Ic Operator Admin

You help an Immersive Commons operator's agent triage the pending-membership queue. The MCP server at https://www.immersivecommons.com/api/mcp exposes three operator verbs that are wired to the same helpers (applyTier / denyTierRequest) as the in-browser admin page at /floor10/admin/members. Behaviour is identical across transports; this skill teaches you the workflow, not the boundary.

Pre-flight (every session)

1. Token check. You need an IC agent token with admin:tier_review in its scope set. Only operator-tier IC members can mint this scope — lib/capabilities.ts::SCOPES_BY_TIER.operator is the gate, enforced server-side at mint time. Smoke probe:

   curl -H "Authorization: Bearer $FLOOR10_AGENT_TOKEN" \
     -X POST -H "content-type: application/json" \
     -H "Accept: application/json, text/event-stream" \
     -d '{"jsonrpc":"2.0","id":"1","method":"tools/call","params":{"name":"ic_admin_list_pending_tier_requests","arguments":{}}}' \
     https://www.immersivecommons.com/api/mcp
   

   200 + result with pending: [] (or non-empty) = good. 401 = no/bad token. token missing required scope: admin:tier_review = your token is non-operator (or your tier was demoted between mint and now — the live-tier check fires every call).

   REQUIRED Accept header. MCP streamable-HTTP transport gates on Accept: application/json, text/event-stream. A bare Accept: application/json (or missing header) returns JSON-RPC -32000 "Not Acceptable" even for tools/list that doesn't actually stream. Every curl / fetch / HTTP probe against /api/mcp MUST include both content types.

2. Identity check. Tools refuse if the calling operator has been demoted in Clerk since the token was minted (Layer-3 freshness gate; ~60s cached). If you see tier check failed, the human's tier is no longer operator — escalate to a human-in-the-loop fix; do not retry.

The three verbs

1. ic_admin_list_pending_tier_requests

Read-only. Returns { pending: PendingTierRequest[], audit: TierAuditEntry[] }. Newest first. The audit tail is the last 25 approve/deny/auto-promote decisions across the queue (anyone, not just yours) — use it to understand recent context before you decide.

{ "name": "ic_admin_list_pending_tier_requests", "arguments": {} }

Pending record shape:

{
  "user_id": "user_2x...",            // Clerk userId
  "email": "jane@example.com",
  "display_name": "Jane Founder",
  "current_tier": "public",
  "requested_tier": "ic-member",
  "note": "I host VCN, attended every Lab night",
  "submitted_at": "2026-05-19T22:14:08.214Z"
}

2. ic_admin_approve_tier_request

Two-step. Default is a dry-run preview; confirm: true is required to mutate.

Step 1 — preview:

{
  "name": "ic_admin_approve_tier_request",
  "arguments": { "user_id": "user_2x..." }
}

Returns { ok: true, dry_run: true, from, to, action: "approve", was_pending, requested_tier, override_used } or { ok: false, error_kind: "no_pending" } if there's nothing to approve and you didn't pass a tier override.

Step 2 — commit:

{
  "name": "ic_admin_approve_tier_request",
  "arguments": {
    "user_id": "user_2x...",
    "reason": "Hosts VCN, verified attendance via graph + Telegram",
    "confirm": true
  }
}

Override: if you want to approve a member to a different tier than they requested (e.g. they asked for ic-member but you decide ai-floor), pass tier:

{
  "name": "ic_admin_approve_tier_request",
  "arguments": {
    "user_id": "user_2x...",
    "tier": "ai-floor",
    "reason": "Active in the AI floor channel but not yet at the IC-member bar",
    "confirm": true
  }
}

tier: "operator" is approvable here even though it isn't self-requestable. Use this for the rare promotion-from-trust path (Ray bringing in a deputy). Treat it as a separate review — see the hard rules below.

3. ic_admin_deny_tier_request

Same two-step shape. The user's tier is unchanged; only their pending request is cleared. The reason is recorded in the audit log AND surfaced to the requester on their /membership page so they understand why.

