| name | sherpamind |
| description | Use for SherpaDesk-related requests: ticket lookup, support-history retrieval, account/user/technician analysis, stale-ticket review, workload questions, operational reporting, and open-ended natural-language questions about SherpaDesk data. This skill is a local SherpaDesk backend plus OpenClaw query layer: it requires SherpaDesk API credentials for live setup, creates workspace-local runtime state under `.SherpaMind/`, and may install an optional user-level background service. Trigger when the user mentions SherpaDesk or asks about tickets, support issues, clients/accounts, technicians, resolution history, recurring incidents, backlog, response timing, or similar support-operations analysis. |
| metadata | {"openclaw":{"emoji":"🧰","homepage":"https://github.com/kklouzal/SherpaMind","requires":{"anyBins":["python3","python"]},"primaryEnv":"SHERPADESK_API_KEY"}} |
SherpaMind
Use SherpaMind as the OpenClaw query/action layer over the local SherpaDesk dataset prepared by the backend service.
Repo root and stable entrypoint
Work from the repo root:
cd {baseDir}
When the repo is installed under an OpenClaw skills/ directory, SherpaMind automatically uses the parent workspace as SHERPAMIND_WORKSPACE_ROOT, so runtime state stays in workspace-level .SherpaMind/ rather than inside the skill checkout. If runtime data already exists somewhere else, set SHERPAMIND_ROOT=/path/to/.SherpaMind before running commands; it takes precedence over workspace-relative discovery.
Before reading generated files directly, resolve the actual runtime paths instead of assuming they live under {baseDir}:
python3 scripts/run.py workspace-layout
Use the stable runner:
python3 scripts/run.py <command> [args...]
Do not invent alternate runtime paths.
Do not treat OpenClaw as the background scheduler for this backend.
Transparency and operator expectations
SherpaMind is not an instruction-only skill.
When installed and configured for live use, it can:
- create workspace-local runtime state under
.SherpaMind/
- create staged runtime dirs under
.SherpaMind/private/config/, .SherpaMind/private/data/, .SherpaMind/private/state/, .SherpaMind/private/logs/, .SherpaMind/private/runtime/, and .SherpaMind/public/
- create a local SQLite database and generated public artifacts
- create a Python runtime venv under
.SherpaMind/private/runtime/venv
- install Python dependencies from PyPI during bootstrap
- require the SherpaDesk API key to be provided via the OpenClaw-managed
SHERPADESK_API_KEY environment surface (the daemon mirrors this from the OpenClaw skill config into its own service env)
- receive optional SherpaDesk API user hints from environment variables or OpenClaw-managed config
- store non-secret connection/runtime settings in
.SherpaMind/private/config/settings.env
- read alert enablement and alert destinations from first-class OpenClaw
sherpamind skill fields
- optionally install and run a user-level
systemd background service
Required live staged credentials/config for real SherpaDesk use:
- OpenClaw-managed
SHERPADESK_API_KEY configured for the sherpamind skill
- staged org/instance settings in
.SherpaMind/private/config/settings.env
- OpenClaw
skills.entries.sherpamind.config alert fields when chat alerting is wanted
If the user only wants query guidance or offline inspection of an existing local dataset, do not imply that fresh credentials or service installation are unnecessary for live sync.
