fix(strix): scope direct OpenAI tool fallback effort

docs/doctoring/strix-openai-fallback-api-base-routing.md

@@ -19,6 +19,18 @@ fallback key file, so every provider chain that ends in
 `openai-direct/gpt-5.4` (NVIDIA NIM primary, OpenRouter primary,
 GitHub Models primary) inherits correct routing automatically.
 
+Strix invokes function tools through chat completions. Direct OpenAI rejects
+those tools when `reasoning_effort` is non-none, so the gate scopes
+`STRIX_REASONING_EFFORT=none` to an explicit direct-OpenAI child using the
+native OpenAI endpoint. NVIDIA, other providers, and standalone explicit-model
+runs targeting a custom OpenAI-compatible endpoint retain their configured
+effort.
+
+Contextual-orchestrator PR #881 run `32967361853` demonstrated this boundary:
+the NVIDIA attempts exhausted with 429/410 provider responses, then the direct
+OpenAI fallback reached the correct endpoint but returned HTTP 400 because
+function tools were combined with `reasoning_effort=high`.
+
 ## Failure this fixes
 
 Required-CI evidence (BandScope PR #1021 strix run 32800796577, 2026-08-25)
@@ -71,11 +83,10 @@ Regression evidence proves that:
 8. the workflow provisions the override file and passes it into the gate env;
 9. the required-workflow smoke contract pins both sides of the wiring; and
 10. the stale `gpt-5.6-luna` expectations left behind by the model rename are
-   aligned with the valid `gpt-5.4` contract in queue-contract tests; and
-11. the direct GPT-5.4 fallback clears reasoning effort to `none` because the
-    pinned Strix scanner uses Chat Completions function tools, a combination
-    the provider rejects when `reasoning_effort` is present. Other models keep
-    the workflow's configured `high` effort.
+    aligned with the valid `gpt-5.4` contract in queue-contract tests; and
+11. explicit direct-OpenAI attempts on the native endpoint receive reasoning
+    effort `none`, while custom compatible endpoints and other providers retain
+    their configured effort.
 
 ## Limitations
 

scripts/ci/strix_quick_gate.sh

@@ -2551,6 +2551,7 @@ run_strix_once() {
 	local rc
 	local llm_api_base_value
 	local child_model
+	local child_reasoning_effort="${STRIX_REASONING_EFFORT:-}"
 	local resolved_target_path
 	local timeout_seconds="$STRIX_PROCESS_TIMEOUT_SECONDS"
 	local total_budget_limited_timeout=0
@@ -2574,20 +2575,19 @@ run_strix_once() {
 		return 2
 	fi
 	child_model="$(child_model_for_api_base "$model" "$llm_api_base_value")"
+	# Strix uses function tools. Direct OpenAI chat-completions rejects those
+	# tools when reasoning_effort is non-none, so scope the supported value to
+	# this provider attempt without weakening reasoning on other providers.
+	if is_explicit_openai_model "$model" &&
+		{ [ -z "$llm_api_base_value" ] || [ "${llm_api_base_value%/}" = "https://api.openai.com/v1" ]; }; then
+		child_reasoning_effort="none"
+	fi
 	if ! resolved_target_path="$(resolve_current_target_path "$TARGET_PATH")"; then
 		return 1
 	fi
 	local start_epoch
 	start_epoch="$(date +%s)"
 	local child_llm_api_key=""
-	local child_reasoning_effort="${STRIX_REASONING_EFFORT:-}"
-	case "$(normalize_model "$model")" in
-	openai-direct/gpt-5.4 | openai_direct/gpt-5.4)
-		# OpenAI rejects function tools plus reasoning_effort for GPT-5.4 on
-		# Chat Completions, the transport used by the pinned Strix scanner.
-		child_reasoning_effort="none"
-		;;
-	esac
 	if ! is_vertex_model "$(normalize_model "$model")"; then
 		child_llm_api_key="$LLM_API_KEY"
 		if is_github_models_model "$(normalize_model "$model")" && [ -n "$STRIX_GITHUB_MODELS_KEY" ]; then

scripts/ci/test_strix_quick_gate.sh

@@ -3332,7 +3332,7 @@ printf '%s\n' "$target_path" >> "${FAKE_STRIX_TARGET_LOG:?}"
 STRIX_REPORTS_DIR="${STRIX_REPORTS_DIR:-strix_runs}"
 
