope

1. Preview all available runbooks via the container:

   docker exec $CONTAINER mdproof --dry-run --report json /workspace/ai_docs/tests/
   

   This returns JSON with every runbook's steps, commands, and expected assertions — no manual markdown parsing needed. Use this to understand what each runbook covers.

2. Identify recent changes (unstaged + recent commits):

   git diff --name-only HEAD~3
   

3. Match changes to relevant runbooks (compare changed file paths against step commands in the JSON output).

Phase 2: Select Tests

Prompt user (via AskUserQuestion):

• Option A: Run existing runbook (list all available + mark those related to recent changes)
• Option B: Auto-generate new test script based on recent changes
• Option C: If $ARGUMENTS specifies a runbook, skip to Phase 3

Phase 3: Prepare & Execute

Running existing runbook:

1. Create isolated environment with auto-initialization:

   ENV_NAME="e2e-$(date +%Y%m%d-%H%M%S)"
   
   # Use --init to automatically run 'ss init -g' with all targets
   docker exec $CONTAINER ssenv create "$ENV_NAME" --init
   

2. Execute the entire runbook via mdproof inside the container:

   docker exec $CONTAINER env SKILLSHARE_DEV_ALLOW_WORKSPACE_PROJECT=1 \
     ssenv enter "$ENV_NAME" -- \
     mdproof --report json \
     /workspace/ai_docs/tests/<runbook_file>.md
   

   mdproof executes each step (bash -c <command>) in the ssenv-isolated HOME, then returns structured JSON:

   {
     "version": "1",
     "runbook": "<runbook_file>.md",
     "duration_ms": 12345,
     "summary": { "total": 7, "passed": 5, "failed": 1, "skipped": 1 },
     "steps": [
       {
         "step": { "number": 1, "title": "...", "command": "...", "expected": ["..."] },
         "status": "passed",    // "passed" | "failed" | "skipped"
         "exit_code": 0,
         "stdout": "...",
         "stderr": "..."
       }
     ]
   }
   

3. Analyze the JSON output:

   • All passed → proceed to Phase 4
   • Any failed → filter for failures only (full JSON can be too large for terminal output):

     mdproof --report json runbook.md 2>&1 | jq '{
       summary: .summary,
       failed: [.steps[] | select(.status == "failed") | {
         step: .step.number, title: .step.title,
         exit_code: .exit_code,
         failed_assertions: [.assertions[]? | select(.matched == false) | .pattern],
         stderr: (.stderr // "" | .[0:200])
       }]
     }'
     

   • Skipped steps (executor=manual) → these need manual verification, run them individually:

     docker exec $CONTAINER env SKILLSHARE_DEV_ALLOW_WORKSPACE_PROJECT=1 \
       ssenv enter "$ENV_NAME" -- <command from step.command>
     

4. For failed steps, debug individually using manual docker exec (same as before):

   docker exec $CONTAINER env SKILLSHARE_DEV_ALLOW_WORKSPACE_PROJECT=1 \
     ssenv enter "$ENV_NAME" -- bash -c '<failed step command>'
   

   • Prefer --json + jq for assertions — see the JSON Reference below

Generating new runbook:

