diff --git a/litellm/README.md b/litellm/README.md
index 6c3afdc1..a915d2f3 100644
--- a/litellm/README.md
+++ b/litellm/README.md
@@ -9,6 +9,11 @@ Messages** format (`POST /v1/messages`), but the target models don't:
 |---|---|---|
 | Bedrock open-weight (GLM, DeepSeek, Kimi) | Converse | yes |
 | OpenRouter (any model) | OpenAI Chat Completions | yes |
+| OpenAI / Azure-OpenAI direct (GPT / Codex) | OpenAI Chat Completions | yes |
+
+The last row is how we test **skill-activation reliability on a GPT/Codex model**:
+route it through this same Claude Code SDK harness instead of the Codex SDK (which
+does not run Claude-Code skills), so the native `Skill` tool fires and can be graded.
 
 **LiteLLM is the translation shim.** The SDK is pointed at LiteLLM via
 `ANTHROPIC_BASE_URL`; LiteLLM translates Anthropic ↔ Converse / OpenAI per model
@@ -46,6 +51,7 @@ environment before launching:
 | `AWS_BEARER_TOKEN_BEDROCK` | Bedrock models | required if you use any `bedrock/*` model |
 | `AWS_REGION` | Bedrock models | defaults to `eu-north-1` |
 | `OPENROUTER_API_KEY` | OpenRouter models | required if you use any `openrouter/*` model |
+| `CODEX_API_KEY` / `CODEX_BASE_URL` | OpenAI / Azure-OpenAI (GPT / Codex) models | required if you use the `gpt-*` OpenAI-compatible model; base URL is reused verbatim (e.g. Azure's `/openai/v1`) |
 | `LITELLM_AUTH_TOKEN` | the virtual key clients present | becomes the proxy's `LITELLM_MASTER_KEY`; falls back to `sk-spike-local`. It can be customly designed.|
 
 ### Start it
diff --git a/litellm/litellm-config.yaml b/litellm/litellm-config.yaml
index 5ec10b8e..4f52ffa1 100644
--- a/litellm/litellm-config.yaml
+++ b/litellm/litellm-config.yaml
@@ -18,8 +18,10 @@
 #   AWS_REGION=eu-north-1      — EU residency (Stockholm).
 #   LITELLM_MASTER_KEY         — the virtual key clients present as LITELLM_AUTH_TOKEN.
 #
-# Run manually:
-#   uvx --from 'litellm[proxy]' litellm --config litellm/litellm-config.yaml --port 4000
+# Run manually (pin fastapi==0.140.0 — a later 0.140.x patch removed a symbol
+# litellm's proxy imports; start-litellm.sh does this for you):
+#   uvx --from 'litellm[proxy]==1.95.0' --with 'fastapi==0.140.0' \
+#     litellm --config litellm/litellm-config.yaml --port 4000
 
 model_list:
   # DeepSeek V3.2 — cost lead ($0.74 / $2.22 per Mtok).
@@ -92,6 +94,37 @@ model_list:
           allow_fallbacks: true
           only: ["streamlake", "gmicloud", "novita", "alibaba"]
 