 case "${FAKE_STRIX_SCENARIO:?}" in
-success|runtime-env-forwarding|vertex-primary-success-timing-message|direct-openai-gpt-does-not-require-github-models-api-base|pr-executable-integrity-mismatch|pr-executable-group-writable)
+success|runtime-env-forwarding|custom-openai-compatible-preserves-effort|vertex-primary-success-timing-message|direct-openai-gpt-does-not-require-github-models-api-base|pr-executable-integrity-mismatch|pr-executable-group-writable)
 		echo "scan ok"
 		exit 0
 		;;
@@ -3416,7 +3416,7 @@ REPORT
 			;;
 		openai/gpt-5.4)
 			if [ "${STRIX_REASONING_EFFORT:-}" != "none" ]; then
-				echo "direct OpenAI GPT-5.4 fallback retained unsupported reasoning effort (${STRIX_REASONING_EFFORT:-<unset>})" >&2
+				echo "direct OpenAI function-tools fallback requires reasoning effort none" >&2
 				exit 29
 			fi
 			if [ "${LLM_API_KEY:-}" != "openai-fallback-token" ]; then
@@ -5660,7 +5660,7 @@ PY
 		STRIX_REPORTS_DIR="$repo_root_dir/strix_runs"
 		STRIX_TARGET_PATH="$effective_target_path"
 	)
-	if [ "$scenario" = "runtime-env-forwarding" ]; then
+	if [ "$scenario" = "runtime-env-forwarding" ] || [ "$scenario" = "custom-openai-compatible-preserves-effort" ]; then
 		env_cmd+=(
 			LLM_TIMEOUT="90"
 			STRIX_MEMORY_COMPRESSOR_TIMEOUT="10"
@@ -5808,6 +5808,7 @@ PY
 			-u STRIX_GEMINI_FALLBACK_MODELS \
 			-u STRIX_FALLBACK_MODELS \
 			-u STRIX_OPENAI_FALLBACK_KEY_FILE \
+			-u STRIX_OPENAI_FALLBACK_API_BASE_FILE \
 			"${env_cmd[@]}" \
 			bash "./scripts/ci/strix_quick_gate.sh" >"$output_log" 2>&1
 	)
@@ -5877,6 +5878,12 @@ PY
 			"LLM_TIMEOUT=90;STRIX_MEMORY_COMPRESSOR_TIMEOUT=10;STRIX_REASONING_EFFORT=minimal;STRIX_LLM_MAX_RETRIES=1;GEMINI_LOCATION=GLOBAL;PYTHONWARNINGS=ignore:Pydantic serializer warnings:UserWarning:pydantic.main;NPM_CONFIG_IGNORE_SCRIPTS=true;PNPM_CONFIG_IGNORE_SCRIPTS=true;YARN_ENABLE_SCRIPTS=false;UNRELATED_SECRET=<unset>" \
 			"scenario=$scenario runtime env forwarding"
 	fi
+	if [ "$scenario" = "custom-openai-compatible-preserves-effort" ]; then
+		assert_file_contains \
+			"$runtime_env_log" \
+			"STRIX_REASONING_EFFORT=minimal" \
+			"scenario=$scenario custom compatible endpoint effort"
+	fi
 
 	if [ "$scenario" = "report-known-internal-warning-sanitized" ]; then
 		assert_file_not_contains \
@@ -6160,6 +6167,18 @@ run_filtered_gate_case_if_requested() {
 			"" \
 			"github_models/deepseek/deepseek-v3-0324 github_models/deepseek/deepseek-r1-0528"
 		;;
+	custom-openai-compatible-preserves-effort)
+		run_gate_case "custom-openai-compatible-preserves-effort" \
+			"openai-direct/gpt-5.4" \
+			"" \
+			"0" \
+			"scan ok" \
+			"1" \
+			"openai/gpt-5.4" \
+			"https://compatible.example/v1" \
+			"openai" \
+			"https://compatible.example/v1"
+		;;
 	nvidia-rate-limit-openai-direct-fallback-clears-api-base)
 		run_gate_case_allow_provider_signal "nvidia-rate-limit-openai-direct-fallback-clears-api-base" \
 			"nvidia_nim/nvidia/rate-limited-primary" \
@@ -12526,6 +12545,17 @@ run_gate_case "github-models-model-prefix-requires-api-base" \
 	"openai" \
 	""
 
+run_gate_case "custom-openai-compatible-preserves-effort" \
+	"openai-direct/gpt-5.4" \
+	"" \
+	"0" \
+	"scan ok" \
+	"1" \
+	"openai/gpt-5.4" \
+	"https://compatible.example/v1" \
+	"openai" \
+	"https://compatible.example/v1"
+
 run_gate_case "github-models-api-base-rejected-for-direct-openai" \
 	"openai/o4-mini" \
 	"" \