feat: add opt-in diagnostics via PostHog

Per-event consent diagnostics that sends anonymous install/update/skill data
to PostHog. Conflict filenames are gated against upstream. Supports --dry-run
to show exact payload before sending, and "never ask again" opt-out via state.yaml.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.claude/skills/_shared/diagnostics.md

@@ -0,0 +1,99 @@
+# Diagnostics (end-of-skill instructions)
+
+Follow these steps **after all other skill steps are complete**.
+
+## 1. Skip check
+
+If this skill was invoked from within another skill (e.g., `/add-whatsapp` called during `/setup`), **skip diagnostics entirely** — the parent skill handles it. Return without doing anything below.
+
+## 2. Determine event type
+
+- `/setup` → `setup_complete`
+- `/update-nanoclaw` → `update_complete`
+- Any other skill → `skill_applied`
+
+## 3. Build event-specific data
+
+Estimate `error_count` from the conversation (how many errors/retries occurred during the run).
+
+**For `setup_complete`:**
+```json
+{
+  "channels_selected": ["whatsapp", "telegram"],
+  "error_count": 0,
+  "failed_step": null,
+  "exit_code": null
+}
+```
+- `channels_selected`: which channels the user chose to install
+- `failed_step`: if setup failed, which step (environment / container / groups / register / mounts / service / verify)
+- `exit_code`: if a step failed, its exit code
+
+**For `skill_applied`:**
+```json
+{
+  "skill_name": "add-telegram",
+  "is_upstream_skill": true,
+  "conflict_files": ["package.json", "src/index.ts"],
+  "error_count": 0
+}
+```
+- `skill_name`: the skill that was run. Use the upstream skill name if it's an upstream skill, otherwise use `"custom"`
+- `is_upstream_skill`: true if this is a known upstream skill
+- `conflict_files`: filenames that had merge conflicts (the script will gate these against upstream — pass them all, the script filters)
+
+**For `update_complete`:**
+```json
+{
+  "version_age_days": 45,
+  "update_method": "merge",
+  "conflict_files": ["package.json"],
+  "breaking_changes_found": false,
+  "breaking_changes_skills_run": [],
+  "error_count": 0
+}
+```
+- `version_age_days`: estimate from the backup tag or commit date how many days old the previous version was
+- `update_method`: "merge" or "rebase"
+- `breaking_changes_found`: whether breaking changes were detected during the update
+- `breaking_changes_skills_run`: which skills had to be re-run to fix breaking changes
+
+## 4. Dry run
+
+Run with `--dry-run` to get the full payload:
+
+```bash
+npx tsx scripts/send-diagnostics.ts --event <event_type> --success --data '<json>' --dry-run
+```
+
+Use `--failure` instead of `--success` if the skill failed.
+
+If the command produces no output, the user has opted out permanently — skip the rest.
+
+## 5. Show the user and ask
+
+Show the JSON output and ask:
+
+> "Would you like to send anonymous diagnostics to help improve NanoClaw? Here's exactly what would be sent:"
+>
+> (show the JSON)
+>
+> **Yes** / **No** / **Never ask again**
+
+Use AskUserQuestion.
+
+## 6. Handle response
+
+- **Yes**: Run the same command without `--dry-run`:
+  ```bash
+  npx tsx scripts/send-diagnostics.ts --event <event_type> --success --data '<json>'
+  ```
+  Confirm: "Diagnostics sent."
+
+- **No**: Do nothing. User will be asked again next time.
+
+- **Never ask again**: Run:
+  ```bash
+  npx tsx -e "import { setNeverAsk } from './scripts/send-diagnostics.ts'; setNeverAsk();"
+  ```
+  Confirm: "Got it — you won't be asked again."