llm-relay 0.68.4 → 0.68.6

package/README.md

@@ -12,10 +12,16 @@ weaker (often free) models.
 npm install -g llm-relay
 llm-relay onboard          # collect free provider keys (NIM, Groq, Gemini, OpenRouter, ...)
 llm-relay onboard --import keys.env  # import dotenv or a FreeLLMAPI export JSON
-llm-relay setup claude-desktop   # or: llm-relay setup claude-cli
+llm-relay setup claude-desktop   # register MCP dispatch; or set up a routed claude-cli
 llm-relay                  # start the proxy — leave it running
 ```
 
+For agent-to-agent work, use the `llm-relay` MCP `dispatch` tool from any host. It returns the
+chosen lane's answer and owns the process details. If MCP is unavailable, use
+`llm-relay dispatch --next-command -t "<task>"` and follow its result. Do not create a `pool/*`
+collaboration child in Codex Desktop: its ChatGPT launcher rejects that model before the relay is
+contacted. A global install registers the MCP server for local Codex automatically.
+
 Then verify:
 
 ```bash
@@ -42,9 +48,9 @@ at login.
   and outages cascade to the next member. Free-model pools update themselves from live catalogs.
 - **Passthrough** — Claude traffic keeps your own credentials and reaches real Anthropic
   untouched, while `pool/*` requests go elsewhere. One proxy, both behaviours.
-- **Opt-in offload** — route Claude/Codex subagents (or whole conversations) to free
-  providers. `llm-relay candidates` compares targets; a `freeOnly` guard ensures rerouted
-  traffic never spends money.
+- **Host-independent offload** — MCP `dispatch` chooses and runs a lane from Claude, Codex,
+  desktop, CLI, or another MCP host; the CLI dispatch contract is the universal fallback. Direct
+  HTTP-routed clients can also opt in to rerouting marked subagents or whole conversations.
 - **Tool-call repair** — malformed tool calls are corrected and re-validated; destructive
   tool calls are refused, never fabricated; unrepairable calls fail clean.
 - **Both API fronts** — Anthropic `/v1/messages` plus OpenAI `/v1/chat/completions` and

package/dist/cli.js

@@ -172,7 +172,7 @@ Usage:
 ${formatTextTable([
     ["llm-relay [options]", "Start proxy."],
     ["llm-relay onboard [--import <file>] [--force]", "Set up or import provider keys."],
-    ["llm-relay setup [target]", "target: claude-cli | claude-desktop."],
+    ["llm-relay setup [target]", "claude-desktop: MCP dispatch; claude-cli: routed CLI wrapper."],
     ["llm-relay keys | keys check | check-keys", "Check every configured credential slot."],
     ["llm-relay keys <action> ...", "action: add|list|rotate|revoke|remove|disable|enable|export|import|unlock."],
     ["llm-relay pools [--probe]", "List members; --probe tests each deployment once."],
@@ -199,7 +199,7 @@ ${formatTextTable([
     ["llm-relay dispatch [lane] [options]", "Choose next dispatch lane; adapts to the calling host."],
     ["llm-relay dispatch --next-command -t <task>", "Print only the runnable command for the next lane."],
     ["llm-relay delegate-gate <diff-file> --repo <root> [--fix]", "Quality-gate a delegated lane's diff before judgment/merge."],
-    ["llm-relay mcp", "Serve the dispatch verb over MCP on stdio, so any MCP host can delegate."],
+    ["llm-relay mcp", "Preferred dispatch entry point for every MCP host; one call returns an answer."],
     ["llm-relay help | --help | -h", "Show help."],
     ["llm-relay version | --version | -v", "Print version."],
 ], "  ")}
@@ -224,6 +224,11 @@ subagent can be rerouted and "@relay:" is inert. "llm-relay dispatch" detects th
 back runnable commands instead; "offload claude on" there also installs a PreToolUse(Agent)
 hook that redirects Agent() calls to the same lane. "offload claude off" removes it.
 
+Codex Desktop is also not a native relay-child path: its ChatGPT collaboration launcher rejects a
+pool/* child model before contacting the custom provider or llm-relay. Use the llm-relay MCP
+dispatch tool there. On any host without MCP, use "dispatch --next-command" and follow its result
+instead of guessing which child mechanism the host supports.
+
 Setup checks: "llm-relay keys" checks every credential slot; "llm-relay pools --probe" spends one
 real completion per unique pool deployment through one serviceable slot, not every credential.
 Environment variables override ~/.llm-relay/.env.