1. Read git diff HEAD~3 to find changed files in cmd/skillshare/ or internal/
2. Read changed files to understand new/modified functionality
3. Validate all CLI flags before writing — for every ss <command> <flag> in the runbook:
   • Grep cmd/skillshare/<command>.go for the exact flag string (e.g. "--force")
   • Run ss <command> --help inside container if needed
   • Common mistakes to avoid:
     • uninstall --yes → wrong, use --force / -f
     • init --target <name> → wrong, init has no --target flag
     • init -p has a completely separate flag set from global init — only supports --targets, --discover, --select, --mode, --dry-run. Global-only flags like --no-copy, --no-skill, --no-git, --all-targets, --force do NOT exist in project mode
     • Audit custom rules: disable by rule ID (e.g. prompt-injection-0, prompt-injection-1), NOT pattern name (e.g. prompt-injection). Rule IDs are in internal/audit/rules.yaml
4. Generate new runbook to ai_docs/tests/<slug>_runbook.md, following existing conventions:
   • YAML-free, pure Markdown
   • Has Scope, Environment, Steps (each with bash + Expected), Pass Criteria
   • Use jq: assertions in Expected blocks for JSON commands — e.g. - jq: .extras | length == 1. This is a native mdproof assertion type, NOT a bash jq pipe
   • Use --json + jq -e in bash for inline verification within multi-command steps
   • Config idempotency — never bare cat >> config.yaml; always prepend sed -i '/^section:/,$d' to remove existing section first, or use CLI commands (ss extras init, ss extras remove --force) that handle duplicates
   • Check ai_docs/tests/runbook.json for project-level config (build, setup, teardown, step_setup, timeout) that affects all runbooks
   • Check .mdproof/lessons-learned.md for known assertion patterns and gotchas
5. Run the runbook quality checklist (see below) before executing
6. Then execute the new runbook (same flow as above)

Phase 4: Cleanup & Report

1. Ask user before cleanup (via AskUserQuestion):

   • Option A: Delete ssenv environment now
   • Option B: Keep for manual debugging (print env name for later ssenv delete)

2. If user chose Option A:

   docker exec $CONTAINER ssenv delete "$ENV_NAME" --force
   

3. Output summary (derived from the runbook JSON output):

   ── E2E Test Report ──
   
   Runbook:  {runbook name}
   Env:      {ENV_NAME}
   Duration: {duration_ms}ms
   
   Step 1: {title}  PASS
   Step 2: {title}  PASS
   Step 3: {title}  FAIL ← exit_code={N}, stderr: {error detail}
   ...
   
   Result: {passed}/{total} passed ({skipped} skipped)
   

   All values come directly from mdproof's JSON output — summary.passed, summary.total, steps[].step.title, steps[].status.

4. If any FAIL → distinguish between runbook bug vs real bug:

   • Runbook bug: wrong flag, wrong file path, stale assertion → fix runbook, re-run step
   • Real bug: CLI misbehavior → analyze cause, provide fix suggestions

5. Retrospective — ask user (via AskUserQuestion):

   Did you encounter any friction during this test run that the skill or runbook could handle better?

   • Option A: Yes, improve e2e skill — review test friction (wrong flags, stale assertions, missing checklist items, unclear instructions), then update SKILL.md and/or runbooks
   • Option B: Yes, but only fix the runbook — fix the specific runbook without changing the skill itself
   • Option C: No, skip

   Improvement targets:

   • SKILL.md: add new checklist items, common-mistake examples, or rule clarifications learned from this run
   • Runbooks: fix stale assertions (e.g. config.yaml → registry.yaml), wrong flags, outdated paths
   • Both: when a systemic issue (e.g. a refactor changed file locations) affects both the skill's guidance and existing runbooks

Runbook Quality Checklist

Before executing a newly generated runbook, verify:

