docs: clarify strict-agentic and codex modes

docs/plugins/sdk-agent-harness.md

@@ -133,6 +133,25 @@ OpenClaw requires Codex app-server `0.118.0` or newer. The Codex plugin checks
 the app-server initialize handshake and blocks older or unversioned servers so
 OpenClaw only runs against the protocol surface it has been tested with.
 
+### Native Codex harness mode
+
+The bundled `codex` harness is the native Codex mode for embedded OpenClaw
+agent turns. Enable the bundled `codex` plugin first, and include `codex` in
+`plugins.allow` if your config uses a restrictive allowlist. It is different
+from `openai-codex/*`:
+
+- `openai-codex/*` uses ChatGPT/Codex OAuth through the normal OpenClaw provider
+  path.
+- `codex/*` uses the bundled Codex provider and routes the turn through Codex
+  app-server.
+
+When this mode runs, Codex owns the native thread id, resume behavior,
+compaction, and app-server execution. OpenClaw still owns the chat channel,
+visible transcript mirror, tool policy, approvals, media delivery, and session
+selection. Use `embeddedHarness.runtime: "codex"` with
+`embeddedHarness.fallback: "none"` when you need to prove that the Codex
+app-server path is used and PI fallback is not hiding a broken native harness.
+
 ## Disable PI fallback
 
 By default, OpenClaw runs embedded agents with `agents.defaults.embeddedHarness`

docs/providers/openai.md

@@ -3,6 +3,7 @@ summary: "Use OpenAI via API keys or Codex subscription in OpenClaw"
 read_when:
   - You want to use OpenAI models in OpenClaw
   - You want Codex subscription auth instead of API keys
+  - You need stricter GPT-5 agent execution behavior
 title: "OpenAI"
 ---
 
@@ -477,6 +478,33 @@ behavior, but it does not receive the hidden OpenAI/Codex attribution headers.
 This preserves current native OpenAI Responses behavior without forcing older
 OpenAI-compatible shims onto third-party `/v1` backends.
 
+### Strict-agentic GPT mode
+
+For `openai/*` and `openai-codex/*` GPT-5-family runs, OpenClaw can use a
+stricter embedded Pi execution contract:
+
+```json5
+{
+  agents: {
+    defaults: {
+      embeddedPi: {
+        executionContract: "strict-agentic",
+      },
+    },
+  },
+}
+```
+
+With `strict-agentic`, OpenClaw no longer treats a plan-only assistant turn as
+successful progress when a concrete tool action is available. It retries the
+turn with an act-now steer, auto-enables the structured `update_plan` tool for
+substantial work, and surfaces an explicit blocked state if the model keeps
+planning without acting.
+
+The mode is scoped to OpenAI and OpenAI Codex GPT-5-family runs. Other providers
+and older model families keep the default embedded Pi behavior unless you opt
+them into other runtime settings.
+
 ### OpenAI Responses server-side compaction
 
 For direct OpenAI Responses models (`openai/*` using `api: "openai-responses"` with