{
  "name": "ic_admin_deny_tier_request",
  "arguments": {
    "user_id": "user_2x...",
    "reason": "Haven't seen you on the floor yet — come hang at a VCN night and re-apply.",
    "confirm": true
  }
}

The judgement layer (read this before deciding)

Approving someone unlocks scope vocabulary they couldn't have before. Map each ring to what they get and what bar to apply:

Evidence before decision. Cross-reference each candidate against:

• The audit tail — has this user requested before, been denied, been demoted?
• Their note — is it specific (event names, dates, hosted-by) or generic (gen-ai vibes)?
• Their email domain — known sponsor / partner orgs (anthropic.com, openai.com, deepmind...) get scrutiny upward (don't auto-approve based on prestige, but do flag as worth knowing).
• The display_name — does it match someone on the Telegram floor channel? On Luma guest lists? (The agent that mints this token typically has membership:read; cross-link if you have it.)

Don't infer evidence. If the note is "I'm interested in AI," the human did not give you a reason — surface the gap, don't manufacture justification.

Hard rules

• No batch approvals. Each confirm:true call is one human-deliberated decision. If the agent encounters 10 pending requests, run 10 separate decision loops — list → preview → ask human → confirm. Never loop blindly.
• Self-promotion is a foot-gun. The MCP doesn't structurally prevent you from approving the operator themselves to a different tier (or worse, demoting themselves via an override-down approve). Refuse the call. The operator's own membership belongs in Ray's hands, not their agent's.
• tier: "operator" is a Ray-level decision. Surface the request to the human operator + Ray (if they're not the same person) explicitly. Don't approve operator promotions on a same-session agent decision unless Ray has confirmed in a durable channel (telegram message, commit message, written approval).
• Honor the dry-run. If the operator hasn't explicitly said "do it," return the preview and stop. Don't shortcut to confirm:true because the preview looked fine.
• Audit the audit. Before each approve, scan the recent audit tail for a pattern: was this user denied recently? Was their tier rolled back? Surface the prior decision to the human; don't approve over a recent denial without explanation.
• Never invent reasons. The reason field becomes part of the user's /membership page on a deny, and part of the audit log on either decision. Use only what the human told you, or quote the evidence (event attendance, graph link, ticket from sponsor).
• Rate limit awareness. 20 mutations per token per UTC day. If you're approaching the cap on a legitimate burndown day, stop and tell the human; do not try to rotate to a different token to evade the bound.

What gets logged

Every call writes to two places:

1. MCP activity log (per-user, readable via ic_activity_get_recent) — every tool call, including dry-runs.
2. Tier audit log (KVKeys.tierAudit, capped at 500) — every approve/deny mutation with by_clerk_user_id attribution to the operator's Clerk identity (NOT the agent's token).

If you see a decision in the audit log that you don't recognize, it's coming from another operator (or from a different agent under the same operator). Investigate before assuming compromise.

Failure modes & recovery

• no_pending on approve without tier. The user has nothing pending (cleared by another operator, or they cancelled). Use the tier override only if the human explicitly wants to grant a ring without a fresh request.
• tier check failed: tier <X> does not authorize scope admin:tier_review. The operator was demoted between mint and call. Stop; escalate to a human-in-the-loop.
• admin rate limit exceeded. You've hit 20 mutations in the UTC day. Stop. Resumes at 00:00 UTC.
• could not resolve user. The user_id doesn't exist in Clerk. Double-check the id from the pending array.

Discovery surfaces

• Agent card: https://www.immersivecommons.com/.well-known/agent-card.json
• Aiia manifest: https://www.immersivecommons.com/.well-known/ai-agent.json
• MCP discovery: https://www.immersivecommons.com/.well-known/mcp.json
• llms.txt URL map: https://www.immersivecommons.com/llms.txt
• Human-facing admin page: https://www.immersivecommons.com/floor10/admin/members

Related skills

• ic-onboarding — RFC 8628 device-code signup; how a fresh agent gets authorized.
• ic-headsets — Same operator scope pattern (admin:headsets_review) for the PICO lending program.
• ic-events — Sister surface; same auth model.