@enderfga/claw-orchestrator 7.17.1 → 7.18.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@enderfga/claw-orchestrator",
-  "version": "7.17.1",
+  "version": "7.18.0",
   "description": "Claw Orchestrator — run Claude Code, Codex, Antigravity, Grok Build, OpenCode, any ACP agent and custom coding CLIs as one unified runtime. Drop into Hermes Agent, Claude Desktop, Cursor, Cline, Continue, Zed, Windsurf, Goose or any Model Context Protocol (MCP) host, serve editors as an ACP agent, install as an OpenClaw plugin, or run standalone. Persistent sessions, durable resumable workflows, acceptance checks the runtime runs itself, fan-out, multi-agent council, Codex /goal runs, autoloop, ultraplan, ultrareview, and a per-turn cost ledger with spend caps.",
   "type": "module",
   "main": "./dist/src/index.js",

package/skills/references/multi-engine.md

@@ -500,6 +500,15 @@ What happens behind the scenes:
 2. Starts a local proxy server on `127.0.0.1` (random port, auto-managed) the first time a Claude session names a non-Claude model, provided a gateway URL was found
 3. Claude Code CLI sends Anthropic-format requests → proxy converts to OpenAI → gateway → any model
 
+### Your own Anthropic-compatible gateway
+
+A gateway that speaks the Anthropic API and serves other models under their own names (an internal model
+gateway, for instance) needs none of this. Set `CLAWO_CLAUDE_GATEWAY=1` in the environment clawo runs in,
+with the gateway in `ANTHROPIC_BASE_URL` (and its token in `ANTHROPIC_AUTH_TOKEN`) or in a session's
+`baseUrl`. A `claude` session then passes any model name to the CLI as given (`--model gw-model-1`), starts no
+proxy, and does not rewrite the base URL. Because it is read from the environment, it applies to every
+`claude` session the process starts, including those of workflow, fan-out and council nodes.
+
 ### Manual Config (optional)
 
 Override with environment variables if needed:

package/skills/references/sessions.md

@@ -284,7 +284,7 @@ Built-in format translation lets Claude Code CLI talk to non-Anthropic models:
 - **Gemini** thought signature caching (round-trip thinking)
 - Auto-detect provider from model name patterns
 
-A `claude` session started with a non-Claude model and no `baseUrl` starts the proxy on a local port and points the session at it. Configuration is in the proxy section of [Multi-Engine](./multi-engine.md); see [`src/proxy/`](https://github.com/Enderfga/claw-orchestrator/tree/main/src/proxy) for implementation details.
+A `claude` session started with a non-Claude model and no `baseUrl` starts the proxy on a local port and points the session at it, unless `CLAWO_CLAUDE_GATEWAY=1` says the caller's own gateway serves that model (see [Multi-Engine](./multi-engine.md)). Configuration is in the proxy section of [Multi-Engine](./multi-engine.md); see [`src/proxy/`](https://github.com/Enderfga/claw-orchestrator/tree/main/src/proxy) for implementation details.
 
 ## Circuit Breaker
 

package/skills/references/workflow.md

@@ -411,10 +411,11 @@ what an unattended run needs around that loop.
   the finished goal and sets a new objective on the same thread, carrying the
   failing checks (as diagnostic data, like an agent node's repair brief) with the
   original objective. The agent keeps its context. On Codex the checks come first,
-  as the definition of done. On Claude they come after the objective as context,
-  and the condition the goal is judged on stays the objective (see below). The
-  original objective is always kept whole; when the rest does not fit in 4,000
-  characters the checks are cut (the full output stays in the run's evidence). A
+  as the definition of done. On Claude they are sent as a message of their own
+  just before the goal, which is set to the original objective word for word (see
+  below). On Codex the original objective is always kept whole; when the rest does
+  not fit in 4,000 characters the checks are cut (the full output stays in the
+  run's evidence). A
   first goal that was never set left nothing on its thread, so its revisit starts
   a new one.
 - **A fresh thread is told why.** Any goal objective set while the run carries a
@@ -459,9 +460,11 @@ stream-json mode:
   asking to stop, and continuing only repeats that). The template's verifier
   still judges the tree. Write the objective as a condition the agent can show
   it met from inside the session; one that depends on something after the
-  agent stops can never be met. That is why a repair objective on Claude does
-  not call the runtime's checks the definition of done: they run after the agent
-  stops, so the judge could never see them pass.
+  agent stops can never be met. Everything after `/goal` is that condition, so a
+  repair never adds to it: steers and the failing checks are sent first as an
+  ordinary message (the agent replies with one word), and `/goal` carries the
+  objective alone. The checks run after the agent stops, where the judge could
+  never see them pass.
 - The session id is the thread id. A new session is started with
   `--session-id` so it is known before the goal is set, and the session is kept
   out of the registry with `skipPersistence` (not `noSessionPersistence`, which