The CLAUDE.md mandatory rule instructed clearing the plugin cache
mid-session, which breaks all MCP tools that were already loaded.
MCP tools are loaded with absolute paths to the venv, and deleting
the cache removes the venv while the session still references it.
Changes:
- CLAUDE.md: Replace "ALWAYS CLEAR CACHE" with "VERIFY AND RESTART"
- verify-hooks.sh: Change cache existence from error to info message
- DEBUGGING-CHECKLIST.md: Add section explaining cache timing
Fixes#145
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Expand parameter schemas for 15 update tools that previously only exposed
the 'id' field. The underlying Python implementation already supported
all fields via **kwargs, but Claude couldn't discover available parameters.
Updated tools:
- virtualization: update_virtual_machine, update_cluster
- dcim: update_site, update_location, update_rack, update_manufacturer,
update_device_type, update_device_role, update_platform,
update_interface, update_cable
- ipam: update_vrf, update_prefix, update_ip_address, update_vlan
Each tool now exposes all commonly-used optional fields matching the
NetBox API, following the pattern established by dcim_update_device.
Fixes#137
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Claude API has a 64-character limit on tool names. Claude Code uses a
36-character prefix (mcp__plugin_cmdb-assistant_netbox__), leaving only
28 characters for the actual tool name.
Shortened 33 tools that exceeded this limit:
- virtualization_* -> virt_* (19 tools)
- circuits_*_circuit_type* -> circ_*_type* (3 tools)
- circuits_*_circuit_termination* -> circ_*_termination* (3 tools)
- wireless_*_wireless_* -> wlan_* (8 tools)
Added TOOL_NAME_MAP to route shortened names to original method names.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The /initial-setup command was generating NETBOX_API_URL without the
/api suffix, causing all MCP tools to fail with JSON parsing errors.
Changes:
- Update example URL to include /api suffix
- Add instruction to auto-append /api if user forgets
- Fix validation test to be consistent with actual usage
Root cause: Setup asked for base URL (https://netbox.company.com) but
NetBoxClient expects API URL (https://netbox.company.com/api). The
validation test incorrectly added /api/ making it pass, while the
actual MCP server failed.
Fixes#126
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Fixes multiple issues from diagnostic #123:
1. create_label_smart type safety (labels.py)
- Add isinstance(result, dict) checks after API calls
- Return structured error dict if API returns unexpected type
- Prevents "list indices must be integers" crash
2. debug-report always uses curl with labels
- Remove MCP option - always use curl for marketplace issues
- Add label ID fetching step (Source/Diagnostic, Type/Bug)
- Include labels in curl POST payload
- Avoids branch protection restrictions on main branch
Fixes#123
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When NetBox MCP tools fail with JSON decode errors, the error message
now includes:
- HTTP status code
- Response content length
- Preview of actual content received (first 200 bytes)
This helps diagnose transient issues like network timeouts or
incomplete responses that result in cryptic "Expecting value" errors.
Fixes#120
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Expose additional parameters in dcim_create_device and dcim_update_device
MCP tools that were already supported by the backend but not exposed:
dcim_create_device:
- platform, primary_ip4, primary_ip6, asset_tag, description, comments
dcim_update_device:
- platform, primary_ip4, primary_ip6, serial, asset_tag, site, rack,
position, description, comments
This enables setting the platform (OS) and primary IP address when
creating or updating devices in NetBox.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
SessionStart hook checks if CLAUDE.md has mandatory rules.
If missing, adds them automatically.
Rules enforced:
- Check everything when user asks
- Believe user when they say something's wrong
- Never say "done" without verification
- Show exactly what user asks for
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add MANDATORY BEHAVIOR RULES to CLAUDE.md (read every session)
- Rules: check everything, believe user, verify before saying done
- post-update.sh now clears plugin cache
- verify-hooks.sh checks all locations for prompt hooks
These rules prevent wasted user time from AI overconfidence.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- post-update.sh now clears plugin cache automatically
- verify-hooks.sh checks ALL locations for prompt hooks
- Prevents cached old hooks from overriding fixed hooks
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
create_label_smart now checks if label already exists before creating.
- Checks both org and repo labels
- Handles format variations (Type/Bug vs Type: Bug)
- Returns {skipped: true} if label already exists
- Prevents duplicate label creation errors
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
ALL hooks now use command type (bash scripts) instead of prompt type.
Prompt hooks are unreliable - model ignores instructions.
Changes:
- projman: SessionStart → startup-check.sh with [projman] prefix
- pr-review: SessionStart → startup-check.sh with [pr-review] prefix
- project-hygiene: cleanup.sh now has [project-hygiene] prefix
- doc-guardian: already fixed (notify.sh with [doc-guardian] prefix)
- code-sentinel: already fixed (security-check.sh with [code-sentinel] prefix)
All hook output now guaranteed to have plugin name prefix.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Same fix as doc-guardian - prompt hooks unreliable.
Command hook guarantees exact behavior.
- Add security-check.sh that skips config/doc files silently
- Only checks code files for hardcoded secrets
- Outputs with [code-sentinel] prefix
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Prompt hook approach didn't work - Claude ignores instructions.
Real fix was switching to command hook type.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Prompt hooks are unreliable - Claude ignores instructions and generates
verbose analysis despite explicit FORBIDDEN rules. Command hooks guarantee
the exact output we want.
- Add notify.sh script that only outputs for config file changes
- Change hooks.json from prompt type to command type
- Script exits silently for non-config files (no blocking)
Fixes#110
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add protected branch detection to /commit command (Step 1)
- Warn users before committing to protected branches
- Offer to create feature branch automatically
- Rewrite doc-guardian hook to be truly non-blocking
- Enforce strict [plugin-name] prefix in all hook outputs
- Add forbidden words list to prevent accidental blocking
Fixes#109, #110
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add [plugin-name] prefix to all hook messages for better identification
- Make doc-guardian hook notification-only (non-blocking)
- Add stale branch detection to /commit-sync with git fetch --prune
- Enhance /branch-cleanup to handle stale branches separately
Closes improvements for hook UX and git workflow
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The MCP server's branch-aware security blocks write operations on
protected branches (main, fix/*, etc). After pushing a feature branch
and creating a PR, we must switch back to development before adding
comments to issues via MCP tools.
When MCP tools are not available in a session (the very scenario
/debug-report is designed to diagnose), the command now falls back to:
1. Check for Gitea credentials at ~/.config/claude/gitea.env
2. Use curl + jq to create the issue via Gitea REST API
3. If no credentials, save report to local file for manual submission
Security measures:
- Uses mktemp -m 600 for restrictive file permissions
- Uses jq --rawfile for safe JSON construction (no command substitution)
- Proper cleanup of temporary files
Fixes#100
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The suggest_labels tool accepted a repo parameter in the implementation
but didn't expose it in the MCP tool schema, causing it to always rely
on auto-detection which failed in some contexts.
Fixes#94
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
MCP servers fail when venvs don't exist in ~/.claude/plugins/marketplaces/.
Claude Code doesn't run setup.sh when installing marketplaces, so users
must run it manually.
Added:
- Critical warning section at top of UPDATING.md
- Step to run setup in installed location after updates
- Troubleshooting for "X MCP servers failed" error
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