Choose the lightest path that answers the question
Exact facts, counts, status, and workload
Start with structured commands:
python3 scripts/run.py dataset-summarypython3 scripts/run.py report-api-usagepython3 scripts/run.py report-enrichment-coveragepython3 scripts/run.py insight-snapshotpython3 scripts/run.py report-ticket-countspython3 scripts/run.py report-status-countspython3 scripts/run.py report-priority-countspython3 scripts/run.py report-technician-countspython3 scripts/run.py report-ticket-log-typespython3 scripts/run.py report-attachment-summarypython3 scripts/run.py report-ticket-classespython3 scripts/run.py report-ticket-class-coveragepython3 scripts/run.py dispatch-ticket-classificationspython3 scripts/run.py refresh-ticket-class-taxonomy --forcepython3 scripts/run.py writeback-ticket-classifications --limit 1python3 scripts/run.py report-ticket-classificationspython3 scripts/run.py recent-ticketspython3 scripts/run.py open-ticket-agespython3 scripts/run.py recent-account-activitypython3 scripts/run.py recent-technician-loadpython3 scripts/run.py account-summary "<account>"python3 scripts/run.py technician-summary "<technician>"python3 scripts/run.py ticket-summary "<ticket-id|ticket-number|ticket-key>"
Examples:
- open-ticket count →
report-status-counts
- technician backlog/load →
technician-summary "<technician>"
- account snapshot →
account-summary "<account>"
- ticket inspection / retrieval-ready context →
ticket-summary "<ticket-id|ticket-number|ticket-key>"
Fuzzy investigation, prior-art lookup, and support-history recall
Use retrieval commands:
python3 scripts/run.py search-ticket-docs "<query>"python3 scripts/run.py search-ticket-docs "<query>" --account "<account>" --status Open --department "<department>"python3 scripts/run.py search-ticket-chunks "<query>"python3 scripts/run.py search-ticket-chunks "<query>" --account "<account>" --status Open --technician "<technician>"python3 scripts/run.py search-ticket-chunks "<query>" --priority High --category "<category>" --class-name "<class>"python3 scripts/run.py search-vector-index "<query>"python3 scripts/run.py search-vector-index "<query>" --account "<account>" --status Openpython3 scripts/run.py search-vector-index "<query>" --technician "<technician>" --priority High --category "<category>"python3 scripts/run.py search-vector-index "<query>" --department "<department>" --class-name "<class>" --submission-category "<channel>" --resolution-category "<resolution>"
Default retrieval workflow:
- Start with keyword/text search when the issue words are concrete.
- Widen to vector search when wording may vary or keyword recall looks thin.
- Use account/technician/status/priority/category/department/class/submission/resolution filters when they materially narrow the search.
- Answer from retrieved evidence instead of jumping to canned conclusions.
Quick factual context from generated artifacts
Read these when a concise derived artifact is enough. First run python3 scripts/run.py workspace-layout, then use the returned docs_root instead of hard-coding {baseDir}/.SherpaMind:
{docs_root}/index.md{docs_root}/insight-snapshot.md{docs_root}/stale-open-tickets.md{docs_root}/recent-account-activity.md{docs_root}/recent-technician-load.md{docs_root}/runtime/status.md{docs_root}/accounts/index.md{docs_root}/technicians/index.md{docs_root}/tickets/index.md{docs_root}/accounts/*.md{docs_root}/technicians/*.md{docs_root}/tickets/ticket_*.md
Preferred answer flow
For broad questions like “what’s been going on with account X lately?” or “have we seen this before?”:
- Pull one structural summary first.
- Pull retrieval evidence second.
- Use generated public docs only when they add concise context.
- Give the user an answer grounded in the retrieved evidence.
Prefer factual retrieval over hand-authored interpretation.
End-to-end install and onboarding on another OpenClaw instance
If the user asks to install SherpaMind properly end-to-end into an OpenClaw instance, first check the host prerequisites and report any missing pieces plainly before continuing.
Minimum prerequisites to check:
python3 is present- Python venv/pip bootstrap works on that host
- the host has network access for Python package installation
systemctl --user is available if background service mode is expected
If any prerequisite is missing, stop and tell the user exactly what is missing and what needs to be fixed.
Then use this flow from the installed skill bundle root:
- audit bootstrap/readiness first
python3 scripts/run.py bootstrap-audit
- bootstrap the skill-local runtime
python3 scripts/bootstrap.py
- run the setup flow
python3 scripts/run.py setup
- verify runtime state
python3 scripts/run.py doctor
- ensure the OpenClaw
sherpamind skill entry has the SherpaDesk API key and any wanted alert fields:
apiKeyconfig.newTicketAlertsEnabledconfig.ticketUpdateAlertsEnabledconfig.newTicketAlertChannelconfig.ticketUpdateAlertChannelconfig.alertModelconfig.alertThinking
- discover organizations/instances
python3 scripts/run.py discover-orgs
- write the chosen org/instance into non-secret settings
python3 scripts/run.py configure --org-key <org> --instance-key <instance>
- seed the local dataset
python3 scripts/run.py seed
- generate/refine the derived artifacts if needed
python3 scripts/run.py generate-public-snapshotpython3 scripts/run.py generate-runtime-status
- confirm the install is actually usable
python3 scripts/run.py dataset-summarypython3 scripts/run.py insight-snapshotpython3 scripts/run.py report-vector-index-status
- only then decide whether unattended background mode is wanted
python3 scripts/run.py install-servicepython3 scripts/run.py service-status
Default expectation on Linux is that setup initializes the DB, cleans up any old SherpaMind cron jobs, and can generate an initial public snapshot. Treat user-level systemd installation as a later, explicit operator choice rather than part of the earliest bootstrap steps.