+  # OpenAI / Azure-OpenAI direct — drive a GPT/Codex-family model through the SAME
+  # Claude Code SDK harness (LiteLLM translates Anthropic Messages <-> OpenAI Chat
+  # Completions) so the native `Skill` tool + progressive disclosure fire and we can
+  # measure skill-activation reliability on a non-Claude model — the thing the Codex
+  # SDK path cannot test (it does not run Claude-Code skills).
+  #
+  # The endpoint is OpenAI-compatible: Azure's `/openai/v1` surface accepts a Bearer
+  # key and OpenAI-format requests, so the generic `openai/` provider + a custom
+  # `api_base` mirrors exactly how the Codex agent already reaches it (bearer auth,
+  # no api-version). `model:` after the `openai/` prefix is the Azure DEPLOYMENT name.
+  # No usage.include/provider block — those are OpenRouter-only; OpenAI/Azure returns
+  # no usage.cost, so cost falls back to the static pricing table (gpt-5.6-terra is
+  # registered), which is exact for a flat per-token rate.
+  - model_name: gpt-5.6-terra
+    litellm_params:
+      model: openai/gpt-5.6-terra
+      api_base: os.environ/CODEX_BASE_URL
+      api_key: os.environ/CODEX_API_KEY
+
+  # Gemini (Google) direct — drive a Gemini model through the SAME Claude Code SDK
+  # harness so the native `Skill` tool fires, to A/B skill activation against Gemini's
+  # own Antigravity SDK. LiteLLM's `gemini/` provider calls the Gemini Developer API
+  # with GEMINI_API_KEY and does the Anthropic<->Gemini translation (incl. tool_use
+  # <-> function-calling, which is what the Skill tool rides). No usage.include block —
+  # that's OpenRouter-only; Gemini returns no usage.cost, so cost falls back to the
+  # static pricing table (gemini-3.1-pro-preview is registered).
+  - model_name: gemini-3.1-pro-preview
+    litellm_params:
+      model: gemini/gemini-3.1-pro-preview
+      api_key: os.environ/GEMINI_API_KEY
+
 litellm_settings:
   # Silently drop provider-unsupported OpenAI/Anthropic params rather than 400.
   drop_params: true
diff --git a/litellm/start-litellm.sh b/litellm/start-litellm.sh
index 169ece47..b495b30c 100755
--- a/litellm/start-litellm.sh
+++ b/litellm/start-litellm.sh
@@ -48,6 +48,15 @@ export AWS_REGION="${AWS_REGION:-$(read_env AWS_REGION)}"
 export AWS_REGION="${AWS_REGION:-eu-north-1}"
 # OpenRouter key for the openrouter/* models in the config (cost-optimization path).
 export OPENROUTER_API_KEY="${OPENROUTER_API_KEY:-$(read_env OPENROUTER_API_KEY)}"
+# OpenAI / Azure-OpenAI key + base URL for the gpt-* models in the config — the
+# Claude-SDK-through-LiteLLM skill-activation path. CODEX_BASE_URL is an
+# OpenAI-compatible endpoint (e.g. Azure's /openai/v1), reused verbatim so the
+# proxy reaches the model the same way the Codex agent does.
+export CODEX_API_KEY="${CODEX_API_KEY:-$(read_env CODEX_API_KEY)}"
+export CODEX_BASE_URL="${CODEX_BASE_URL:-$(read_env CODEX_BASE_URL)}"
+# Gemini key for the gemini/* model in the config — the Gemini-via-Claude-SDK
+# skill-activation A/B against the native Antigravity harness.
+export GEMINI_API_KEY="${GEMINI_API_KEY:-$(read_env GEMINI_API_KEY)}"
 # Master key = the key clients present. Default to .env's LITELLM_AUTH_TOKEN so the
 # client and proxy match; fall back to the local dev key.
 export LITELLM_MASTER_KEY="${LITELLM_MASTER_KEY:-$(read_env LITELLM_AUTH_TOKEN)}"
@@ -92,8 +101,17 @@ Set these in coder_eval's .env (or shell) to use it:
   LITELLM_BASE_URL=http://localhost:$PORT
   LITELLM_AUTH_TOKEN=$LITELLM_MASTER_KEY
   LITELLM_COST_LOG=$LITELLM_COST_LOG   # SAME path as above, so the run joins ACTUAL per-call cost (else static pricing)
-  # then run:  coder-eval run <task> --model zai.glm-5   (or deepseek.v3.2 / moonshotai.kimi-k2.5)
+  # then run:  coder-eval run <task> --model zai.glm-5   (or deepseek.v3.2 / moonshotai.kimi-k2.5 / gpt-5.6-terra)
 
 EOF
 
-exec uvx --from 'litellm[proxy]' litellm --config "$CONFIG" --host 127.0.0.1 --port "$PORT"
+# Pin the proxy deps. `uvx --from 'litellm[proxy]'` unpinned drifts: litellm 1.95.0
+# declares `fastapi>=0.136.3,<1.0`, so uvx grabs the newest fastapi — but fastapi
+# dropped `get_flat_dependant` (which litellm's proxy still imports) in a 0.140.x
+# PATCH (0.140.0 has it, 0.140.13 doesn't), so a range cap isn't enough and startup
+# dies with a (masked) `ModuleNotFoundError: proxy_server`. Pin fastapi to an exact
+# verified-good version and pin litellm so the sidecar can't silently re-break.
+# Override for an upgrade: LITELLM_SPEC='litellm[proxy]==<ver>' LITELLM_FASTAPI='fastapi==<ver>'.
+LITELLM_SPEC="${LITELLM_SPEC:-litellm[proxy]==1.95.0}"
+LITELLM_FASTAPI="${LITELLM_FASTAPI:-fastapi==0.140.0}"
+exec uvx --from "$LITELLM_SPEC" --with "$LITELLM_FASTAPI" litellm --config "$CONFIG" --host 127.0.0.1 --port "$PORT"