profile-summon

name: profile-summon description: >- Use when a directory has no .cue.profile, or the user says "summon", "load", "apply", or "pull in" a profile, or "no profile here". Soft-loads a profile's persona and skills into the LIVE session and pins it, no restart. tags: [meta, cue, routing, profile] category: meta version: 1.0.0 requires_mcps: [] allowed-tools: Bash(cue:), Bash(jq:), Read triggers: - "summon a profile" - "load a profile" - "apply the vercel profile" - "pull in the X profile" - "no profile here" - "bind a profile" - "which profile for this repo"

profile-summon

Prerequisites

• cue CLI on PATH (the profile manager that ships this skill).
• jq (apt install jq / brew install jq) to read cue summon --json.

Open a directory with no .cue.profile and the right profile's skills and MCPs normally need a pin plus a full claude restart, because CLAUDE_CONFIG_DIR, the Skill() list, MCP servers, /slash commands, and plugins are all frozen at boot. This skill is the bridge: it binds a profile into the running session now and hands back a warm re-exec for the parts a soft load cannot provide.

Two tiers, honestly

1. Now, zero restart (soft-load). cue summon lists the profile's persona and each skill's SKILL.md path. You Read those paths and apply the persona plus the relevant skill playbooks inline. Same mechanism as meta/smart-loader, just whole-profile.
2. Durable plus full fidelity (warm re-exec). cue summon pins .cue.profile so the next launch is correct, and prints claude --continue. Running that resumes THIS conversation under the fully materialized profile, the only path to its MCP servers and /slash commands.

Iron contract

1. Never fake the harness. MCP tools, /slash commands, and plugins cannot be hot-added. Do not pretend to call an MCP tool or run a /command that is not in your actual tool list. The warm re-exec is the sanctioned path to them.
2. Only follow SKILL.md you actually read. Read each path from cue summon output this session before applying it. No paraphrasing from memory.
3. Honest framing. Tell the user once, plainly, what soft-loaded and what still needs claude --continue (see the example below).
4. Propose, then apply. On a no-profile directory, show the detected profile and what it will soft-load in 3 to 4 lines, then apply on the user's OK. Do not inject a profile's persona silently.
5. One summon per topic per session. Once a profile is soft-loaded, it stays applied. Do not re-run for the same directory.

When to summon

Summon when any holds:

• The directory has no .cue.profile (the first-time launch block points here).
• The user names a profile to load: "summon vercel", "apply the backend profile".
• The user asks "which profile for this repo" and wants it active now, not later.

Do not summon when a .cue.profile is already pinned and active, or when a loaded skill already covers the task (use Skill() directly).

The recipe

Step 1, resolve and inspect

Auto-detect from the repo, or pass an explicit profile:

cue summon --json            # auto-detect from cwd
cue summon vercel --json     # force a known profile

Parse the result. The useful fields:

cue summon vercel --json | jq '{profile, detected, confidence,
  soft_loadable: [.skills[] | select(.mcp_status=="ok") | {id, path}],
  gated: [.skills[] | select(.mcp_status!="ok") | {id, mcp_status}],
  mcps, commands, plugins, reexec_cmd}'

--json writes the .cue.profile pin by default. Add --no-pin to inspect first, --pick to list candidates without acting, --dry-run to compute and write nothing.

Step 2, propose

When you auto-detected (no profile named), show the user a short proposal before applying:

No .cue.profile here. Detected: vercel (97% match: vercel.json, @vercel dep).
Summon it? Soft-loads persona + N skills now, pins .cue.profile, no restart.
The Vercel MCP and /deploy need `claude --continue` after.

Step 3, apply inline

On the user's OK, apply the persona, then Read each mcp_status:"ok" skill path and follow the ones the task needs. Skills with mcp_status:"missing:...", npx skills, /slash commands, and plugins do not soft-load: list them as available after the warm re-exec.

Step 4, hand back the warm re-exec

Close with the honest summary:

Soft-loaded the vercel persona plus N skill playbooks into this session and pinned .cue.profile. The Vercel MCP (mcp.vercel.com) and /deploy /env commands ride in the plugin, which needs the harness. To get them, run claude --continue: it resumes this exact conversation under the full profile.

What can never be faked (needs claude --continue)

Example: Vercel summon

User opens a Next.js plus Vercel repo (has vercel.json) with no .cue.profile.

cue summon --json | jq '{profile, confidence, reasons}'
# → { "profile": "vercel", "confidence": 0.97,
#     "reasons": ["vercel.json", "package.json @vercel/* or vercel"] }

Propose vercel, apply on OK: persona plus the soft-loadable skills (agent-browser, the framework playbooks) apply now. The hosted MCP, /deploy, and /env are flagged as needing claude --continue. The pin is written so that re-exec lands the full profile and resumes the chat.

What this skill is NOT

• Not an MCP shim. It never calls mcp.vercel.com for you.
• Not a slash-command emulator. /deploy stays a harness command.
• Not a search engine. To find one off-profile skill for a topic, use meta/smart-loader. profile-summon binds a whole profile.

Linking

• Related: [[smart-loader]], soft-loads a single off-profile skill by topic.
• Related: [[profile-suggest]], suggests a profile when none is pinned.
• Related: [[focus]], routes among already-loaded skills.