mlclaw 0.5.0 → 0.6.1

package/.agents/skills/mlclaw/SKILL.md

@@ -290,6 +290,22 @@ settings to set a Space Secret; do not put the user's broad Hub token into the
 Space just so the browser app can mutate secrets. The key must never be logged
 or returned to the browser.
 
+Deployment-scoped ChatGPT login uses OpenAI's Codex device-code protocol
+directly and does not require the Codex CLI:
+
+```bash
+mlclaw credentials codex login <agent>
+```
+
+After login, ML Claw imports the deployment credential into a managed native
+OpenClaw OAuth profile and enables OpenClaw's bundled Codex app-server harness
+for `openai/*` agent models. OpenClaw contacts OpenAI directly, refreshes the
+OAuth credential in its private auth store, and discovers the account-visible
+catalog, including `openai/gpt-5.6-sol` when entitled. ML Claw does not proxy
+model discovery or inference and does not install or invoke a Codex CLI from
+`PATH`. Normal `OPENAI_API_KEY` profiles can coexist with subscription OAuth;
+OpenAI agent turns use the Codex harness with either credential type.
+
 ## Branding
 
 ML Claw white-labels the browser gateway at the proxy layer. App Spaces default

package/Dockerfile

@@ -2,7 +2,7 @@ ARG OPENCLAW_VERSION=2026.7.1
 ARG OPENCLAW_BASE_IMAGE=ghcr.io/openclaw/openclaw:${OPENCLAW_VERSION}
 ARG BROKERKIT_PLUGIN_VERSION=0.4.1
 ARG BROKERKIT_VERSION=hf-broker/v0.6.2
-ARG MLCLAW_RUNTIME_IMAGE=ghcr.io/huggingface/mlclaw:0.5.0-openclaw-2026.7.1
+ARG MLCLAW_RUNTIME_IMAGE=ghcr.io/huggingface/mlclaw:0.6.1-openclaw-2026.7.1
 
 FROM golang:1.26.5-bookworm AS hf-broker-build
 ARG BROKERKIT_VERSION

package/README.md

@@ -371,6 +371,40 @@ never returned to the browser, exposed to the agent except as its intended
 OpenAI credential, or stored as plaintext in the state bucket. App Spaces do
 not keep broad Hugging Face authority in the web control process.
 
+To connect an OpenAI/ChatGPT account through Codex device login instead of an
+API key, target the specific deployment. ML Claw performs the device-code flow
+directly over HTTPS. It does not install, invoke, or depend on the Codex CLI.
+
+```bash
+mlclaw credentials codex login mlclaw
+mlclaw credentials codex status mlclaw
+mlclaw credentials codex logout mlclaw
+```
+
+`login` requests a one-time device code from OpenAI, prints the verification
+URL and code, polls for approval, encrypts the OAuth credential into the
+selected deployment's private state bucket, and restarts the deployment.
+OpenClaw then discovers the models visible to that ChatGPT account under its
+native `openai/*` provider. For example, an entitled account may expose
+`openai/gpt-5.6-sol` alongside the other live Codex models.
+
+At startup, ML Claw imports the encrypted login into a managed native OpenClaw
+OAuth profile and enables OpenClaw's bundled Codex app-server harness for
+`openai/*` agent models. Codex owns the low-level model loop, native thread
+state, tool continuation, and compaction; OpenClaw continues to own channels,
+approvals, dynamic tools, model selection, and the visible transcript mirror.
+OpenClaw manages its compatible app-server binary, so ML Claw still does not
+install or invoke a Codex CLI from `PATH`.
+
+OpenClaw contacts OpenAI directly for model discovery, inference, and token
+refresh; ML Claw is not in the inference path. OpenClaw keeps its own model
+names, metadata, transport, and refreshed credential, and a normal
+`OPENAI_API_KEY` profile can coexist with the ChatGPT route. OpenAI agent models
+also use the Codex harness with API-key auth. `status` and `logout` inspect or
+revoke the deployment-scoped login without exposing tokens. Raw OAuth tokens
+are never printed, stored in the Space repository or model configuration, or
+included in workspace files.
+
 ## How State Works
 
 OpenClaw runs against local disk inside the active runtime. In Space gateway