feedback

name

feedback

description

Drop a feedback message into the claude-tools feedback inbox. Use when the user explicitly requests it ('/inbox-feedback', 'log this as feedback', 'send this to claude-tools', 'report this back', 'feedback to the toolmaker') AND proactively, on your own initiative, when you observe friction with a claude-tools-shipped component during a session — a taskmaster skill that errored, a guard-hook that blocked the wrong command, an MCP tool that returned a confusing shape, a skill description that didn't match its behaviour, a slash command that did the wrong thing. Do not fire for generic project friction or external-tool friction; only for claude-tools components (taskmaster, reflect-auto-improve, guard-hooks, statusline, feedback-inbox itself, ue5-materials, etc.). Silent on success — no user confirmation needed; the user reviews during /inbox triage.

feedback — Producer

You drop one feedback message into the configured inbox. One invocation = one message. Never bundle multiple unrelated complaints into one message.

Scope guard (read before deciding to invoke proactively)

Fire ONLY for friction with components shipped from claude-tools:

• taskmaster (any skill, command, MCP tool, hook)
• reflect-auto-improve
• guard-hooks
• statusline
• feedback-inbox itself
• ue5-materials, image-gen, codex-dispatch, reality-reprojection, shader-nodes

DO NOT fire for:

• Bugs in the user's project code.
• Friction with external tools (git, pytest, Node, language servers).
• General confusion that isn't tied to a specific claude-tools component.

If in doubt, don't fire. The user can always invoke /inbox-feedback explicitly.

Procedure

1. Resolve the target. The plugin is installed at ${CLAUDE_PLUGIN_ROOT} in every session that has it enabled, regardless of cwd. Probe the config with:

   python -c "import sys, json; sys.path.insert(0, r'${CLAUDE_PLUGIN_ROOT}'); from scripts.resolve_target import resolve_target; r = resolve_target(); print(json.dumps({'enabled': r.enabled, 'inbox': str(r.inbox) if r.inbox else None, 'reason': r.reason}))"
   

   If enabled: false, emit ONE line to the conversation: feedback-inbox: <reason> — skipping. Then stop. Do not raise, do not retry.

2. Compose the message. Pick:

   • source: claude if you're invoking proactively; user if the user explicitly dictated text.
   • category: bug (something is broken) | friction (works, but awkward) | idea (improvement suggestion) | praise (something worked well) | question (you want the toolmaker to clarify something).
   • summary: one-line, ≤ 120 chars. Lead with the component name (e.g. "taskmaster:pick-task hangs on long worktree paths").
   • component: the specific claude-tools subsystem, e.g. taskmaster/pick-task, guard-hooks, reflect-auto-improve/harvest. Required when source=claude.
   • project + project_path: derive from the current working directory — Path.cwd().name and the absolute cwd. Leave empty if you can't determine them; do not block.
   • body: freeform markdown. Two sections is the convention:
     • ## What happened — concrete observation, with exact tool calls or error messages if available.
     • ## Suggested fix — your best guess at what would help. Optional but valued.

3. Write the message. Run:

   python -c "
   import sys, json
   from datetime import datetime, timezone
   from pathlib import Path
   sys.path.insert(0, r'${CLAUDE_PLUGIN_ROOT}')
   from scripts.resolve_target import resolve_target
   from scripts.write_message import MessagePayload, write_message
   r = resolve_target()
   if not r.enabled:
       print(f'feedback-inbox: {r.reason} — skipping.'); sys.exit(0)
   payload = MessagePayload(
       source='<claude|user>',
       category='<bug|friction|idea|praise|question>',
       summary='<your one-line summary>',
       body='''<your multi-line body>''',
       project=Path.cwd().name,
       project_path=str(Path.cwd()),
       component='<claude-tools/component>',
       created=datetime.now(tz=timezone.utc),
   )
   path = write_message(r.inbox, payload)
   print(f'feedback-inbox: logged \"{payload.summary}\" → {path.name}')
   "
   

   Substitute the angle-bracketed placeholders with your composed values. Be careful to escape any single quotes inside the body, or use a heredoc / temp file if the body is long.

4. Emit one line to the conversation. Whatever the write script prints. Do not say more. The user reviews during /inbox triage, not now.

Failure handling

• If the write script errors (permissions, disk full, malformed input), emit one line: feedback-inbox: write failed — <error>. Do not retry. Do not raise.
• If you can't determine component for a source=claude message, don't fire. The whole point of the scope guard is to tie observations to specific subsystems; an untagged message is noise.

Anti-patterns

• Bundling multiple unrelated observations into one message. One issue per message.
• Firing for friction with the user's project code (not a claude-tools component).
• Asking the user "should I log this as feedback?" — proactive invocation is silent by design.
• Re-logging the same observation a second time in the same session. If you already wrote one, you're done.