claude-skills

Upgrade the pw plugin to **v2.12.0 or later**. In v2.9.0, .mcp.json automatically starts pw-testrail and pw-browserstack, but their @modelcontextprotocol/sdk dependency is never installed, causing ERR_MODULE_NOT_FOUND and permanent CONNECTION_CLOSED errors. v2.12.0 moves these servers to integrations/mcp-servers.example.json, so they no longer fail for every user.

If you need TestRail or BrowserStack integration, explicitly opt in:

cd "$CLAUDE_PLUGIN_ROOT/integrations/testrail-mcp"
npm install

cd "$CLAUDE_PLUGIN_ROOT/integrations/browserstack-mcp"
npm install

Then copy or merge integrations/mcp-servers.example.json into the plugin's .mcp.json, configure the required TESTRAIL_* or BROWSERSTACK_* environment variables, and restart Claude Code.

On v2.9.0, the immediate workaround is to remove the unused pw-testrail and pw-browserstack entries from .mcp.json, or run npm install in each integration directory before using them.

The hook command runs from the session's current working directory, not the plugin root, so a cwd-relative path cannot resolve. In engineering-team/self-improving-agent/hooks/hooks.json, replace "command": "./hooks/error-capture.sh" with:

"command": "\"${CLAUDE_PLUGIN_ROOT}/hooks/error-capture.sh\""

${CLAUDE_PLUGIN_ROOT} points to the installed plugin directory, and the escaped quotes support plugin paths containing spaces. Also update the same relative path in the plugin's settings.json install example and the script header comment for consistency.

The Windows console is using cp1252, which cannot encode characters such as ╔, ═, or ≥. Run Python in UTF-8 mode before executing the script:

- Command Prompt:

set PYTHONUTF8=1
  python your_script.py

$env:PYTHONUTF8="1"
  python your_script.py

To make it permanent for future Command Prompt sessions:

setx PYTHONUTF8 1

For an application-level fix, add this near the top of each CLI entry point before printing Unicode output:

import sys

if hasattr(sys.stdout, "reconfigure"):
    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
if hasattr(sys.stderr, "reconfigure"):
    sys.stderr.reconfigure(encoding="utf-8", errors="replace")

errors="replace" prevents crashes by substituting unsupported characters instead of raising UnicodeEncodeError. Python 3.7 or later supports UTF-8 mode and reconfigure(). If available, update to a release containing the cp1252 hardening fix for the affected scripts.

This is a known false-positive bug fixed in claude-skills v2.13.0 (PR #1013, landed via #1038). Two root causes in skill_review_checklist_runner.py: (1) the frontmatter parser only matched bare > and | YAML block scalars, so descriptions written with chomping/indent indicators (>-, >+, |-, |2) caused the function to return the indicator itself as the description, making the trigger check unpassable; (2) the trigger patterns were a fixed two-word list (use when, invoke before, etc.), rejecting valid phrasing like 'Use whenever', 'This skill should be used when', 'Use this before', or 'Load before'. Fix: upgrade to v2.13.0+, which parses all block scalar variants via re.fullmatch(r"[>|][0-9]*[-+]?|[>|][-+]?[0-9]*", val) and replaces the fixed list with a verb-plus-cue regex (verbs: use/used/invoke/apply/run/load/fire/trigger; cues: when/whenever/before/after/during/while/at the start of/on every). Workaround on older versions: use bare > instead of >- in frontmatter, and phrase triggers exactly as 'use when/for/before/during/after/while' or 'invoke when/before/after'.

In versions before v2.13.0, scripts/derive_counters.py --check only validated absolute totals in five named files (README.md, CLAUDE.md headline lines, marketplace.json, mkdocs.yml, .codex-plugin/plugin.json). It never read CHANGELOG.md, and its CLAIM_PATTERNS only matched absolute totals, not delta claims of the form X → Y — so a fabricated delta like skills 999 → 1234 in the changelog passed CI green. Fixed in v2.13.0 (via PR #1010, landed through #1038): the gate now reads CHANGELOG.md, recognizes delta patterns (A → B and A -> B), and asserts the right-hand side (B) equals the derived value. Upgrade to v2.13.0 or later: https://github.com/alirezarezvani/claude-skills/releases/tag/v2.13.0. Note: the proposed --write auto-generation mode (option 3) was not implemented and may be filed as a separate issue.