tempguru-pi — TempGuru event staffing for Pi and Prime Agent

This package documents the runtime-adapted skill and native-tool behavior included in version 1.7.2. Publication status is tracked in the repository, not inside this immutable npm artifact.

One package gives either Pi or Prime Agent both layers:

• 8 skills (Markdown, generated from canonical content/skills/*.md by gen-skill-digests.mjs with a Pi runtime-routing preamble and native tool-name substitutions; CI fails if either layer drifts): ordering, compliance, event-brief extraction, urgent backfill, agency partnering, multi-city activation, procurement, and pro operations.
• 9 native read-only tools (extensions/tempguru.ts) calling TempGuru's hosted REST API with runtime-aware attribution (?source=pi in Pi and ?source=prime-agent in Prime Agent): cities, roles, availability, pricing, compliance, policies, saved plans, quote status, and tempguru_request_quote. The quote tool accepts only a saved plan ID and returns a TempGuru form the buyer submits personally; it never accepts PII or creates a CRM lead.

This closes most of the gap the audit flagged: neither runtime gains API access from a Markdown skill alone, so the package installs an attributed native action layer with the skills. No MCP bridge is required for the 9 granular operations. The full plan_staffing planner, explicit save_staffing_plan artifact write, and get_rate_benchmark remain MCP-only in version 1.7.2. Pi can attach the remote MCP through a compatible client; Prime Agent has the authless-MCP limitation described below. See llms-install.md until native REST parity ships in a follow-up release.

The generated runtime copies perform this mapping inside every installed skill, so the model sees it at runtime (the README is not relied on as hidden context):

| Canonical name | Installed native name | |---|---| | get_cities | tempguru_get_cities | | get_roles | tempguru_get_roles | | check_availability | tempguru_check_availability | | get_role_pricing | tempguru_get_role_pricing | | get_compliance_by_state | tempguru_get_compliance | | get_policies | tempguru_get_policies | | get_plan | tempguru_get_plan | | get_quote_status | tempguru_quote_status | | request_quote | tempguru_request_quote |

There is no native mapping yet for plan_staffing, save_staffing_plan, or get_rate_benchmark. Every generated package skill therefore contains the same explicit fallback: use the remote MCP when attached; otherwise compose a transparent straight-time plan from the granular native tools, never invent a saved plan_id, and use city pricing or the public OKF Rate Index instead of inventing planner/benchmark output.

Pi can use a separately attached trusted MCP client for those three operations. Prime Agent v0.7.0's stock Python MCP integration currently requires OAuth or a bearer token, so it cannot attach TempGuru's authless server as-is. Do not add an ineffective mcpServers entry in Prime; those three operations remain unavailable there until the native adapter gains parity or Prime supports authless MCP.

Install (users)

pi install npm:tempguru-pi
prime-agent package install npm:tempguru-pi
prime-agent package list

There is no separate Prime Agent marketplace or TempGuru Prime package. Prime intentionally consumes the same pi.skills and pi.extensions manifest.

Layout

package.json            # shared pi manifest: skills + extensions
extensions/tempguru.ts  # native tools -> /api/v1/* with Pi/Prime attribution
skills/<slug>/SKILL.md  # 8 runtime-adapted canonical skills, generated — never hand-edit

Regenerate the skills after any content/skills/*.md change (runs inside npm run build:okf):

node scripts/gen-skill-digests.mjs

Publish (maintainer)

The canonical publisher is the manual .github/workflows/publish-pi.yml GitHub Action. It validates the requested version, package identity, generated skill digests, and tarball contents before publishing with npm Trusted Publishing (OIDC). From the repository:

gh workflow run publish-pi.yml \
  --repo Tempguru-co/tempguru-mcp \
  --ref main \
  -f version=1.7.2

Notes:

• Before the first OIDC release, configure the existing tempguru-pi package's npm Trusted Publisher with this exact identity: organization Tempguru-co, repository tempguru-mcp, workflow filename publish-pi.yml, environment Production, allowed action npm publish. The workflow must be run from main.
• Do not run npm publish locally. The GitHub workflow is the release path and requires no long-lived npm token.
• The name is unscoped tempguru-pi to match tempguru-mcp (no npm org required). If you later create the @tempguru npm org, publish as @tempguru/pi-event-staffing and deprecate this name in favor of it.
• Pi package patch versions may move independently from the MCP CLI when the native adapter changes. Always bump to an unpublished version before release, and record the published Pi version in distribution/assistants/README.md.
• The pi-package keyword is required for Pi discovery and Prime package compatibility; prime-agent and agent-skills make that support explicit.
• Pi and Prime Agent both provide the core typebox runtime; keep it as a "*" peer dependency rather than bundling a second copy.
• After publishing, verify the promise the README makes in both runtimes. Run pi install npm:tempguru-pi, then in Pi run /skills. Also run prime-agent package install npm:tempguru-pi and prime-agent package list. Each runtime must expose 8 TempGuru skills and the 9 tools tempguru_get_cities … tempguru_request_quote.
• Confirm npm independently before calling the release live: npm view [email protected] version.
• Record the listing in distribution/assistants/README.md's status tracker.

Safety posture (mirrors the MCP server)

• tempguru_request_quote is read-only. It requires a saved plan_id, returns a prefilled https://mcp.tempguru.co/request-quote URL, and never accepts or transmits contact details. The buyer must review and submit the form personally; only the REST form submission creates a lead or TG reference.
• Rates are all-inclusive W-2 bill rates presented as planning estimates, never binding quotes; availability responses are guidance, never reservations.
• Compliance data is operational guidance, not legal advice.