The repo-as-Claude-Code-plugin manifest at plugin/.claude-plugin/plugin.json
declares "agents": "./agents", but the symlink at plugin/agents pointed
at ../v2/.claude/agents — a path that never existed at repo root and
was definitively gone after archive/v2/ moves. Result: anyone installing
this repo via /plugin got zero agents loaded.
Repointed to ../.claude/agents (33 active md files: coder, reviewer,
tester, planner, researcher, security-architect, etc). The hooks /
commands / skills symlinks were already correct.
Verified: ls -L plugin/agents resolves to 33 entries; matches what
.claude/agents/ contains.
Co-Authored-By: RuFlo <ruv@ruv.net>
The Liberation of Bajor team's report (#1747) is a precise diagnosis:
\`plugin/hooks/hooks.json\` interpolated \`\$TOOL_INPUT_command\` (and
\`\$TOOL_INPUT_*\`, \`\$PROMPT\`, \`\$TOOL_NAME\` etc.) directly into
double-quoted command strings. When the bash command being executed
contained both \" and >, the wrapping quotes broke and the shell
re-parsed the inner content. Any > became a redirect, creating empty
files at CWD with names like '3', '40)', 'maxWidth)', etc.
After ~24h of normal usage on Windows + Git Bash, dozens of empty
files accumulated at every project root.
Fix: rewrite ALL 13 command hooks to use the stdin-jq-xargs pattern
that the companion file \`.claude-plugin/hooks/hooks.json\` already
uses safely:
cat | jq -r '.tool_input.command // empty' | tr '\\\\n' '\\\\0' \\
| xargs -0 -I {} npx claude-flow@alpha hooks pre-command --command '{}'
Reading via stdin + jq + xargs -0 -I {} bypasses shell re-parsing
entirely. The user input becomes a single argv element, never
re-interpreted as shell tokens.
Validation (script in commit):
safe (stdin-jq): 13 unsafe (\$-interp): 0
Affected hooks (all rewritten):
PreToolUse: Bash, Write|Edit|MultiEdit, Task, Grep|Glob|Read,
mcp__claude-flow__*
PostToolUse: same 5 matchers
UserPromptSubmit: route by .prompt
SessionStart: by .session_id
Notification: by .message
Dropped non-essential \$TOOL_SUCCESS / \$TOOL_EXIT_CODE flags from
post-* hooks — they're internal booleans/numbers (not user input,
so not the file-creation vector) but the cleaner stdin-only pattern
is preferable over mixed inline/stdin.
Stop / SubagentStop / PermissionRequest hooks unchanged — they don't
interpolate user input.
This is a plugin-file change; no npm publish needed. The fix is live
when the PR merges and users pull main / re-clone.
Closes#1747.
Co-authored-by: Reuven <cohen@ruv-mac-mini.local>
The PreToolUse:Task hook used $TOOL_INPUT_prompt which expands to the
full agent prompt (thousands of characters), overflowing the shell's
argument length limit. Switched to $TOOL_INPUT_description (the short
3-5 word field) and truncated to 200 chars.
Fixes#1160
