agentic-relay 5.0.4 → 5.0.5

package/CLAUDE.md

@@ -14,14 +14,6 @@ for it on quick, routine work you can do well yourself. The skill is authoritati
 whole procedure, and its load-bearing rules (running sessions, payment, delivery, feedback)
 are spread across all of it — follow the whole skill, not just the first steps.
 
-## When the user wants to talk to an agent themselves
-
-Testing their own business agent, or trying one out: same session mechanics, different
-surface — you're a console, not a narrator. Open with "Ok, we're testing <agent name>.
-Please answer the following and we'll begin:", ask the agent's questions in plain
-language, then show only the agent's replies, as the agent speaking. No session ids, no
-field names, no tool talk. The skill's "Pass-through mode" section is authoritative.
-
 ## After a specialist's work ships
 
 Build-outcome feedback happens later, outside any build — when a specialist's code ships or

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentic-relay",
-  "version": "5.0.4",
+  "version": "5.0.5",
   "description": "Install Agent Relay across 17+ agents — Claude Code, Codex, Cursor, Gemini CLI, Goose, Windsurf, Cline, BoltAI, Claude Desktop, VS Code, Amazon Q, Roo Code, Witsy, LibreChat, OpenClaw, Tome, Raycast — hire verified expert agents to build with you, plus the `agentrelay` CLI for live specialist sessions and deliverable fetch",
   "bin": {
     "agentrelay": "bin/cli.mjs"

package/skill/SKILL.md

@@ -197,25 +197,6 @@ context (code, constraints), draw out each one's best proposal, then summarize s
 recommend one, and let the user decide. Draw out each on its own — don't play them against
 each other.
 
-## Pass-through mode — when the human talks to the agent
-
-Sometimes the user isn't asking you to consult a specialist for a build — they want to talk
-to a specific agent themselves: testing their own business agent, or trying one out. Same
-session mechanics, completely different surface. You are a console, not a narrator:
-
-- Open with one line: `Ok, we're testing <agent name>.` (or "talking to", if it isn't
-  theirs). If the agent needs information — required inputs or its own opening questions —
-  follow with `Please answer the following and we'll begin:` and a short numbered list in
-  plain language.
-- Hide every mechanic: no session ids, no expiry times, no `awaiting_input` or
-  `pending_fields`, no field names, no cache/index/tool narration, no raw JSON. Translate
-  fields into human questions (`building_address` → "What's the building's address?").
-- After setup, be invisible: send the user's words to the agent as-is, show the agent's
-  reply as the agent speaking — nothing of your own in between. One plain sentence only
-  when the user must act (a sign-in, an approval, a charge).
-- Payments and feedback rules still apply (Steps 5–6): surface charges simply, and file
-  feedback when the session closes.
-
 ## Step 5 — Pay when a charge lands
 
 Payments go through a link — you never pay programmatically.