• All CLI flags exist — every ss <cmd> --flag was grep-verified against source
• --init interaction — if runbook has ss init, account for ssenv create --init already initializing (add --force to re-init, or skip init step)
• --init creates default extras — ssenv create --init creates a rules extra by default. Runbooks that assume an empty extras list must add cleanup first: ss extras remove rules --force -g 2>/dev/null || true + rm -rf ~/.claude/rules
• Correct confirmation flags — uninstall uses --force (not --yes); init re-run needs no flag (just fails gracefully)
• Skill data in registry.yaml — assertions about installed skills check registry.yaml, NOT config.yaml; config.yaml should never contain skills:
• File existence timing — registry.yaml is only created after first install/reconcile, not on ss init
• Project mode paths — project commands use .skillshare/ not ~/.config/skillshare/
• Project init flags — init -p only supports --targets, --discover, --select, --mode, --dry-run; global-only flags (--no-copy, --no-skill, --no-git, --all-targets, --force) are not available
• Audit rule IDs — custom rules in audit-rules.yaml use rule IDs (e.g. prompt-injection-0), not pattern names (e.g. prompt-injection). Verify IDs against internal/audit/rules.yaml
• Use --json for assertions — if the command supports --json, use it with jq instead of grepping human-readable output. Text output changes between versions; JSON structure is stable
• Expected = actual substrings, NOT descriptions — the runbook assertion engine does case-insensitive substring matching. Write - Installed or - cangjie-docs-navigator, NOT - Install completes without error or - Output contains at least one skill. Negation: use Not <substring> prefix (e.g. - Not cangjie-docs-navigator)
• Skill name ≠ repo name — after ss install <repo>, the actual skill name may differ from the repo name (e.g. repo cangjie-docs-mcp → skill cangjie-docs-navigator). Always verify the installed skill name via ss list before writing uninstall/check steps
• /tmp/ cleanup — ssenv only isolates $HOME; /tmp/ is shared across runs. Any step using /tmp/<path> must start with rm -rf /tmp/<path> to avoid stale state from previous runs
• echo > symlink writes through — echo "content" > path where path is a symlink writes to the symlink's target, it does NOT replace the symlink with a real file. To create a local (non-managed) file at a symlinked path: either use a different filename, or rm the symlink first then echo
• cat >> is not idempotent — appending to config files (cat >> config.yaml) will duplicate sections on re-run. Prefer ss extras init (which validates duplicates) or full file replacement over cat >> when possible
• Extras source path layout — extras use ~/.config/skillshare/extras/<name>/ (not the legacy flat path ~/.config/skillshare/<name>/). Symlink assertions must include extras/ in the path regex (e.g. regex: skillshare/extras/rules/tdd\.md)
• Prefer jq: over python3 -c — for JSON output validation, use mdproof's native jq: assertion type (e.g. - jq: .extras | length == 1) instead of piping to python3 -c. It's one line vs 10, and mdproof handles failure reporting automatically
• Config append idempotency — when appending YAML sections with cat >>, always prepend sed -i '/^section_key:/,$d' to remove existing section. Or prefer CLI commands (ss extras init, ss extras remove --force) over manual config editing
• Check lessons-learned — read .mdproof/lessons-learned.md before writing new runbooks for known gotchas and proven assertion patterns

Runbook Assertion Types

mdproof supports 6 assertion types under Expected: blocks. Use the most specific type for each check:

jq: best practices:

# Simple field check
- jq: .name == "rules"

# Array length
- jq: .extras | length == 3

# Sorted array comparison
- jq: [.extras[].name] | sort | . == ["a","b","c"]

# Null/missing field (omitempty)
- jq: .extras == null

# Nested access
- jq: .[0].targets[0].status == "synced"

# Boolean
- jq: .source_exists == true

Rules

• Always execute inside devcontainer — use docker exec, never run CLI on host
• Always use ssenv for HOME isolation — don't pollute container default HOME
• Always create fresh ssenv environments — never reuse an environment from a previous run; stale config/state causes confusing cascade failures (e.g. duplicate YAML keys, "already exists" errors)
• ssenv only isolates $HOME — /tmp/, /var/, and other system paths are shared across all environments. Runbook steps using /tmp/ must include rm -rf cleanup at the start
• Verify every step — never skip Expected checks
• Don't abort on failure — record FAIL, continue to next step, summarize at end
• Ask before cleanup — Phase 4 must prompt user before deleting ssenv environment
• ss = skillshare — same binary in runbooks
• ~ = ssenv-isolated HOME — ssenv enter auto-sets HOME
• Use --init — simplify setup by using ssenv create <name> --init
• --init already runs init — the env is pre-initialized; runbook steps calling ss init again will fail unless the step explicitly resets state first

ssenv Quick Reference

• For interactive debugging: ssnew <env> then exit when done
• For deterministic automation: prefer ssenv enter <env> -- <command> one-liners

Test Command Policy

When running Go tests inside devcontainer (not via runbook):

# ssenv changes HOME, so always cd to /workspace first for Go test commands
cd /workspace
go build -o bin/skillshare ./cmd/skillshare
SKILLSHARE_TEST_BINARY="$PWD/bin/skillshare" go test ./tests/integration -count=1
go test ./...