If service installation fails because the target host lacks usable systemctl --user, continue the bootstrap/config/seed flow anyway, report the service limitation clearly, and use the worker-specific fallback commands (hot-watch-run-once, alert-dispatch-run-once, maintenance-run-once, or the corresponding foreground loops) instead of pretending the split services installed.
If install/runtime/use issues or meaningful feature gaps are discovered while operating SherpaMind, check https://github.com/kklouzal/SherpaMind/issues. If a matching issue exists, add supporting detail; otherwise open a new issue with clear reproduction/context. Keep issue content anonymized and public-safe.
Lifecycle and maintenance commands
Use these for setup/maintenance, not routine user queries:
python3 scripts/bootstrap.pypython3 scripts/run.py workspace-layoutpython3 scripts/run.py doctorpython3 scripts/run.py backfill-technician-stubspython3 scripts/run.py backfill-ticket-entity-stubspython3 scripts/run.py bootstrap-auditpython3 scripts/run.py setuppython3 scripts/run.py migrate-legacy-statepython3 scripts/run.py archive-legacy-statepython3 scripts/run.py cleanup-legacy-cronpython3 scripts/run.py discover-orgspython3 scripts/run.py configure --org-key <org> --instance-key <instance>python3 scripts/run.py install-servicepython3 scripts/run.py restart-servicepython3 scripts/run.py service-statuspython3 scripts/run.py hot-watch-run-oncepython3 scripts/run.py alert-dispatch-run-oncepython3 scripts/run.py maintenance-run-oncepython3 scripts/run.py generate-public-snapshotpython3 scripts/run.py generate-runtime-statuspython3 scripts/run.py sync-ticket-classespython3 scripts/run.py report-ticket-classespython3 scripts/run.py report-ticket-class-coveragepython3 scripts/run.py dispatch-ticket-classificationspython3 scripts/run.py refresh-ticket-class-taxonomy --forcepython3 scripts/run.py writeback-ticket-classifications --limit 1python3 scripts/run.py report-ticket-classifications
Guarded write-backs
SherpaMind now has one deliberately narrow SherpaDesk write-back lane: while normal backend daemon work is already touching old closed tickets, any observed ticket with status=Closed, closed_time at least 365 days old, and is_confirmed=false is automatically updated with PUT /tickets/{id} and is_confirmed=true. This rides the slow rolling warm/cold/enrichment cadence rather than scanning aggressively.
Manual commands remain available:
python3 scripts/run.py confirm-stale-unconfirmed-closed-tickets — dry-run list from the local DBpython3 scripts/run.py confirm-stale-unconfirmed-closed-tickets --apply — live write-back for local DB candidates, then refreshes/materializes touched tickets
Boundaries
- Treat SherpaMind as read-only except for explicitly supported guarded write-back behavior: automatic 365-day stale confirmation during normal daemon processing; active-leaf ticket classification write-back after local event classification; plus the manual stale-confirmation and classification write-back command surfaces.
- Keep attachment handling metadata-only by default.
- Do not auto-download attachment bodies by default.
- Treat docs, chunks, vector rows, and public Markdown artifacts as replaceable derived caches.
- Let SherpaMind prepare and expose data; let OpenClaw interpret it at answer time.
- For automatic LLM-backed work, keep the accuracy-to-token contract tight: event-triggered only, compact bounded payloads, locally retrieved evidence when useful, no broad searches from alert/classification prompts, and only narrow extra SherpaMind retrieval for alerts when it materially improves accuracy.
References
Read these only when needed. Keep the action layer in this file lean; use the reference files for deeper architecture, retrieval, automation, and API details.
{baseDir}/README.md — current live project overview and command surface{baseDir}/references/openclaw-query-model.md — query/retrieval model; prefer compact/default snippets first and use --full only after narrowing{baseDir}/references/architecture-doctrine.md — backend vs skill-front boundary{baseDir}/references/retrieval-architecture.md — retrieval and vector design{baseDir}/references/bootstrap-onboarding.md — audit-first install/onboarding model{baseDir}/references/automation.md — service/install/update model{baseDir}/references/delta-sync-strategy.md — hot/warm/cold sync behavior{baseDir}/references/api-reference.md — verified API/auth behavior