Always run in devcontainer unless there is a documented exception. Note: ssenv enter changes HOME, which may affect Go module resolution — always cd /workspace before running go test or go build.

--json Quick Reference

Most commands support --json for structured output, making assertions more reliable than text matching.

Key behaviors:

• --json that implies --force / --all skips interactive prompts — safe for automation
• Output goes to stdout only (progress/spinners suppressed)
• audit prefers --format json; --json still works but is the deprecated form
• log --json outputs JSONL (newline-delimited), not a JSON array

Assertion Patterns with jq

# Count installed skills
ss list --json | jq 'length'

# Check a specific skill exists
ss list --json | jq -e '.[] | select(.name == "my-skill")'

# Verify target is configured
ss target list --json | jq -e '.[] | select(.name == "claude")'

# Assert no critical audit findings
ss audit --format json | jq -e '.summary.critical == 0'

# Check update availability
ss check --json | jq -e '.tracked_repos | length > 0'

# Verify sync succeeded (zero errors)
ss sync --json | jq -e '.errors == 0'

# Install and verify result
ss install https://github.com/user/repo --json | jq -e '.skills | length > 0'

When a jq -e expression fails (exit code 1 = false, 5 = no output), the step FAILs — no ambiguous text matching needed.

Container Command Templates

# Single command
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- ss status

# JSON assertion (preferred for verification)
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- bash -c '
  ss list --json | jq -e ".[] | select(.name == \"my-skill\")"
'

# Multi-line compound command (use bash -c) — global mode flags
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- bash -c '
  ss init --no-copy --all-targets --no-git --no-skill
  ss status
'

# Project mode init (different flag set!)
docker exec $CONTAINER env SKILLSHARE_DEV_ALLOW_WORKSPACE_PROJECT=1 \
  ssenv enter "$ENV_NAME" -- bash -c '
  cd /tmp/test-project && ss init -p --targets claude
'

# Check files (HOME is set to isolated path by ssenv)
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- bash -c '
  cat ~/.config/skillshare/config.yaml
'

# With environment variables
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- bash -c '
  TARGET=~/.claude/skills
  ls -la "$TARGET"
'

# Go tests (must cd /workspace because ssenv changes HOME)
docker exec $CONTAINER ssenv enter "$ENV_NAME" -- bash -c '
  cd /workspace
  go test ./internal/install -run TestParseSource -count=1
'

Relationship with /mdproof Skill

This skill (/cli-e2e-test) and the /mdproof skill are complementary, not competing:

How they work together

1. Writing a new runbook → invoke /mdproof first for format guidance (assertion types, jq: patterns, snapshot usage), then /cli-e2e-test to execute it in isolation
2. Improving existing runbooks → invoke /mdproof for assertion quality review (python3 → jq:, idempotency), then /cli-e2e-test to verify changes pass
3. Debugging failures → /cli-e2e-test Phase 3 step 4 handles manual docker exec; /mdproof lessons-learned captures recurring patterns
4. After a test run → /mdproof Self-Learning section guides recording discoveries to .mdproof/lessons-learned.md

Rule of thumb

• Need to run tests or debug in devcontainer? → /cli-e2e-test
• Need to write assertions or improve runbook quality? → /mdproof
• User says "run extras E2E" → /cli-e2e-test
• User says "improve runbook assertions" → /mdproof then /cli-e2e-test to verify